Merge pull request 'Give Stage1Model/Stage2OneShot/Stage2Autoregressive a shared StageModel base (gitea #39)' (#56) from fix/issue-39 into master

Reviewed-on: #56

.gitea/workflows/ci.yml

@@ -82,7 +82,11 @@ jobs:
           echo "UV_CACHE_DIR=/uv-cache" >> "$GITHUB_ENV"
           echo "UV_LINK_MODE=copy" >> "$GITHUB_ENV"
       - run: uv sync --extra cpu --extra dev
-      - run: uv run pytest
+      - run: uv run pytest --cov --cov-report=term-missing --cov-report=xml
+      - uses: actions/upload-artifact@v3
+        with:
+          name: coverage-report
+          path: coverage.xml
 
   sync-version-on-tag:
     name: Sync project version with tag

.gitignore

@@ -20,3 +20,8 @@ checkpoints/
 
 # giant analyze run directories (shared.json, reduced/, plots/, condor logs)
 /analysis_runs/
+
+# Coverage artifacts
+.coverage
+coverage.xml
+htmlcov/

CLAUDE.md

@@ -22,7 +22,7 @@ giant analyze render <run_dir> --gallery                   # render PDFs + HTML
 dwarf --help                                     # dataset/tooling CLI: convert, migrate, bump-gen,
                                                   # bump-schema, status, update-manifest, create-manifest,
                                                   # make-root, build-geometry-oracle, warm-cache, hparam-scan
-                                                  # (see scripts/dwarf.py)
+                                                  # (see giant/tools/dwarf.py)
 ```
 
 `cpu` and `cuda` are mutually exclusive — pick one to select the torch build (pinned to 2.3.x; newer torch requires newer NVIDIA drivers). Plain `uv sync` with no extra will not install torch at all; uv has no concept of a "default extra", so `--extra cpu` should always be included unless you need GPU support.
@@ -91,4 +91,6 @@ GIANT is a conditional generative surrogate for the Geant4 step function. It rep
 
 A sampling-calorimeter (multi-material) dataset is still a planned future direction, not yet built. See the knowledge base (`/home/lars/knowledge-base/meta/roadmap.md`).
 
+**v0.3.0 — Stage-2 autoregressive redesign (designed, not implemented; branch `v0.3.0-stage2-autoregressive`):** the 2026-08-03 WGAN rollout benchmark failed specifically at the secondary-species level (zero photon secondaries, ~4M hallucinated `-14` muon antineutrinos). The agreed response pivots Stage 2 to **autoregressive generation** in descending-energy order with teacher forcing, and switches the particle-type representation back to **categorical** (top N−1 by training-set count + an "other" bucket), reversing the 2026-07-17 continuous `(log-mass, charge)` target. This requires a config break: `[conditioning]` / `[stage1_model]` / `[stage2_model]` / `[train]` blocks replace the single global `train.mode` + `[model]`, so per-stage generators (`stage1 = flow` + `stage2 = wgan`), stage-2-only training, and one-shot-vs-autoregressive comparison are all expressible. `network.py` is refactored from ten permutation classes into composable parts (encoder × trunk × objective), which also makes routed WGAN work for the first time. This config break is why v0.2-shaped configs/checkpoints need migrating at all (`config.migrate_config`, `model.network._migrate_legacy_model_config`, both drawing on shared facts in `giant/_migration.py`) — v0.2 checkpoint-loading support has **no expiry decided yet**: `/ceph` still holds pre-v0.3.0 checkpoints and analysis runs referencing them, so don't delete or substantially alter either migration function or `tests/legacy/network_v02_snapshot.py` (the frozen v0.2 snapshot they're tested against) without an explicit decision to do so first.
+
 **Condor-submitted GPU training/rollout (in progress, `condor-gpu-train-rollout` branch, not yet merged):** moves `giant train`/`giant rollout` off the shared portal GPU dev machines (see Compute environment) onto remote-GPU HTCondor submission on TOpAS/NEMO2 (`giant/condor.py`). Partway between "needs major features" and feature-complete — not ready to merge yet.

README.md

@@ -1,20 +1,29 @@
 # giant
 
-**G**eant4 **I**nference via **A**utoregressive **N**eural s**T**ep surrogate — a play on *Geant4* and the step function being the computationally heaviest part of the simulation.
+**G**eant4 **I**nference via **A**utoregressive **N**eural s**T**ep surrogate.
 
-Conditional generative surrogate for the Geant4 step function. Given a pre-step particle state, the model samples a physically plausible post-step outcome — the primary's continuation plus the variable-length list of secondary particles it produces — replacing the stochastic Geant4 physics engine with a trained generative model. A trained checkpoint autoregressively rolls out full showers, stepping each primary and pushing secondaries as new tracks.
+A conditional generative model that replaces the Geant4 step function: given a pre-step particle state it samples a post-step outcome — the primary's continuation plus its secondary particles — and autoregressively rolls that out into full showers. Trained entirely from parquet dumps of the miniCaloSim steps tree; no Geant4 runtime dependency.
 
-Training is driven entirely from parquet files of the miniCaloSim steps tree. No Geant4 runtime dependency.
+## Quick start
+
+```bash
+uv sync --extra cpu               # install deps (CPU torch; use --extra cuda for GPU)
+
+giant new-run --hidden-dim 512 --lr 3e-4          # scaffold config.toml + run dir
+giant train path/to/steps.parquet                 # train (flow + wgan by default)
+giant predict path/to/steps.parquet --checkpoint checkpoints/.../best.pt
+
+dwarf build-geometry-oracle path/to/steps.parquet --out oracle.pkl   # needed for rollout
+giant rollout path/to/steps.parquet --checkpoint checkpoints/.../best.pt --geometry oracle.pkl
+```
+
+Every command takes `--help` for the full flag list, and `--config config.toml` for anything not exposed as a flag.
 
 ## Architecture
 
-A **two-stage model**, both stages checkpointed together, with a choice of generative mode per stage (`--mode`):
+A **two-stage model**, checkpointed together. Either stage's outcome can be produced by one of three interchangeable generative objectives (`--stage1-generator`/`--stage2-generator`, or `--mode` to set both at once): `flow` (conditional flow matching, ODE-sampled in ~10 steps), `ddpm` (denoising diffusion), or `wgan` (single-pass WGAN-GP generator/critic).
 
-- **`flow`** (default) — conditional flow matching (Lipman et al. 2022): an MLP learns a vector field mapping noise → step outcomes, sampled via ODE integration in ~10 steps.
-- **`ddpm`** — a standard denoising diffusion baseline for comparison (`giant/model/schedule.py:CosineSchedule`).
-- **`wgan`** — a single-pass Wasserstein-GAN-GP generator/critic (`giant/model/wgan.py`), trading iterative sampling for one forward pass; implemented, not yet validated against the flow-matching baseline.
-
-**Stage 1 — primary (9D, `giant/constants.py:LOCAL_TARGET_NAMES`):**
+**Stage 1 — primary step.** Predicts the 9D post-step outcome (`giant/constants.py:LOCAL_TARGET_NAMES`) from the pre-step conditioning:
 
 | Index | Variable | Encoding |
 |-------|----------|----------|
@@ -23,36 +32,29 @@ A **two-stage model**, both stages checkpointed together, with a choice of gener
 | 3–5 | `post_dir` in local frame | unit vector |
 | 6–8 | `travel_dir` (`post_pos − pre_pos`) in local frame | unit vector |
 
-The two energy logits decode via softmax over `[edep_logit, sec_logit, 0]` × `pre_E`, so `edep + e_sec + post_E == pre_E` exactly — **energy conservation is built into the parametrization**, not left to the loss (`energy_simplex_decode`). Stage 1 also has a classifier head (`predict_n_sec`) predicting the number of secondaries `n_sec ∈ {0..K_MAX}` (`K_MAX = 15`) from the conditioning alone, no diffusion noise involved.
+- Energy logits decode via softmax over `[edep_logit, sec_logit, 0]` × `pre_E`, so `edep + e_sec + post_E == pre_E` exactly — conservation is architectural, not learned.
+- `post_dir`/`travel_dir` live in the frame where `pre_dir = ẑ`. `post_pos` isn't a target — it's reconstructed as `pre_pos + step_length * world_frame(travel_dir)`.
 
-Both `post_dir` and `travel_dir` are expressed in the coordinate frame where `pre_dir = ẑ`, making the scattering distribution nearly azimuthally symmetric. `post_pos` itself is not a raw target — it's reconstructed at inference as `pre_pos + step_length * world_frame(travel_dir)`, so the two stay consistent by construction instead of being learned (and potentially diverging) independently.
+**Stage 2 — secondaries.** Conditioned on the pre-step state and Stage 1's outcome, it generates the variable-length list of secondary particles. Two decoding strategies (`--stage2-decoder`):
 
-**Stage 2 — secondaries (`SecondaryDecoder`):** conditioned on the pre-step state *and* the Stage-1 outcome, a second net generates all `K_MAX` secondary slots at once — `(stick-breaking energy logit, local-frame direction, log-mass, charge)` per slot, ordered by descending energy; slots beyond the predicted `n_sec` are masked. Secondary energies are a stick-breaking partition of the `e_sec` budget from Stage 1, so the whole chain conserves energy. A secondary's mass/charge are regressed directly against its ground-truth PDG code's physical values (`giant.particles.particle_mass_charge`) and used as-is at inference — including for its own conditioning if it takes further steps in a rollout. No snapping to a known PDG code happens in the model path; `giant.particles.nearest_known_pdg` is a reporting-only lookup used to populate a nominal `pdg` label on output rows.
+- `autoregressive` — emits secondaries one at a time in descending-energy order, each token conditioned on a running history of prior tokens (`markov`: previous token only, or `attention`: causal self-attention, KV-cached at inference)
+- `one_shot` — all `K_MAX` slots generated in a single forward pass, masked past the predicted `n_sec`
 
-**Conditioning (`--conditioning`, per-checkpoint):** pre-step position, log(pre-energy), pre-step direction, layer ID, plus particle/material physical properties — mass/charge (`giant/particles.py`) and Z_eff/A_eff/density/X0/λ_int (`giant/materials.py`). Two mutually exclusive modes:
+Either way, secondary energies stick-break the `e_sec` budget handed down from Stage 1, so the full chain conserves energy. A secondary's particle identity is represented as `onehot` (categorical, top-N PDG codes + "other"), `physical` (continuous log-mass/charge), or `embedding` (nearest-neighbour lookup).
 
-- **`physical`** (default) — the physical-property columns are routed through small MLPs, computable for any PDG code / material, letting the surrogate generalize to species/materials outside the training menu.
-- **`embedding`** — the original design: a learned `nn.Embedding` per PDG code / material, kept as a generalization-comparison baseline (memorizes the training menu).
+**Conditioning.** Pre-step position/energy/direction/layer, plus particle mass/charge and material Z_eff/A_eff/density/X0/λ_int, encoded the same three ways as particle identity above (`--conditioning`) — the `physical` representation generalizes to species/materials outside the training menu since it's computed rather than looked up. `n_sec`/`e_sec` are always model outputs, never conditioning inputs.
 
-`n_sec` and `e_sec` are model outputs, not conditioning inputs — a rollout is self-contained and never injects ground truth.
-
-**Mixture-of-experts routing (`--router`, opt-in):** `giant/model/network.py` also implements a pluggable `Router` contract (`ROUTER_REGISTRY`: `energy`, `pdg`, `process`, plus a `composed` router combining several axes) that splits `DenoisingMLP`/`SecondaryDecoder` into per-expert trunks, soft-gated in training and top-1 dispatched at eval. Implemented; first rollout benchmark needs a retrain with a load-balancing loss and better-seeded router centers (see Roadmap). See `--router-type`/`--n-experts`/`--router-axis` on `giant train`/`giant new-run`.
-
-## Roadmap
-
-**Phase 1 (done):** `n_sec` and total secondary energy `e_sec` were conditioning inputs; the model predicted only the 9D primary post-step (energy-conservation PoC).
-
-**Phase 2 (implemented — baseline):** the two-stage model above predicts `n_sec` and each secondary's energy, direction, and species jointly with the primary, so a shower rollout is fully self-contained.
-
-**Physical-property conditioning (implemented):** replaces learned PDG/material embeddings with physical-property MLPs (see above); Stage 2 predicts a secondary's mass/charge directly instead of a snapped species embedding. Not yet done: the held-out-material/species generalization comparison against the `embedding` baseline — the natural dataset for that is the 34GB multi-material dataset at the repo root (6 materials, 237 PDG codes).
-
-**Faster-eval architectures (implemented, validation in progress):** both target a ~10× native-Geant4 eval budget. WGAN-GP (`--mode wgan`) has no rollout-vs-reference analysis run against it yet. The MoE router (`--router`) had its first rollout benchmark diverge from Geant4 despite matching bulk deposited energy — the experts weren't specializing (near-uniform gating), traced to a missing load-balance loss and a center-init that didn't match the real energy distribution; both are now fixable via `lambda_balance > 0` and quantile-seeded router centers, but a re-run to confirm hasn't happened yet.
-
-A multi-material sampling-calorimeter dataset is a planned future direction, not yet built.
+**MoE routing** (`--router`, either stage): a pluggable `Router` (`energy`/`pdg`/`process`/`composed` axes) top-1-dispatches each row to one of several small expert trunks at eval time, instead of running one monolithic trunk.
 
 ## Data
 
-Input: parquet files produced by [miniCaloSim](https://gitlab.etp.kit.edu/lbogner/minicalosim), or converted from a ROOT file via `dwarf convert`. Each row is one Geant4 step. Train/val split is by `event_id` (not row shuffle, and `--seed`-controlled) to avoid leaking correlated steps from the same shower.
+- Input: parquet files produced by [miniCaloSim](https://gitlab.etp.kit.edu/lbogner/minicalosim), or converted from ROOT via `dwarf convert`. One row = one Geant4 step.
+- **Conditioning (pre-step) columns:** `event_id`, `pdg`, `pre_x`/`pre_y`/`pre_z`, `pre_E`, `pre_dx`/`pre_dy`/`pre_dz` (direction), `material`, `layer_id`.
+- **Primary outcome (post-step) columns:** `post_x`/`post_y`/`post_z`, `post_E`, `post_dx`/`post_dy`/`post_dz`, `step_length`, `edep` (energy deposited in this step), `e_sec` (total energy carried off by secondaries), `child_track_ids` (its length gives `n_sec`).
+- **Secondary columns**, one variable-length list per step: `sec_pdg_list`, `sec_E_list`, `sec_dx_list`/`sec_dy_list`/`sec_dz_list` — padded/truncated to `K_MAX` (15) slots on load, ordered by descending energy.
+- **Optional:** `process` — the physics process that produced the step (e.g. `compt`, `phot`, `eBrem`); a post-step label used only as classifier supervision (`ProcessRouter`), never as conditioning.
+- Train/val split is by `event_id` (`--seed`-controlled), not row shuffle, so correlated steps from the same shower never leak across the split.
+- Loading a directory or `.manifest` of multiple parquet files (each one Geant4 job, `event_id` restarting from 0) offsets each file's `event_id`s by a fixed per-file stride so ids stay globally unique across files.
 
 ## Project structure
 
@@ -64,15 +66,20 @@ giant/
 │   │   ├── transforms.py   # log transforms, local-frame rotation, energy simplex, secondary encode/decode
 │   │   └── dataset.py      # StepsDataset / StreamingStepsDataset (PyTorch)
 │   ├── model/
-│   │   ├── network.py      # ConditionEncoder, DenoisingMLP, SecondaryDecoder, Router/MoE, WGAN generator/critic
+│   │   ├── network.py      # ConditionEncoder, Stage1Model, Stage2OneShot/Stage2Autoregressive, Router/MoE, CriticModel
 │   │   ├── schedule.py     # CosineSchedule (DDPM) and flow matching utilities
 │   │   └── wgan.py         # WGAN-GP gradient penalty / critic / generator losses
 │   ├── constants.py        # output/conditioning dims, K_MAX, secondary slot layout, schema keys
-│   ├── particles.py        # PDG → (mass, charge) decode, incl. nuclear/ion codes; nearest-known-PDG lookup
+│   ├── particles.py        # PDG → (mass, charge) decode, incl. nuclear/ion codes; onehot/embedding secondary-identity decode
 │   ├── materials.py        # material name → (Z_eff, A_eff, density, X0, λ_int)
 │   ├── config.py           # default hyperparameters, TOML config merging, device autodetect
 │   ├── pipeline.py         # builds datasets/normalizers and kicks off a training run (with setup-stage caching)
-│   ├── train.py            # two-stage training loop, checkpointing, graceful shutdown, W&B logging
+│   ├── training/           # two-stage training: loop, per-stage trainers, metrics, checkpointing
+│   │   ├── loop.py         #   epoch loop, graceful shutdown, best-checkpoint selection
+│   │   ├── trainers.py     #   StageSpec + flow/ddpm and WGAN-GP per-stage trainers
+│   │   ├── stage2_inputs.py#   ground-truth stage-2 targets + autoregressive/teacher-forcing inputs
+│   │   ├── metrics.py      #   MetricsCollector: metrics.csv columns, W&B logging, progress/summary
+│   │   └── checkpoint.py   #   checkpoint assembly/restore (format unchanged since v0.2)
 │   ├── sample.py           # DDPM / DDIM / flow matching / WGAN samplers + secondary sampling
 │   ├── geometry.py         # GeometryOracle: position → (material, layer_id, escaped) for rollout
 │   ├── rollout.py          # autoregressive shower rollout driver
@@ -86,7 +93,7 @@ giant/
 │   │   ├── condor.py       #   prep / compute-one / submit-description plumbing
 │   │   └── render.py       #   PDFs + HTML gallery (only module importing plotstyle/LaTeX)
 │   └── cli.py              # `giant train` / `new-run` / `predict` / `rollout` / `analyze` Typer app
-├── scripts/                            # dataset/tooling logic, unified under the `dwarf` CLI (`dwarf --help`)
+├── giant/tools/                        # dataset/tooling logic, unified under the `dwarf` CLI (`dwarf --help`)
 │   ├── dwarf.py                        # Typer app: convert, migrate, bump-gen, bump-schema, status,
 │   │                                    #   update-manifest, create-manifest, make-root,
 │   │                                    #   build-geometry-oracle, warm-cache, hparam-scan
@@ -106,39 +113,48 @@ giant/
 ## Setup
 
 ```bash
-uv sync --extra cpu               # CPU-only torch (use --extra cuda for CUDA 11.8 instead)
-uv sync --extra cpu --extra dev   # add dev tools (pytest, ruff, ty)
+uv sync --extra cpu                   # CPU-only torch (use --extra cuda for CUDA 11.8 instead)
+uv sync --extra cpu --extra dev       # add dev tools (pytest, ruff, ty)
 uv sync --extra cpu --extra geometry  # add scikit-learn, for `dwarf build-geometry-oracle` / rollout
 ```
 
-`cpu` and `cuda` are mutually exclusive extras selecting the torch build (pinned to 2.3.x); plain `uv sync` installs no torch at all. See `CLAUDE.md` for details.
+`cpu` and `cuda` are mutually exclusive — pick one to select the torch build (pinned to 2.3.x). Plain `uv sync` installs no torch at all. See `CLAUDE.md` for details.
 
-## Training, prediction, and rollout
+## Training, prediction, rollout
 
 ```bash
-giant new-run --hidden-dim 512 --lr 3e-4 --comment "..."   # scaffold a config.toml + run dir for a new run
-giant train path/to/steps.parquet --mode flow               # train (flow matching; also --mode ddpm / wgan)
+giant new-run --hidden-dim 512 --lr 3e-4 --comment "..."   # scaffold a config.toml + run dir
+giant train path/to/steps.parquet                          # train (flow stage 1 + wgan stage 2, default)
 giant predict path/to/steps.parquet --checkpoint checkpoints/.../best.pt
 
-# Full-shower rollout needs a geometry oracle (position → material/layer_id):
-dwarf build-geometry-oracle path/to/steps.parquet --out oracle.pkl
+dwarf build-geometry-oracle path/to/steps.parquet --out oracle.pkl   # position → material/layer_id
 giant rollout path/to/steps.parquet --checkpoint checkpoints/.../best.pt --geometry oracle.pkl
 ```
 
-`train`/`predict` accept a TOML config file (`--config`) and CLI overrides for hyperparameters; see `--help` on any command for the full option list. `giant train --wandb` logs per-epoch metrics (the same ones written to `metrics.csv`) to Weights & Biases; requires `uv sync --extra wandb`. A repeat `giant train` against the same dataset (e.g. a hyperparameter sweep) reuses a cached setup-stage sidecar (vocab maps, event split, normalizer stats) unless `--no-cache-setup`/`--rebuild-setup-cache`; `dwarf warm-cache` precomputes it ahead of time. `giant rollout` seeds showers from the highest-energy entry step per event, then autoregressively steps the two-stage model to completion — pushing secondaries as new tracks and looking up `material`/`layer_id` from the oracle at each step. Tracks terminate on energy cutoff, per-track max steps, detector escape, or natural end; energy is deposited locally on every stop except escape (leakage), so showers conserve energy by construction.
+Useful flags on `giant train`:
+
+- `--mode {flow,ddpm,wgan}` sets both stages' objective at once; `--stage1-generator`/`--stage2-generator` override per stage
+- `--stage2-decoder {autoregressive,one_shot}` — Stage 2 decoding strategy (see Architecture)
+- `--conditioning {physical,embedding,onehot}` — conditioning representation
+- `--router` / `--router-type` / `--n-experts` / `--router-axis` — MoE routing
+- `--wandb` — log per-epoch metrics to Weights & Biases (needs `uv sync --extra wandb`); metric names are `<stage>/<split>/<metric>` plus an unprefixed run-level tail, all derived from `giant/training/trainers.py` `MetricSpec`s
+- `--no-cache-setup` / `--rebuild-setup-cache` — control the setup-stage sidecar cache (vocab maps, event split, normalizer stats); `dwarf warm-cache` precomputes it
+
+Config-file-only knobs (no CLI flag — use `--config config.toml`): `stage2_model.autoregressive.teacher_forcing`/`.history`, `stage2_model.particle_type.target`. v0.2 flat-schema configs and checkpoints load fine (auto-migrated).
+
+`giant rollout` seeds showers from each event's highest-energy entry step, then autoregressively steps the model to completion, pushing secondaries as new tracks and looking up `material`/`layer_id` from the geometry oracle each step. Tracks terminate on energy cutoff, max steps, detector escape, or natural end; energy is deposited locally on every stop except escape, so showers conserve energy by construction.
 
 ## Validation and analysis
 
-`giant.validate.validate_marginals` runs step-level marginal and KL-divergence checks during training (`--validate-every`).
-
-For deeper rollout-vs-reference diagnostics — marginals stratified by energy/pdg/material, per-event totals, shower profiles, species share, leakage, and secondaries — `giant analyze` runs a streaming compute/render pipeline against a `giant rollout` YAML sidecar:
+- `giant.validate.validate_marginals` — step-level marginal + KL-divergence checks during training (`--validate-every`)
+- `giant analyze` — deeper rollout-vs-reference diagnostics (marginals by energy/pdg/material, per-event totals, shower profiles, species share, leakage, secondaries):
 
 ```bash
 giant analyze submit rollout.yaml --accounting-group cms   # prep + one HTCondor job per plot (compute only)
 giant analyze render <run_dir> --gallery                   # local: styled PDFs + HTML gallery (needs LaTeX)
 ```
 
-`<run_dir>` is derived next to the rollout parquet (`analyze prep`/`submit` print it). Compute jobs are polars/numpy only; only `render` imports plotstyle/LaTeX, so it always runs locally.
+`<run_dir>` is derived next to the rollout parquet (`analyze prep`/`submit` print it). Compute jobs are polars/numpy only; only `render` needs LaTeX, so it always runs locally.
 
 ## Development
 

giant/_migration.py

@@ -0,0 +1,70 @@
+"""Shared v0.2 -> v0.3 migration knowledge.
+
+v0.3.0 broke the config format (single `[train]` + `[model]` -> `[conditioning]`/
+`[stage1_model]`/`[stage2_model]`/`[train]`), and that break has to be absorbed by two
+independent migration surfaces: `giant.config.migrate_config` (a v0.2 `config.toml`) and
+`giant.model.network._migrate_legacy_model_config` (a v0.2 checkpoint's flat
+`model_config` dict). Both translate the same v0.2 facts into the same v0.3 shape, so
+the facts live here once rather than as two hand-maintained copies — see issues.md
+Issue 6.
+
+A dependency-free leaf module so neither `config.py` nor `network.py` has to import the
+other to share this.
+"""
+
+# v0.2 model-shaped keys (config.toml's [model] table, or a checkpoint's flat
+# model_config dict — same key names in both) applied identically to both v0.3 stage
+# blocks, because v0.2 had only one trunk shape shared by both stages.
+V02_MODEL_KEY_TO_STAGES: tuple[tuple[str, str], ...] = (
+    ("hidden_dim", "hidden_dim"),
+    ("n_blocks", "n_res_blocks"),
+    ("dropout", "dropout"),
+)
+
+# v0.2 architectural facts that had no corresponding config key at all — always true of
+# a v0.2 model, so both migration surfaces inject them unconditionally. Keyed by dotted
+# path relative to the migrated dict's root. NOTE: conditioning.*.n_layers (2) differs
+# from the v0.3 *default* (1) — not a typo, v0.2's conditioning MLP was always 2 layers
+# deep.
+V02_FIXED_FACTS: dict[str, object] = {
+    "conditioning.out_dim": 128,
+    "conditioning.particle.n_layers": 2,
+    "conditioning.material.n_layers": 2,
+    "stage1_model.active": True,
+    "stage1_model.flow.time_dim": 64,
+    "stage1_model.ddpm.time_dim": 64,
+    "stage2_model.active": True,
+    "stage2_model.flow.time_dim": 64,
+    "stage2_model.ddpm.time_dim": 64,
+    "stage2_model.context_dim": 64,
+    "stage2_model.decoder": "one_shot",
+    "stage2_model.particle_type.target": "physical",
+}
+
+
+def reject_legacy_router_expert_sizing(router_cfg: dict, *, source: str) -> None:
+    """Pop and validate v0.2's per-expert width/depth override, in place.
+
+    v0.3.0 removed per-expert sizing — experts always inherit the stage's
+    hidden_dim/n_res_blocks — so a v0.2 router config/checkpoint that set a non-default
+    `expert_hidden_dim`/`expert_n_blocks` describes experts with a different width/depth
+    than the monolith, and can only be reproduced by v0.2 code. Silently dropping these
+    keys (a router builder's kwarg filtering would do this for free) would resize the
+    experts instead of refusing, so this raises loudly.
+
+    Always pops both keys, whether or not they were non-default, so callers can go on
+    to use the (now-cleaned) `router_cfg` unconditionally. `source` names what's being
+    migrated (e.g. "v0.2 config's model.router" or "this checkpoint's
+    model_config.router") for the error message.
+    """
+    expert_hidden_dim = router_cfg.pop("expert_hidden_dim", 0)
+    expert_n_blocks = router_cfg.pop("expert_n_blocks", 0)
+    if not (expert_hidden_dim or expert_n_blocks):
+        return
+    raise ValueError(
+        f"{source} sets expert_hidden_dim/expert_n_blocks to a non-default value "
+        f"({expert_hidden_dim!r}, {expert_n_blocks!r}); v0.3.0 removed per-expert "
+        "sizing (experts always inherit the stage's hidden_dim/n_res_blocks), so "
+        "this router's experts have a different width/depth than the monolith. "
+        "This checkpoint/config can only be loaded by v0.2 code."
+    )

giant/analysis/catalog.py

@@ -58,6 +58,7 @@ from giant.analysis.router_gating import (
     compute_router_share_by_process,
 )
 from giant.analysis.sources import Side, open_side, physical_steps, secondaries
+from giant.analysis.type_embedding_distance import compute_type_embedding_l1_distance
 from giant.analysis.variables import RANGED_VARS, cos_scatter_expr
 
 
@@ -71,6 +72,11 @@ class Bundle:
     r_phys: pl.LazyFrame  # rollout, physical steps only
     t_phys: pl.LazyFrame  # reference, physical steps only
     checkpoint: str | None = None  # from the rollout YAML; router_gating only
+    # Diagnostic pre-aggregated at rollout time (giant.rollout.
+    # L1DistCollector.summary()) — from the rollout YAML, type_embedding_l1_distance
+    # only. Unlike checkpoint/router_gating, this needs no live model: it's
+    # already a finished histogram, just passed through.
+    type_embedding_l1_dist: dict | None = None
 
     @classmethod
     def open(
@@ -80,6 +86,7 @@ class Bundle:
         ctx: Context,
         checkpoint=None,
         chunk: tuple[int, int] | None = None,
+        type_embedding_l1_dist: dict | None = None,
     ) -> "Bundle":
         """Open both sides, optionally restricted to one event-disjoint chunk.
 
@@ -103,6 +110,7 @@ class Bundle:
             r_phys=physical_steps(r_all, Side.rollout),
             t_phys=physical_steps(t_all, Side.reference),
             checkpoint=checkpoint,
+            type_embedding_l1_dist=type_embedding_l1_dist,
         )
 
 
@@ -158,9 +166,7 @@ def _finalize_counts(merged: dict[str, list], key, nbins: int) -> list[int]:
     return list(merged.get(str(key), [0] * nbins))
 
 
-def _np_hist_pair(
-    r: np.ndarray, t: np.ndarray, nbins: int
-) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+def _np_hist_pair(r: np.ndarray, t: np.ndarray, nbins: int) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
     """Shared-edge histogram of two small per-event arrays (robust range)."""
     both = np.concatenate([r, t]) if (len(r) or len(t)) else np.array([0.0, 1.0])
     lo, hi = float(np.quantile(both, 0.001)), float(np.quantile(both, 0.999))
@@ -232,9 +238,7 @@ def _marginal_overall_finalize(parts: list[dict], ctx: Context, var: str) -> Red
 
 def _energy_group_expr(lf: pl.LazyFrame, edges: np.ndarray) -> pl.Expr:
     ids, bins = event_energy_bins(lf, edges)
-    return pl.col("event_id").replace_strict(
-        ids, bins, default=-1, return_dtype=pl.Int64
-    )
+    return pl.col("event_id").replace_strict(ids, bins, default=-1, return_dtype=pl.Int64)
 
 
 def _marginal_grouped_partial(b: Bundle, var: str, axis: str) -> dict:
@@ -257,9 +261,7 @@ def _marginal_grouped_partial(b: Bundle, var: str, axis: str) -> dict:
     }
 
 
-def _marginal_grouped_finalize(
-    parts: list[dict], ctx: Context, var: str, axis: str
-) -> Reduced:
+def _marginal_grouped_finalize(parts: list[dict], ctx: Context, var: str, axis: str) -> Reduced:
     label, _ = _var(var)
     edges = _marginal_edges(ctx, var)
     nb = len(edges) - 1
@@ -309,9 +311,7 @@ def _event_scalar_partial(b: Bundle, col: str, use_all: bool) -> dict:
     return {"r": r.tolist(), "t": t.tolist()}
 
 
-def _event_scalar_finalize(
-    parts: list[dict], ctx: Context, spec_id: str, title: str, xlabel: str
-) -> Reduced:
+def _event_scalar_finalize(parts: list[dict], ctx: Context, spec_id: str, title: str, xlabel: str) -> Reduced:
     r = np.concatenate([np.asarray(p["r"], dtype=float) for p in parts])
     t = np.concatenate([np.asarray(p["t"], dtype=float) for p in parts])
     edges, rc, tc = _np_hist_pair(r, t, ctx.n_marginal_bins)
@@ -480,9 +480,7 @@ def _leakage_partial(b: Bundle) -> dict:
 
 def _leakage_finalize(parts: list[dict], ctx: Context) -> Reduced:
     frac = np.concatenate([np.asarray(p["frac"], dtype=float) for p in parts])
-    edges = np.linspace(
-        0.0, max(float(frac.max()) if len(frac) else 1.0, 1e-3), ctx.n_marginal_bins + 1
-    )
+    edges = np.linspace(0.0, max(float(frac.max()) if len(frac) else 1.0, 1e-3), ctx.n_marginal_bins + 1)
     counts = np.histogram(frac, edges)[0]
     return Reduced(
         id="leakage_fraction",
@@ -513,18 +511,8 @@ def _sec_frames(b: Bundle):
 
 def _sec_count_per_event_partial(b: Bundle) -> dict:
     r_sec, t_sec = _sec_frames(b)
-    r = (
-        r_sec.group_by("event_id")
-        .agg(pl.len().alias("n"))
-        .collect(engine="streaming")["n"]
-        .to_numpy()
-    )
-    t = (
-        t_sec.group_by("event_id")
-        .agg(pl.len().alias("n"))
-        .collect(engine="streaming")["n"]
-        .to_numpy()
-    )
+    r = r_sec.group_by("event_id").agg(pl.len().alias("n")).collect(engine="streaming")["n"].to_numpy()
+    t = t_sec.group_by("event_id").agg(pl.len().alias("n")).collect(engine="streaming")["n"].to_numpy()
     return {"r": r.tolist(), "t": t.tolist()}
 
 
@@ -560,9 +548,7 @@ def _sec_count_per_species_partial(b: Bundle) -> dict:
 def _sec_count_per_species_finalize(parts: list[dict], ctx: Context) -> Reduced:
     r = sum_merge([p["r"] for p in parts])
     t = sum_merge([p["t"] for p in parts])
-    keys = sorted(set(r) | set(t), key=lambda k: -(r.get(k, 0) + t.get(k, 0)))[
-        : len(ctx.top_pdgs)
-    ]
+    keys = sorted(set(r) | set(t), key=lambda k: -(r.get(k, 0) + t.get(k, 0)))[: len(ctx.top_pdgs)]
     return Reduced(
         id="sec_count_per_species",
         family="secondaries",
@@ -609,11 +595,9 @@ def _sec_energy_finalize(parts: list[dict], ctx: Context) -> Reduced:
 
 def _sec_cos_angle_partial(b: Bundle) -> dict:
     edges = np.linspace(-1.0, 1.0, b.ctx.n_sec_bins + 1)
-    cos = (
-        pl.col("sdx") * pl.col("axis_x")
-        + pl.col("sdy") * pl.col("axis_y")
-        + pl.col("sdz") * pl.col("axis_z")
-    ).clip(-1.0, 1.0)
+    cos = (pl.col("sdx") * pl.col("axis_x") + pl.col("sdy") * pl.col("axis_y") + pl.col("sdz") * pl.col("axis_z")).clip(
+        -1.0, 1.0
+    )
 
     def _side(sec_lf: pl.LazyFrame, steps_lf: pl.LazyFrame) -> dict[str, list[int]]:
         ea = entry_axis(steps_lf)
@@ -651,13 +635,14 @@ _router_gating_partial, _router_gating_finalize = _unchunkable(
     lambda b: compute_router_gating(b.checkpoint, b.r_phys, b.t_phys)
 )
 _router_share_pdg_partial, _router_share_pdg_finalize = _unchunkable(
-    lambda b: compute_router_share_by_pdg(
-        b.checkpoint, b.r_phys, b.t_phys, b.ctx.top_pdgs
-    )
+    lambda b: compute_router_share_by_pdg(b.checkpoint, b.r_phys, b.t_phys, b.ctx.top_pdgs)
 )
 _router_share_process_partial, _router_share_process_finalize = _unchunkable(
     lambda b: compute_router_share_by_process(b.checkpoint, b.t_phys)
 )
+_type_embedding_l1_distance_partial, _type_embedding_l1_distance_finalize = _unchunkable(
+    lambda b: compute_type_embedding_l1_distance(b.type_embedding_l1_dist)
+)
 
 
 # ---------------------------------------------------------------------------
@@ -678,9 +663,7 @@ def build_catalog() -> list[PlotSpec]:
                 f"marginal_{var}",
                 "marginals",
                 compute_partial=lambda b, v=var: _marginal_overall_partial(b, v),
-                finalize=lambda parts, ctx, v=var: _marginal_overall_finalize(
-                    parts, ctx, v
-                ),
+                finalize=lambda parts, ctx, v=var: _marginal_overall_finalize(parts, ctx, v),
             )
         )
         for axis in GROUPING_AXES:
@@ -688,12 +671,8 @@ def build_catalog() -> list[PlotSpec]:
                 PlotSpec(
                     f"marginal_{var}_by_{axis}",
                     "marginals",
-                    compute_partial=lambda b, v=var, a=axis: _marginal_grouped_partial(
-                        b, v, a
-                    ),
-                    finalize=lambda parts, ctx, v=var, a=axis: (
-                        _marginal_grouped_finalize(parts, ctx, v, a)
-                    ),
+                    compute_partial=lambda b, v=var, a=axis: _marginal_grouped_partial(b, v, a),
+                    finalize=lambda parts, ctx, v=var, a=axis: _marginal_grouped_finalize(parts, ctx, v, a),
                 )
             )
 
@@ -701,9 +680,7 @@ def build_catalog() -> list[PlotSpec]:
         PlotSpec(
             "event_total_edep",
             "event",
-            compute_partial=lambda b: _event_scalar_partial(
-                b, "total_edep", use_all=True
-            ),
+            compute_partial=lambda b: _event_scalar_partial(b, "total_edep", use_all=True),
             finalize=lambda parts, ctx: _event_scalar_finalize(
                 parts,
                 ctx,
@@ -721,9 +698,7 @@ def build_catalog() -> list[PlotSpec]:
         PlotSpec(
             "event_mean_length",
             "event",
-            compute_partial=lambda b: _event_scalar_partial(
-                b, "mean_length", use_all=False
-            ),
+            compute_partial=lambda b: _event_scalar_partial(b, "mean_length", use_all=False),
             finalize=lambda parts, ctx: _event_scalar_finalize(
                 parts,
                 ctx,
@@ -735,9 +710,7 @@ def build_catalog() -> list[PlotSpec]:
         PlotSpec(
             "event_n_steps",
             "event",
-            compute_partial=lambda b: _event_scalar_partial(
-                b, "n_steps", use_all=False
-            ),
+            compute_partial=lambda b: _event_scalar_partial(b, "n_steps", use_all=False),
             finalize=lambda parts, ctx: _event_scalar_finalize(
                 parts,
                 ctx,
@@ -762,9 +735,7 @@ def build_catalog() -> list[PlotSpec]:
         PlotSpec(
             "shower_transverse",
             "shower",
-            compute_partial=lambda b: _profile_partial(
-                b, transverse_expr, "transverse_edges"
-            ),
+            compute_partial=lambda b: _profile_partial(b, transverse_expr, "transverse_edges"),
             finalize=lambda parts, ctx: _profile_finalize(
                 parts,
                 ctx,
@@ -831,6 +802,13 @@ def build_catalog() -> list[PlotSpec]:
             finalize=_router_share_process_finalize,
             chunkable=False,
         ),
+        PlotSpec(
+            "type_embedding_l1_distance",
+            "model",
+            compute_partial=_type_embedding_l1_distance_partial,
+            finalize=_type_embedding_l1_distance_finalize,
+            chunkable=False,
+        ),
     ]
     return specs
 

giant/analysis/condor.py

@@ -83,6 +83,12 @@ _PLOT_META_KEYS = (
     "best_val_loss",
     "training_config",
     "training_meta",
+    # Diagnostic — only present when giant rollout ran under
+    # stage2_model.particle_type.target="embedding" (see giant/cli.py's
+    # rollout command and giant.rollout.L1DistCollector); absent otherwise,
+    # which the type_embedding_l1_distance PlotSpec (catalog.py) reads as
+    # "not applicable to this checkpoint".
+    "type_embedding_l1_dist",
 )
 
 
@@ -155,9 +161,7 @@ class RunMeta:
         return cls(**json.loads(Path(path).read_text()))
 
 
-def _rows_per_chunk(
-    rollout: str | Path, reference: str | Path, n_chunks: int
-) -> list[int]:
+def _rows_per_chunk(rollout: str | Path, reference: str | Path, n_chunks: int) -> list[int]:
     """Rollout+reference row count of each ``event_id % n_chunks`` chunk.
 
     One cheap streaming ``group_by`` per side (just the ``event_id`` column) —
@@ -247,6 +251,7 @@ def compute_reduced(
     checkpoint: str | None = None,
     chunk_index: int = 0,
     n_chunks: int = 1,
+    type_embedding_l1_dist: dict | None = None,
 ) -> Path:
     """Core: run one (plot, chunk)'s partial reduction against explicit paths.
 
@@ -261,11 +266,15 @@ def compute_reduced(
     effective_n = n_chunks if spec.chunkable else 1
     if not (0 <= chunk_index < effective_n):
         raise ValueError(
-            f"{spec_id}: chunk_index={chunk_index} out of range for "
-            f"n_chunks={effective_n} (chunkable={spec.chunkable})"
+            f"{spec_id}: chunk_index={chunk_index} out of range for n_chunks={effective_n} (chunkable={spec.chunkable})"
         )
     bundle = Bundle.open(
-        rollout, reference, ctx, checkpoint=checkpoint, chunk=(chunk_index, effective_n)
+        rollout,
+        reference,
+        ctx,
+        checkpoint=checkpoint,
+        chunk=(chunk_index, effective_n),
+        type_embedding_l1_dist=type_embedding_l1_dist,
     )
     partial = Partial(
         id=spec_id,
@@ -291,6 +300,7 @@ def compute_one(spec_id: str, run_dir: str | Path, chunk_index: int = 0) -> Path
         checkpoint=meta.plot_meta.get("checkpoint"),
         chunk_index=chunk_index,
         n_chunks=meta.n_chunks,
+        type_embedding_l1_dist=meta.plot_meta.get("type_embedding_l1_dist"),
     )
 
 
@@ -314,10 +324,7 @@ def merge_one(spec_id: str, run_dir: str | Path) -> Path:
     effective_n = meta.n_chunks if spec.chunkable else 1
 
     partial_dir = run_path / "reduced_partial"
-    found = {
-        p.chunk: p
-        for p in (Partial.load(jf) for jf in partial_dir.glob(f"{spec_id}__*.json"))
-    }
+    found = {p.chunk: p for p in (Partial.load(jf) for jf in partial_dir.glob(f"{spec_id}__*.json"))}
     missing = sorted(set(range(effective_n)) - set(found))
     if missing:
         raise FileNotFoundError(
@@ -362,11 +369,7 @@ exec {giant_exe} analyze compute-one --id "$1" --chunk "$2" --run-dir {run_dir}
 
 
 def _submit_description(cfg: SubmitConfig, wrapper: Path, jobs_file: Path) -> str:
-    reqs_attrs = (
-        "+RemoteJob = True\n"
-        if cfg.remote
-        else "requirements = TARGET.ProvidesETPResources\n"
-    )
+    reqs_attrs = "+RemoteJob = True\n" if cfg.remote else "requirements = TARGET.ProvidesETPResources\n"
     return (
         "universe = docker\n"
         f"docker_image = {cfg.docker_image}\n"
@@ -386,9 +389,7 @@ def _submit_description(cfg: SubmitConfig, wrapper: Path, jobs_file: Path) -> st
     )
 
 
-def _job_walltimes(
-    run_dir: Path, ids: list[str], n_chunks: int
-) -> list[tuple[str, int, int]]:
+def _job_walltimes(run_dir: Path, ids: list[str], n_chunks: int) -> list[tuple[str, int, int]]:
     """``(spec_id, chunk, walltime_s)`` for every job, sized from ``run_meta.json``.
 
     Row counts come from ``prep``'s ``RunMeta.rows_per_chunk``/``total_rows``;
@@ -464,9 +465,7 @@ def write_submit(cfg: SubmitConfig, ids: list[str] | None = None) -> Path:
     (run_dir / "reduced_partial").mkdir(parents=True, exist_ok=True)
 
     wrapper = run_dir / "run_compute.sh"
-    wrapper.write_text(
-        _WRAPPER.format(repo_dir=cfg.repo_dir, giant_exe=giant_exe, run_dir=run_dir)
-    )
+    wrapper.write_text(_WRAPPER.format(repo_dir=cfg.repo_dir, giant_exe=giant_exe, run_dir=run_dir))
     wrapper.chmod(0o755)
 
     jobs = _job_walltimes(run_dir, ids, cfg.n_chunks)

giant/analysis/context.py

@@ -74,9 +74,7 @@ def _row_subsample(lf: pl.LazyFrame, sample_rows: int, seed: int) -> pl.LazyFram
     return lf.filter((pl.col("pre_E").hash(seed=seed) % 2**32) < threshold)
 
 
-def _combined_quantiles(
-    r_vals: np.ndarray, t_vals: np.ndarray, lo_q: float, hi_q: float
-) -> tuple[float, float]:
+def _combined_quantiles(r_vals: np.ndarray, t_vals: np.ndarray, lo_q: float, hi_q: float) -> tuple[float, float]:
     """Robust (lo_q, hi_q) range over the union of two value samples."""
     both = np.concatenate([r_vals, t_vals])
     lo, hi = float(np.quantile(both, lo_q)), float(np.quantile(both, hi_q))
@@ -104,31 +102,15 @@ def build_context(
 
     # Ranged marginal variables: robust ranges over a shared row subsample.
     exprs = [e.alias(n) for n, (_, e) in RANGED_VARS.items()]
-    r_s = (
-        _row_subsample(r_lf, sample_rows, seed)
-        .select(exprs)
-        .collect(engine="streaming")
-    )
-    t_s = (
-        _row_subsample(t_lf, sample_rows, seed)
-        .select(exprs)
-        .collect(engine="streaming")
-    )
+    r_s = _row_subsample(r_lf, sample_rows, seed).select(exprs).collect(engine="streaming")
+    t_s = _row_subsample(t_lf, sample_rows, seed).select(exprs).collect(engine="streaming")
     var_ranges = {
-        name: _combined_quantiles(
-            r_s[name].to_numpy(), t_s[name].to_numpy(), _LO_Q, _HI_Q
-        )
-        for name in RANGED_VARS
+        name: _combined_quantiles(r_s[name].to_numpy(), t_s[name].to_numpy(), _LO_Q, _HI_Q) for name in RANGED_VARS
     }
 
     # Energy-bin edges from exact per-event incident energies (cheap group_by).
     def _incident(lf: pl.LazyFrame) -> np.ndarray:
-        return (
-            lf.group_by("event_id")
-            .agg(pl.col("pre_E").max())
-            .collect(engine="streaming")["pre_E"]
-            .to_numpy()
-        )
+        return lf.group_by("event_id").agg(pl.col("pre_E").max()).collect(engine="streaming")["pre_E"].to_numpy()
 
     r_inc, t_inc = _incident(r_lf), _incident(t_lf)
     energy_edges = energy_bin_edges(np.concatenate([r_inc, t_inc]), n_energy_bins)
@@ -145,8 +127,7 @@ def build_context(
     )
     top_pdgs = [int(x) for x in pdg_counts["pdg"].to_list()[:top_k_pdg]]
     materials = sorted(
-        set(_counts(r_lf, "material")["material"].to_list())
-        | set(_counts(t_lf, "material")["material"].to_list())
+        set(_counts(r_lf, "material")["material"].to_list()) | set(_counts(t_lf, "material")["material"].to_list())
     )
 
     # Shower depth / transverse ranges from a subsampled proxy.

giant/analysis/grouping.py

@@ -68,9 +68,7 @@ def energy_bin_edges(incident_E: np.ndarray, n_bins: int = 4) -> np.ndarray:
 
 def energy_bin_labels(edges: np.ndarray) -> list[str]:
     """``E in [lo, hi)`` labels for each bin defined by ``edges`` (MeV)."""
-    return [
-        f"E in [{edges[i]:.3g}, {edges[i + 1]:.3g}) MeV" for i in range(len(edges) - 1)
-    ]
+    return [f"E in [{edges[i]:.3g}, {edges[i + 1]:.3g}) MeV" for i in range(len(edges) - 1)]
 
 
 def digitize_expr(value: pl.Expr, edges: np.ndarray) -> pl.Expr:
@@ -88,9 +86,7 @@ def digitize_expr(value: pl.Expr, edges: np.ndarray) -> pl.Expr:
     return idx.clip(0, n_bins - 1)
 
 
-def event_energy_bins(
-    lf: pl.LazyFrame, edges: np.ndarray
-) -> tuple[np.ndarray, np.ndarray]:
+def event_energy_bins(lf: pl.LazyFrame, edges: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
     """Per-event incident-energy bin: ``(event_ids, bin_idx)`` numpy arrays.
 
     Incident energy is ``max(pre_E)`` per event (the primary). One bounded

giant/analysis/reduce.py

@@ -142,9 +142,7 @@ def attach_entry_axis(lf: pl.LazyFrame, entry: pl.DataFrame) -> pl.LazyFrame:
     """
     ids = entry["event_id"].to_numpy()
     return lf.with_columns(
-        pl.col("event_id")
-        .replace_strict(ids, entry[col].to_numpy(), return_dtype=pl.Float64)
-        .alias(col)
+        pl.col("event_id").replace_strict(ids, entry[col].to_numpy(), return_dtype=pl.Float64).alias(col)
         for col in _ENTRY_AXIS_COLS
     )
 
@@ -254,10 +252,7 @@ def leakage_fraction(lf: pl.LazyFrame) -> np.ndarray:
         lf.group_by("event_id")
         .agg(
             pl.col("edep").sum().alias("deposited"),
-            pl.col("pre_E")
-            .filter(pl.col("termination_reason") == TERM_ESCAPED)
-            .sum()
-            .alias("escaped"),
+            pl.col("pre_E").filter(pl.col("termination_reason") == TERM_ESCAPED).sum().alias("escaped"),
         )
         .collect(engine="streaming")
     )

giant/analysis/render.py

@@ -41,11 +41,47 @@ def _overlay(ax, edges: np.ndarray, series: dict[str, list], log_y: bool) -> Non
         ax.set_yscale("log")
 
 
-def _router_summary(model_config: dict) -> str:
-    r = model_config.get("router") or {}
-    if not r.get("enabled"):
+def _router_summary(router_cfg: dict) -> str:
+    if not router_cfg.get("enabled"):
         return "off"
-    return f"{r.get('type', '?')}×{r.get('n_experts', '?')}"
+    return f"{router_cfg.get('type', '?')}×{router_cfg.get('n_experts', '?')}"
+
+
+def _figure_params_v2(mc: dict, run_meta: dict) -> dict:
+    """`_figure_params` for a new-shape (nested) `model_config` — has a
+    `stage1_model` key. Reports stage 1's architecture (the headline
+    generator); stage 2's generator is only added (`mode_s2`) when it
+    differs from stage 1's, since a mixed run (the `stage1=flow` +
+    `stage2=wgan` case) is the interesting exception, not
+    the common case."""
+    s1 = mc["stage1_model"]
+    s2 = mc.get("stage2_model") or {}
+    mode = s1.get("generator")
+    params: dict = {}
+    if s1.get("hidden_dim") is not None:
+        params["hidden_dim"] = s1["hidden_dim"]
+    if s1.get("n_res_blocks") is not None:
+        params["n_res_blocks"] = s1["n_res_blocks"]
+    if mode is not None:
+        params["mode"] = mode
+    s2_mode = s2.get("generator")
+    if s2_mode is not None and s2_mode != mode:
+        params["mode_s2"] = s2_mode
+    particle_type = ((mc.get("conditioning") or {}).get("particle") or {}).get("type")
+    if particle_type is not None:
+        params["conditioning"] = particle_type
+    params["router"] = _router_summary(s1.get("router") or {})
+    if run_meta.get("training_epoch") is not None:
+        params["epoch"] = run_meta["training_epoch"]
+    if run_meta.get("best_val_loss") is not None:
+        params["best_val_loss"] = round(run_meta["best_val_loss"], 4)
+    if mode == "wgan":
+        noise_dim = (s1.get("wgan") or {}).get("noise_dim")
+        if noise_dim is not None:
+            params["noise_dim"] = noise_dim
+    elif run_meta.get("steps") is not None:
+        params["steps"] = run_meta["steps"]
+    return params
 
 
 def _figure_params(run_meta: dict) -> dict:
@@ -59,8 +95,14 @@ def _figure_params(run_meta: dict) -> dict:
     architecture-conditional: flow/ddpm runs show the ODE ``steps`` used for
     this rollout, wgan runs show ``noise_dim`` instead since wgan sampling is
     single-pass and has no ODE step count.
+
+    Handles both a v0.2 checkpoint's flat ``model_config`` and a v0.3.0
+    nested one (has a ``stage1_model`` key — see ``_figure_params_v2``).
     """
     mc = run_meta.get("model_config") or {}
+    if "stage1_model" in mc:
+        return _figure_params_v2(mc, run_meta)
+
     mode = mc.get("mode")
     params: dict = {}
     if mc.get("hidden_dim") is not None:
@@ -71,7 +113,7 @@ def _figure_params(run_meta: dict) -> dict:
         params["mode"] = mode
     if mc.get("conditioning") is not None:
         params["conditioning"] = mc["conditioning"]
-    params["router"] = _router_summary(mc)
+    params["router"] = _router_summary(mc.get("router") or {})
     if run_meta.get("training_epoch") is not None:
         params["epoch"] = run_meta["training_epoch"]
     if run_meta.get("best_val_loss") is not None:
@@ -97,11 +139,11 @@ def _render_overlay(r: Reduced, params: dict):
 def _render_single(r: Reduced, params: dict):
     edges = np.asarray(r.payload["edges"])
     fig, ax = ps.new_figure("thesis-single", title=r.title, params=params)
-    ax.stairs(
-        _density(r.payload["rollout"], edges), edges, label=_SERIES_LABELS["rollout"]
-    )
+    ax.stairs(_density(r.payload["rollout"], edges), edges, label=_SERIES_LABELS["rollout"])
     if r.payload.get("log_y"):
         ax.set_yscale("log")
+    if r.payload.get("log_x"):
+        ax.set_xscale("log")
     ax.set_xlabel(r.xlabel)
     ax.set_ylabel("density")
     ps.style_legend(ax, title="source")
@@ -143,9 +185,7 @@ def _render_profile(r: Reduced, params: dict):
         mean = np.asarray(r.payload[f"{key}_mean"])
         std = np.asarray(r.payload[f"{key}_std"])
         (line,) = ax.plot(centers, mean, label=_SERIES_LABELS[key])
-        ax.fill_between(
-            centers, mean - std, mean + std, alpha=0.2, color=line.get_color()
-        )
+        ax.fill_between(centers, mean - std, mean + std, alpha=0.2, color=line.get_color())
     ax.set_xlabel(r.xlabel)
     ax.set_ylabel(r.payload.get("ylabel", "mean deposited energy [MeV]"))
     ps.style_legend(ax, title="source")
@@ -157,9 +197,7 @@ def _render_bar(r: Reduced, params: dict):
     x = np.arange(len(labels))
     width = 0.4
     fig, ax = ps.new_figure("thesis-single", title=r.title, params=params)
-    ax.bar(
-        x - width / 2, r.payload["reference"], width, label=_SERIES_LABELS["reference"]
-    )
+    ax.bar(x - width / 2, r.payload["reference"], width, label=_SERIES_LABELS["reference"])
     ax.bar(x + width / 2, r.payload["rollout"], width, label=_SERIES_LABELS["rollout"])
     ax.set_xticks(x)
     ax.set_xticklabels(labels, rotation=45, ha="right")
@@ -171,9 +209,7 @@ def _render_bar(r: Reduced, params: dict):
 def _render_router_gating(r: Reduced, params: dict):
     n_experts = r.payload["n_experts"]
     log_x = r.payload.get("log_x", False)
-    fig, axes = ps.new_figure(
-        "slide-16x9", title=r.title, params=params, nrows=1, ncols=2, squeeze=False
-    )
+    fig, axes = ps.new_figure("slide-16x9", title=r.title, params=params, nrows=1, ncols=2, squeeze=False)
     flat = axes.ravel()
     for ax, key in zip(flat, ("rollout", "reference")):
         side = r.payload.get(key, {})
@@ -182,9 +218,7 @@ def _render_router_gating(r: Reduced, params: dict):
         if len(centers) and means.size:
             cum = np.zeros(len(centers))
             for i in range(n_experts):
-                ax.fill_between(
-                    centers, cum, cum + means[:, i], alpha=0.7, label=f"expert {i}"
-                )
+                ax.fill_between(centers, cum, cum + means[:, i], alpha=0.7, label=f"expert {i}")
                 cum = cum + means[:, i]
         if log_x:
             ax.set_xscale("log")
@@ -303,9 +337,7 @@ def render_all(
         families.add(r.family)
         fig = render(r, run_meta)
         ps.savefig(fig, str(family_dir / r.id), formats=("pdf",))
-        (family_dir / f"{r.id}.yaml").write_text(
-            yaml.safe_dump(_plot_metadata(r, run_meta), sort_keys=False)
-        )
+        (family_dir / f"{r.id}.yaml").write_text(yaml.safe_dump(_plot_metadata(r, run_meta), sort_keys=False))
         pdfs.append(family_dir / f"{r.id}.pdf")
         import matplotlib.pyplot as plt
 
@@ -326,9 +358,7 @@ def render_all(
     )
     for fam in families:
         (out_dir / fam / "metadata.yaml").write_text(
-            yaml.safe_dump(
-                {"title": fam, "description": f"{fam} plots."}, sort_keys=False
-            )
+            yaml.safe_dump({"title": fam, "description": f"{fam} plots."}, sort_keys=False)
         )
 
     if run_gallery:
@@ -356,6 +386,4 @@ def render_run(run_dir: str | Path, *, run_gallery: bool = False) -> list[Path]:
         "reference": meta.reference,
         **meta.plot_meta,
     }
-    return render_all(
-        run_dir / "reduced", run_dir / "plots", run_meta, run_gallery=run_gallery
-    )
+    return render_all(run_dir / "reduced", run_dir / "plots", run_meta, run_gallery=run_gallery)

giant/analysis/router_gating.py

@@ -61,7 +61,8 @@ class _RouterHandle:
     pdg_map: dict[int, int]
     mat_map: dict[str, int]
     cond_normalizer: "Normalizer"
-    conditioning: str
+    particle_conditioning: str
+    material_conditioning: str
     router_type: str
 
 
@@ -69,32 +70,43 @@ def load_router(checkpoint: str | Path) -> _RouterHandle | None:
     """Load a checkpoint's Stage-1 router, or None if it isn't a MoE checkpoint."""
     import torch
 
+    from giant.checkpoint_io import conditioning_axes
     from giant.data.transforms import Normalizer
     from giant.model.network import build_models
 
     ckpt = torch.load(checkpoint, map_location="cpu", weights_only=False)
     model_cfg = ckpt.get("model_config") or {}
-    router_cfg = model_cfg.get("router")
+    # New nested shape (has a "stage1_model" key) vs. a v0.2 checkpoint's
+    # flat model_config.
+    router_cfg = (
+        (model_cfg.get("stage1_model") or {}).get("router") if "stage1_model" in model_cfg else model_cfg.get("router")
+    )
     if not router_cfg or not router_cfg.get("enabled"):
         return None
 
-    stage1, _ = build_models(model_cfg)
+    built = build_models(model_cfg)
+    stage1 = built["stage1"]
+    if stage1 is None:
+        return None
     stage1.load_state_dict(ckpt["model"])
     stage1.eval()
+    router = stage1.trunk.router
+    if router is None:
+        return None
 
+    particle_conditioning, material_conditioning = conditioning_axes(model_cfg)
     return _RouterHandle(
-        router=stage1.router,
+        router=router,
         pdg_map={int(k): v for k, v in ckpt["pdg_map"].items()},
         mat_map={str(k): v for k, v in ckpt["mat_map"].items()},
         cond_normalizer=Normalizer.from_dict(ckpt["normalizer"]["cond"]),
-        conditioning=model_cfg.get("conditioning", "embedding"),
+        particle_conditioning=particle_conditioning,
+        material_conditioning=material_conditioning,
         router_type=router_cfg["type"],
     )
 
 
-def _subsample(
-    lf: pl.LazyFrame, n: int, seed: int, extra_cols: tuple = ()
-) -> pl.DataFrame:
+def _subsample(lf: pl.LazyFrame, n: int, seed: int, extra_cols: tuple = ()) -> pl.DataFrame:
     total = lf.select(pl.len()).collect(engine="streaming").item()
     if total > n:
         threshold = int(n / total * 2**32)
@@ -102,9 +114,7 @@ def _subsample(
     return lf.select(*_COLS, *extra_cols).collect(engine="streaming")
 
 
-def _gate_for_df(
-    handle: _RouterHandle, df: pl.DataFrame
-) -> tuple[pl.DataFrame, np.ndarray]:
+def _gate_for_df(handle: _RouterHandle, df: pl.DataFrame) -> tuple[pl.DataFrame, np.ndarray]:
     """(filtered df, gate_weights) for rows in ``df`` with a known pdg/material.
 
     Rows whose species or material never appeared in the checkpoint's
@@ -128,13 +138,9 @@ def _gate_for_df(
     df = df.filter(pl.Series(known, dtype=pl.Boolean))
 
     data = {
-        "pre_pos": np.column_stack(
-            [df["pre_x"].to_numpy(), df["pre_y"].to_numpy(), df["pre_z"].to_numpy()]
-        ),
+        "pre_pos": np.column_stack([df["pre_x"].to_numpy(), df["pre_y"].to_numpy(), df["pre_z"].to_numpy()]),
         "pre_E": df["pre_E"].to_numpy(),
-        "pre_dir": np.column_stack(
-            [df["pre_dx"].to_numpy(), df["pre_dy"].to_numpy(), df["pre_dz"].to_numpy()]
-        ),
+        "pre_dir": np.column_stack([df["pre_dx"].to_numpy(), df["pre_dy"].to_numpy(), df["pre_dz"].to_numpy()]),
         "layer_id": df["layer_id"].to_numpy(),
         "pdg": df["pdg"].to_numpy(),
         "material": df["material"].to_numpy(),
@@ -144,12 +150,11 @@ def _gate_for_df(
         handle.pdg_map,
         handle.mat_map,
         cond_normalizer=handle.cond_normalizer,
-        conditioning=handle.conditioning,
+        particle_conditioning=handle.particle_conditioning,
+        material_conditioning=handle.material_conditioning,
     )
     with torch.no_grad():
-        gate = handle.router.gate(
-            torch.from_numpy(cond_cont).float(), torch.from_numpy(cond_cat).long()
-        ).numpy()
+        gate = handle.router.gate(torch.from_numpy(cond_cont).float(), torch.from_numpy(cond_cat).long()).numpy()
     return df, gate
 
 
@@ -172,9 +177,7 @@ def _quantile_bins(x: np.ndarray, gate: np.ndarray, n_bins: int) -> dict:
     return {"centers": centers[valid].tolist(), "means": means[valid].tolist()}
 
 
-def _top1_shares(
-    categories: np.ndarray, idx: np.ndarray, order: list, n_experts: int
-) -> dict[str, list[float]]:
+def _top1_shares(categories: np.ndarray, idx: np.ndarray, order: list, n_experts: int) -> dict[str, list[float]]:
     """Fraction of each category's rows hard-dispatched to each expert.
 
     Uses `Router.top1` (argmax), not the soft `gate` mean — grouped top-1
@@ -194,10 +197,7 @@ def _top1_shares(
     return shares
 
 
-_NOTE_NOT_MOE = (
-    "checkpoint has no enabled MoE router (model.router.enabled is "
-    "false/absent) — nothing to show"
-)
+_NOTE_NOT_MOE = "checkpoint has no enabled MoE router (model.router.enabled is false/absent) — nothing to show"
 
 _TITLES = {
     "router_gating": "Router gating (mixture-of-experts decision boundaries)",
@@ -233,9 +233,7 @@ def compute_router_gating(
         df = _subsample(lf, _SAMPLE_ROWS, seed)
         df, gate = _gate_for_df(handle, df)
         x = df["pre_E"].to_numpy()
-        sides[name] = (
-            _quantile_bins(x, gate, _N_BINS) if len(x) else {"centers": [], "means": []}
-        )
+        sides[name] = _quantile_bins(x, gate, _N_BINS) if len(x) else {"centers": [], "means": []}
 
     return Reduced(
         id="router_gating",
@@ -271,9 +269,7 @@ def compute_router_share_by_pdg(
         df, gate = _gate_for_df(handle, df)
         if len(df):
             idx = gate.argmax(axis=1)
-            shares = _top1_shares(
-                df["pdg"].to_numpy(), idx, top_pdgs, handle.router.n_experts
-            )
+            shares = _top1_shares(df["pdg"].to_numpy(), idx, top_pdgs, handle.router.n_experts)
         else:
             shares = {str(p): [0.0] * handle.router.n_experts for p in top_pdgs}
         sides[name] = {labels[i]: shares[str(p)] for i, p in enumerate(top_pdgs)}
@@ -318,9 +314,7 @@ def compute_router_share_by_process(
         counts = df["process"].value_counts().sort("count", descending=True)
         order = counts["process"].to_list()[:top_k]
         idx = gate.argmax(axis=1)
-        shares = _top1_shares(
-            df["process"].to_numpy(), idx, order, handle.router.n_experts
-        )
+        shares = _top1_shares(df["process"].to_numpy(), idx, order, handle.router.n_experts)
     else:
         order, shares = [], {}
 

giant/analysis/runtime_estimate.py

@@ -27,7 +27,7 @@ would then wrongly scale up with a bigger dataset. `RUNTIME_SAFETY_MARGIN` is
 deliberately generous (4x total) specifically to absorb that kind of
 contention spike instead. Rerun this calibration (pull fresh
 `condor_history`/`run_meta.json`, refit) if the catalog changes or timings
-drift — a synthetic local rebaseline via `scripts/profile_analysis_costs.py`
+drift — a synthetic local rebaseline via `giant/tools/profile_analysis_costs.py`
 is a reasonable fallback when no real cluster data is available yet, but
 undershoots real wall time badly (it can't see docker pull / `/ceph` I/O
 latency), which is exactly why this file moved off it.
@@ -54,9 +54,7 @@ _FIXED_OVERHEAD_S = 60.0
 # scan. Calibrated from the 3 real router jobs' observed wall times (119, 66,
 # 124s) — max minus _FIXED_OVERHEAD_S, on top of it.
 _ROUTER_FIXED_S = 64.0
-_ROUTER_IDS = frozenset(
-    {"router_gating", "router_share_by_pdg", "router_share_by_process"}
-)
+_ROUTER_IDS = frozenset({"router_gating", "router_share_by_pdg", "router_share_by_process"})
 
 # Conservative fallback for any catalog id not in _COST_MODEL (e.g. a plot
 # added after the last calibration run) — the most expensive fitted per-row

giant/analysis/sources.py

@@ -93,10 +93,7 @@ def _check_rollout_metadata(path: Path) -> None:
     metadata = pq.read_schema(path).metadata or {}
     coord = metadata.get(PREDICT_COORD_METADATA_KEY.encode())
     if coord is not None and coord.decode() != ROLLOUT_COORD_VALUE:
-        raise ValueError(
-            f"{path} is not a rollout file (coord={coord.decode()!r}); "
-            "expected `giant rollout` output"
-        )
+        raise ValueError(f"{path} is not a rollout file (coord={coord.decode()!r}); expected `giant rollout` output")
 
 
 def open_side(source: str | Path | pl.LazyFrame, side: Side) -> pl.LazyFrame:
@@ -121,11 +118,7 @@ def open_side(source: str | Path | pl.LazyFrame, side: Side) -> pl.LazyFrame:
     else:
         # The reference (a rollout's seed `dataset`) may be a directory of
         # parquet shards rather than a single file — scan them all.
-        lf = (
-            pl.scan_parquet(str(path / "**/*.parquet"))
-            if path.is_dir()
-            else pl.scan_parquet(path)
-        )
+        lf = pl.scan_parquet(str(path / "**/*.parquet")) if path.is_dir() else pl.scan_parquet(path)
     return lf.with_columns(pl.col("pdg").cast(pl.Int64))
 
 
@@ -137,9 +130,7 @@ def physical_steps(lf: pl.LazyFrame, side: Side) -> pl.LazyFrame:
     """
     if side is Side.reference:
         return lf
-    return lf.filter(
-        ~pl.col("termination_reason").is_in(list(SYNTHETIC_TERMINATION_REASONS))
-    )
+    return lf.filter(~pl.col("termination_reason").is_in(list(SYNTHETIC_TERMINATION_REASONS)))
 
 
 def secondaries(lf: pl.LazyFrame, side: Side) -> pl.LazyFrame:

giant/analysis/type_embedding_distance.py

@@ -0,0 +1,70 @@
+"""Secondary-type embedding-distance diagnostic.
+
+Unlike every other diagnostic in this package, the data isn't derivable from
+a rollout/reference parquet at all — it's the L1 distance between each
+emitted secondary's *raw* predicted embedding vector (under
+`stage2_model.particle_type.target = "embedding"`) and the nearest row of the
+conditioning's embedding table it snapped to, which only exists transiently
+inside `giant rollout` (`giant.rollout.decode_secondary_identity`), never
+written to a column. So it's accumulated once, at rollout time
+(`giant.rollout.L1DistCollector`), and stashed as a pre-finished histogram
+summary in the rollout YAML sidecar (`type_embedding_l1_dist`) — this module
+just turns that summary into a `Reduced`, no parquet scan involved (a
+`chunkable=False` spec, like `router_gating`, but even cheaper: no live model
+call either).
+
+A heavy right tail means the decoder is emitting vectors off the embedding
+manifold — the direct analogue of the species-collapse symptom the v0.3.0
+redesign exists to fix.
+"""
+
+from __future__ import annotations
+
+from giant.analysis.reduced import Reduced
+
+_NOTE_NOT_APPLICABLE = (
+    "not applicable: this rollout's checkpoint doesn't use "
+    "stage2_model.particle_type.target='embedding' (or generated no "
+    "secondaries), so giant rollout recorded no type_embedding_l1_dist "
+    "diagnostic in its YAML sidecar"
+)
+
+
+def compute_type_embedding_l1_distance(l1_dist: dict | None) -> Reduced:
+    """`Reduced` for the type-embedding-distance figure, or an explanatory
+    note if this checkpoint never populated the diagnostic.
+
+    `l1_dist`: `giant.rollout.L1DistCollector.summary()`'s dict, as recorded
+    in the rollout YAML's `type_embedding_l1_dist` key (`Bundle.
+    type_embedding_l1_dist`) — `{"n", "mean", "std", "min", "max",
+    "hist_edges", "hist_counts"}`.
+    """
+    if l1_dist is None:
+        return Reduced(
+            id="type_embedding_l1_distance",
+            family="model",
+            kind="unavailable",
+            title="Secondary-type embedding L1 distance",
+            xlabel="n/a",
+            payload={"note": _NOTE_NOT_APPLICABLE},
+        )
+
+    return Reduced(
+        id="type_embedding_l1_distance",
+        family="model",
+        kind="single_hist",
+        title="Secondary-type embedding L1 distance (predicted vector -> nearest PDG row)",
+        xlabel="L1 distance",
+        payload={
+            "edges": l1_dist["hist_edges"],
+            "rollout": l1_dist["hist_counts"],
+            "log_y": True,
+            "log_x": True,
+            "note": (
+                f"n={l1_dist['n']:,}  mean={l1_dist['mean']:.4g}  "
+                f"std={l1_dist['std']:.4g}  min={l1_dist['min']:.4g}  "
+                f"max={l1_dist['max']:.4g}; rollout only, no reference "
+                "concept for a raw pre-decode vector"
+            ),
+        },
+    )

giant/checkpoint_io.py

@@ -0,0 +1,234 @@
+"""Load a trained checkpoint into ready-to-run models (giant.cli's `predict`/`rollout`).
+
+Both commands need the same ~15 steps to go from a checkpoint path to two
+`eval()`-mode models plus their normalizers/vocab maps: load the pickle,
+validate it carries what current code expects, resolve which conditioning
+mode each axis was trained with, restore the top-N vocab maps (if the
+checkpoint used one-hot conditioning), rebuild the normalizers, construct the
+model from `model_config`, and load the requested (raw or EMA) weights. This
+used to be duplicated near-verbatim in both commands (issues.md Issue 5) —
+`load_for_inference` is the single implementation.
+
+This module intentionally has no Typer dependency, so it can be unit-tested
+directly and imported from non-CLI code (`giant.analysis.router_gating`,
+lazily — see that module's docstring for why). Failures raise
+`CheckpointCompatibilityError` with the same wording the CLI has always
+shown; the CLI layer catches it and does the `typer.echo`/`Exit(1)`.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+
+import torch
+from torch import nn
+
+from giant import config as gconfig
+from giant.constants import K_MAX
+from giant.data.loader import TopNMap
+from giant.data.setup_cache import topnmap_from_json
+from giant.data.transforms import Normalizer
+from giant.model.network import build_models
+
+
+class CheckpointCompatibilityError(Exception):
+    """Checkpoint is missing something `load_for_inference` needs."""
+
+
+def conditioning_axes(model_cfg: dict, default: str = "embedding") -> tuple[str, str]:
+    """(particle_conditioning, material_conditioning) for
+    `giant.data.transforms.build_cond_features`/`build_features` — from
+    either a v0.2 checkpoint's flat `model_config["conditioning"]` (one
+    shared string, same for both axes) or a new-format one (independent
+    `model_config["conditioning"]["particle"/"material"]["type"]` — the two
+    axes are configured independently and may differ)."""
+    raw = model_cfg.get("conditioning", default)
+    if isinstance(raw, dict):
+        return (
+            raw.get("particle", {}).get("type", default),
+            raw.get("material", {}).get("type", default),
+        )
+    return raw, raw
+
+
+def stage_cfg(model_cfg: dict, stage: str) -> dict:
+    """`model_cfg[f"{stage}_model"]` for a new-format model_config, `{}` for
+    a v0.2 flat one (whose ddpm schedule always used `CosineSchedule`'s own
+    default `T=1000` — never a config key — and which never had
+    `particle_type` at all, so `{}` is the correct fallback for both
+    `ddpm_steps`/`particle_type_other_policy` below)."""
+    val = model_cfg.get(f"{stage}_model")
+    return val if isinstance(val, dict) else {}
+
+
+def ddpm_steps(model_cfg: dict, stage: str) -> int:
+    return stage_cfg(model_cfg, stage).get("ddpm", {}).get("n_steps", 1000)
+
+
+def particle_type_other_policy(model_cfg: dict) -> str:
+    return stage_cfg(model_cfg, "stage2").get("particle_type", {}).get("other_policy", "sample")
+
+
+def load_pdg_topn_map(ckpt: dict) -> TopNMap | None:
+    """`ckpt["pdg_topn_map"]` as a `giant.data.loader.TopNMap`, or `None` if
+    this checkpoint's conditioning/particle_type never needed one (see
+    `giant.pipeline.run_setup_stage`, which only populates it when
+    `conditioning.particle.type` or `stage2_model.particle_type.target` is
+    `"onehot"`)."""
+    raw = ckpt.get("pdg_topn_map")
+    return topnmap_from_json(raw, axis="pdg") if raw is not None else None
+
+
+def load_mat_topn_map(ckpt: dict) -> TopNMap | None:
+    """`ckpt["mat_topn_map"]` as a `giant.data.loader.TopNMap`, or `None` if
+    this checkpoint's `conditioning.material.type` was never `"onehot"` (see
+    `giant.pipeline.run_setup_stage`)."""
+    raw = ckpt.get("mat_topn_map")
+    return topnmap_from_json(raw, axis="material") if raw is not None else None
+
+
+def load_sec_type_topn_map(ckpt: dict) -> TopNMap | None:
+    """`ckpt["sec_type_topn_map"]` as a `giant.data.loader.TopNMap`, or
+    `None` if this checkpoint's `stage2_model.particle_type.target` was never
+    `"onehot"` (see `giant.pipeline.run_setup_stage`).
+
+    Pre-gitea-#29 checkpoints have no `sec_type_topn_map` key at all — before
+    #29, the secondary-species decode map and the conditioning PDG onehot map
+    were always numerically the same map, saved once under `pdg_topn_map`.
+    For those, fall back to `load_pdg_topn_map` to reproduce that exact
+    behavior; a current checkpoint always has the key (possibly `null`, if
+    `particle_type.target != "onehot"`), so this fallback never fires for one."""
+    if "sec_type_topn_map" in ckpt:
+        raw = ckpt["sec_type_topn_map"]
+        return topnmap_from_json(raw, axis="pdg") if raw is not None else None
+    return load_pdg_topn_map(ckpt)
+
+
+@dataclass(frozen=True)
+class InferenceContext:
+    """Everything needed to run a trained checkpoint forward, resolved once."""
+
+    stage1: nn.Module | None
+    stage2: nn.Module | None
+    cond_norm: Normalizer
+    tgt_norm: Normalizer
+    sec_phys_norm: Normalizer
+    pdg_map: dict[int, int]
+    mat_map: dict[str, int]
+    pdg_topn_map: TopNMap | None
+    mat_topn_map: TopNMap | None
+    sec_type_topn_map: TopNMap | None
+    particle_conditioning: str
+    material_conditioning: str
+    k_max: int
+    stage1_ddpm_steps: int
+    stage2_ddpm_steps: int
+    other_policy: str
+    model_config: dict
+    epoch: int | None
+    best_val_loss: float | None
+
+
+def load_for_inference(
+    checkpoint: Path,
+    device: torch.device,
+    command_name: str,
+    weights: str = "raw",
+    require_stage2: bool = True,
+) -> InferenceContext:
+    """Load *checkpoint* and reconstruct everything `predict`/`rollout` need
+    to run it forward, on *device*, in `eval()` mode.
+
+    *command_name* (e.g. `"predict"`/`"rollout"`) only feeds the "needs both"
+    error message below. *weights* is `"raw"` (the live training weights) or
+    `"ema"` (the EMA shadow copy, see `--ema-decay`). *require_stage2*
+    controls whether a checkpoint with an inactive stage 2
+    (`stage2_model.active = false`) is an error (both current callers need
+    both stages) or an acceptable `stage2 = None` result — kept as a real
+    parameter since `stage{1,2}_model.active` is a real, if currently
+    stage1+stage2-only-in-practice, config option.
+    """
+    ckpt = torch.load(checkpoint, map_location="cpu", weights_only=False)
+    for key in ("model_config", "sec_decoder"):
+        if key not in ckpt:
+            raise CheckpointCompatibilityError(f"checkpoint has no {key} — retrain with the current code")
+
+    if "sec_phys" not in ckpt.get("normalizer", {}):
+        raise CheckpointCompatibilityError("checkpoint has no normalizer.sec_phys — retrain with the current code")
+
+    gconfig.warn_if_checkpoint_config_mismatch(checkpoint)
+
+    model_cfg = ckpt["model_config"]
+    particle_conditioning, material_conditioning = conditioning_axes(model_cfg)
+    pdg_topn_map = load_pdg_topn_map(ckpt)
+    mat_topn_map = load_mat_topn_map(ckpt)
+    if particle_conditioning == "onehot" and pdg_topn_map is None:
+        raise CheckpointCompatibilityError(
+            "checkpoint's conditioning.particle.type='onehot' but has no pdg_topn_map — retrain with the current code"
+        )
+    if material_conditioning == "onehot" and mat_topn_map is None:
+        raise CheckpointCompatibilityError(
+            "checkpoint's conditioning.material.type='onehot' but has no mat_topn_map — retrain with the current code"
+        )
+    sec_type_topn_map = load_sec_type_topn_map(ckpt)
+    particle_type_target = stage_cfg(model_cfg, "stage2").get("particle_type", {}).get("target", "onehot")
+    if particle_type_target == "onehot" and sec_type_topn_map is None:
+        raise CheckpointCompatibilityError(
+            "checkpoint's stage2_model.particle_type.target='onehot' but has no "
+            "sec_type_topn_map — retrain with the current code"
+        )
+    other_policy = particle_type_other_policy(model_cfg)
+    stage1_ddpm_steps = ddpm_steps(model_cfg, "stage1")
+    stage2_ddpm_steps = ddpm_steps(model_cfg, "stage2")
+    k_max = stage_cfg(model_cfg, "stage2").get("k_max", K_MAX)
+    pdg_map = {int(k): v for k, v in ckpt["pdg_map"].items()}
+    mat_map = {str(k): v for k, v in ckpt["mat_map"].items()}
+    cond_norm = Normalizer.from_dict(ckpt["normalizer"]["cond"])
+    tgt_norm = Normalizer.from_dict(ckpt["normalizer"]["target"])
+    sec_phys_norm = Normalizer.from_dict(ckpt["normalizer"]["sec_phys"])
+
+    built = build_models(model_cfg)
+    stage1, stage2 = built["stage1"], built["stage2"]
+    if require_stage2 and (stage1 is None or stage2 is None):
+        raise CheckpointCompatibilityError(
+            f"checkpoint has an inactive stage1 or stage2 — {command_name} needs both (see stage{{1,2}}_model.active)"
+        )
+
+    if weights == "raw":
+        model_key, sec_key = "model", "sec_decoder"
+    else:
+        model_key, sec_key = "model_ema", "sec_decoder_ema"
+        if model_key not in ckpt or sec_key not in ckpt:
+            raise CheckpointCompatibilityError(
+                f"{checkpoint} has no EMA weights (trained before --ema-decay, "
+                "or with --ema-decay 0) — use --weights raw"
+            )
+    if stage1 is not None:
+        stage1.load_state_dict(ckpt[model_key])
+        stage1.to(device).eval()
+    if stage2 is not None:
+        stage2.load_state_dict(ckpt[sec_key])
+        stage2.to(device).eval()
+
+    return InferenceContext(
+        stage1=stage1,
+        stage2=stage2,
+        cond_norm=cond_norm,
+        tgt_norm=tgt_norm,
+        sec_phys_norm=sec_phys_norm,
+        pdg_map=pdg_map,
+        mat_map=mat_map,
+        pdg_topn_map=pdg_topn_map,
+        mat_topn_map=mat_topn_map,
+        sec_type_topn_map=sec_type_topn_map,
+        particle_conditioning=particle_conditioning,
+        material_conditioning=material_conditioning,
+        k_max=k_max,
+        stage1_ddpm_steps=stage1_ddpm_steps,
+        stage2_ddpm_steps=stage2_ddpm_steps,
+        other_policy=other_policy,
+        model_config=model_cfg,
+        epoch=ckpt.get("epoch"),
+        best_val_loss=ckpt.get("best_val_loss"),
+    )

giant/cond_layout.py

@@ -0,0 +1,103 @@
+"""Single source of truth for the conditioning arrays' column layout (gitea #37).
+
+`cond_cont` and `cond_cat` are built in `giant.data.transforms` and consumed in
+`giant.model.encoders` / `giant.model.routers`. Their column order used to be
+written down independently on each side, kept in sync only by parallel comments
+— so getting it wrong produced silently mis-indexed columns rather than an
+exception, and adding a conditioning axis meant a coordinated multi-file edit.
+
+`CondLayout` owns that order. Both sides construct one from the same
+`conditioning.particle.type` / `conditioning.material.type` pair and read named
+slices off it, so the layout is stated exactly once. This module depends only on
+`giant.constants`, so both the data and model packages can import it.
+"""
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+from giant.constants import COND_DIM, COND_DIM_BASE, MATERIAL_PHYS_DIM, PARTICLE_PHYS_DIM
+
+# The three per-axis conditioning modes. Mirrors giant.config.Conditioning,
+# which this module deliberately does not import (giant.config pulls in the
+# whole model package).
+AXIS_TYPES = ("physical", "embedding", "onehot")
+
+
+@dataclass(frozen=True)
+class CondLayout:
+    """Column layout of `cond_cont`/`cond_cat` for one (particle, material) mode pair.
+
+    `cond_cont` is unconditionally `COND_DIM` wide regardless of mode: the base
+    block, then the particle physical block, then the material physical block.
+    An axis that isn't `"physical"` gets its block zero-filled and never reads
+    it (see `giant.data.transforms._physical_cond_columns`), so the widths are
+    mode-independent and only the *meaning* of a block changes.
+
+    `cond_cat` is 2 to 4 wide. Columns `PDG_COL`/`MAT_COL` are always the dense
+    training-vocab index; an axis in `"onehot"` mode appends one more column
+    holding its top-N-plus-other class index, particle before material.
+    """
+
+    particle_type: str
+    material_type: str
+
+    # cond_cat's dense-vocab columns, present in every mode. Under
+    # "physical"/"onehot" they are a reporting/router convenience the
+    # ConditionEncoder never reads; under "embedding" they are the signal.
+    PDG_COL: ClassVar[int] = 0
+    MAT_COL: ClassVar[int] = 1
+
+    def __post_init__(self) -> None:
+        if self.particle_type not in AXIS_TYPES:
+            raise ValueError(f"unknown conditioning.particle.type {self.particle_type!r}")
+        if self.material_type not in AXIS_TYPES:
+            raise ValueError(f"unknown conditioning.material.type {self.material_type!r}")
+
+    @classmethod
+    def from_types(cls, particle_type: str, material_type: str) -> "CondLayout":
+        """Named constructor — the entry point both sides use."""
+        return cls(particle_type=particle_type, material_type=material_type)
+
+    # --- cond_cont ---------------------------------------------------------
+
+    @property
+    def base(self) -> slice:
+        """pre_pos(3), log(pre_E)(1), pre_dir(3), layer_id(1)."""
+        return slice(0, COND_DIM_BASE)
+
+    @property
+    def particle_phys(self) -> slice:
+        """log(mass), charge — see `giant.particles`."""
+        return slice(COND_DIM_BASE, COND_DIM_BASE + PARTICLE_PHYS_DIM)
+
+    @property
+    def material_phys(self) -> slice:
+        """Z_eff, A_eff, log(density), log(X0), log(lambda_int) — see `giant.materials`."""
+        start = COND_DIM_BASE + PARTICLE_PHYS_DIM
+        return slice(start, start + MATERIAL_PHYS_DIM)
+
+    @property
+    def cont_dim(self) -> int:
+        return COND_DIM
+
+    # --- cond_cat ----------------------------------------------------------
+
+    @property
+    def particle_topn_col(self) -> int | None:
+        """Column of the particle top-N class index, or `None` if not `"onehot"`."""
+        return self.MAT_COL + 1 if self.particle_type == "onehot" else None
+
+    @property
+    def material_topn_col(self) -> int | None:
+        """Column of the material top-N class index, or `None` if not `"onehot"`.
+
+        Comes after the particle top-N column when both axes are `"onehot"`.
+        """
+        if self.material_type != "onehot":
+            return None
+        return self.MAT_COL + (2 if self.particle_type == "onehot" else 1)
+
+    @property
+    def cat_dim(self) -> int:
+        """Total `cond_cat` width: 2, plus one column per `"onehot"` axis."""
+        return self.MAT_COL + 1 + (self.particle_type == "onehot") + (self.material_type == "onehot")

giant/data/dataset.py

@@ -1,15 +1,48 @@
 from __future__ import annotations
 
 from pathlib import Path
+from typing import NamedTuple
 
 import numpy as np
 import torch
 from torch.utils.data import IterableDataset
 
+from giant.constants import K_MAX
 from giant.data.loader import event_id_offset, iter_file_chunks
 from giant.data.transforms import Normalizer, build_features, sorted_membership
 
 
+class StepBatch(NamedTuple):
+    """One training batch, as yielded by `StreamingStepsDataset`. Field order
+    is load-bearing for existing positional unpacking elsewhere (`trainers.py`,
+    `validate.py`, test fixtures) — append only, never insert or reorder.
+
+    cond_cont:  (B, COND_DIM) float32
+    cond_cat:   (B, 2/3/4) int64 — width 2 unless conditioning="onehot"
+    target_s1:  (B, 9) float32 — normalised Stage-1 primary target
+    n_sec:      (B,) int64 — true secondary count per step
+    sec_cont:   (B, k_max, SEC_SLOT_DIM) float32 — [stick_logit,
+                local_dir, log_mass, charge] per slot (mass/charge
+                normalised iff `sec_phys_normalizer` was given); always
+                computed the same way regardless of
+                stage2_model.particle_type.target, only actually used
+                downstream under target="physical"
+    proc_idx:   (B,) int64 — process-class label (ProcessRouter supervision
+                only; zeros when `proc_map` is None)
+    sec_type_idx: (B, k_max) int64 — per-slot class index into
+                `sec_type_class_map`, for particle_type.target in
+                ("onehot", "embedding"); zeros (unused) otherwise
+    """
+
+    cond_cont: torch.Tensor
+    cond_cat: torch.Tensor
+    target_s1: torch.Tensor
+    n_sec: torch.Tensor
+    sec_cont: torch.Tensor
+    proc_idx: torch.Tensor
+    sec_type_idx: torch.Tensor
+
+
 def make_event_split(
     all_event_ids: np.ndarray,
     val_fraction: float = 0.1,
@@ -38,18 +71,11 @@ class StreamingStepsDataset(IterableDataset):
     rather than single rows, so the batch is assembled with vectorized
     numpy slicing instead of a per-row Python loop in the default collate.
 
-    Each batch is a tuple:
-        (cond_cont, cond_cat, target_s1, n_sec, sec_cont, proc_idx)
-    where:
-        cond_cont:  (B, COND_DIM) float32
-        cond_cat:   (B, 2) int64
-        target_s1:  (B, 9) float32 — normalised Stage-1 primary target
-        n_sec:      (B,) int64 — true secondary count per step
-        sec_cont:   (B, K_MAX, SEC_SLOT_DIM) float32 — [stick_logit,
-                    local_dir, log_mass, charge] per slot (mass/charge
-                    normalised iff `sec_phys_normalizer` was given)
-        proc_idx:   (B,) int64 — process-class label (ProcessRouter supervision
-                    only; zeros when `proc_map` is None)
+    Each batch is a `StepBatch` — see its docstring for field meanings.
+
+    `k_max` (constructor arg, default the module constant) should match
+    `stage2_model.k_max` — it sets the padded
+    width of `sec_cont`/`sec_type_idx` above.
     """
 
     def __init__(
@@ -64,8 +90,13 @@ class StreamingStepsDataset(IterableDataset):
         shuffle_buffer: int = 65536,
         shuffle: bool = True,
         proc_map: dict[str, int] | None = None,
-        conditioning: str = "embedding",
+        particle_conditioning: str = "embedding",
+        material_conditioning: str = "embedding",
         sec_phys_normalizer: Normalizer | None = None,
+        pdg_topn_map: dict[int, int] | None = None,
+        mat_topn_map: dict[str, int] | None = None,
+        sec_type_class_map: dict | None = None,
+        k_max: int = K_MAX,
     ) -> None:
         self.files = list(files)
         self._offsets = {path: event_id_offset(i) for i, path in enumerate(self.files)}
@@ -79,8 +110,13 @@ class StreamingStepsDataset(IterableDataset):
         self.shuffle_buffer = max(shuffle_buffer, batch_size)
         self.shuffle = shuffle
         self.proc_map = proc_map
-        self.conditioning = conditioning
+        self.particle_conditioning = particle_conditioning
+        self.material_conditioning = material_conditioning
         self.sec_phys_normalizer = sec_phys_normalizer
+        self.pdg_topn_map = pdg_topn_map
+        self.mat_topn_map = mat_topn_map
+        self.sec_type_class_map = sec_type_class_map
+        self.k_max = k_max
 
     def __iter__(self):
         worker_info = torch.utils.data.get_worker_info()
@@ -98,25 +134,17 @@ class StreamingStepsDataset(IterableDataset):
         buf_nsec: list[np.ndarray] = []
         buf_sec: list[np.ndarray] = []
         buf_proc: list[np.ndarray] = []
+        buf_type: list[np.ndarray] = []
         buf_n = 0
 
         for path in files:
-            for chunk in iter_file_chunks(path, offset=self._offsets[path]):
+            for chunk in iter_file_chunks(path, offset=self._offsets[path], k_max=self.k_max):
                 mask = sorted_membership(chunk["event_id"], self._events_arr)
                 if not mask.any():
                     continue
                 chunk = {k: v[mask] for k, v in chunk.items()}
 
-                (
-                    cond_cont,
-                    cond_cat,
-                    target_s1,
-                    n_sec,
-                    sec_cont,
-                    proc_idx,
-                    _,
-                    _,
-                ) = build_features(
+                feats = build_features(
                     chunk,
                     self.pdg_map,
                     self.mat_map,
@@ -125,15 +153,21 @@ class StreamingStepsDataset(IterableDataset):
                     sec_phys_normalizer=self.sec_phys_normalizer,
                     proc_map=self.proc_map,
                     require_secondaries=True,
-                    conditioning=self.conditioning,
+                    particle_conditioning=self.particle_conditioning,
+                    material_conditioning=self.material_conditioning,
+                    pdg_topn_map=self.pdg_topn_map,
+                    mat_topn_map=self.mat_topn_map,
+                    sec_type_class_map=self.sec_type_class_map,
+                    k_max=self.k_max,
                 )
-                buf_cont.append(cond_cont)
-                buf_cat.append(cond_cat)
-                buf_tgt.append(target_s1)
-                buf_nsec.append(n_sec)
-                buf_sec.append(sec_cont)
-                buf_proc.append(proc_idx)
-                buf_n += len(cond_cont)
+                buf_cont.append(feats.cond_cont)
+                buf_cat.append(feats.cond_cat)
+                buf_tgt.append(feats.target_s1)
+                buf_nsec.append(feats.n_sec)
+                buf_sec.append(feats.sec_cont)
+                buf_proc.append(feats.proc_idx)
+                buf_type.append(feats.sec_type_idx)
+                buf_n += len(feats.cond_cont)
 
                 if buf_n >= self.shuffle_buffer:
                     (
@@ -143,6 +177,7 @@ class StreamingStepsDataset(IterableDataset):
                         buf_nsec,
                         buf_sec,
                         buf_proc,
+                        buf_type,
                         buf_n,
                     ) = yield from self._flush(
                         buf_cont,
@@ -151,6 +186,7 @@ class StreamingStepsDataset(IterableDataset):
                         buf_nsec,
                         buf_sec,
                         buf_proc,
+                        buf_type,
                         final=False,
                     )
 
@@ -162,6 +198,7 @@ class StreamingStepsDataset(IterableDataset):
                 buf_nsec,
                 buf_sec,
                 buf_proc,
+                buf_type,
                 final=True,
             )
 
@@ -173,6 +210,7 @@ class StreamingStepsDataset(IterableDataset):
         buf_nsec: list[np.ndarray],
         buf_sec: list[np.ndarray],
         buf_proc: list[np.ndarray],
+        buf_type: list[np.ndarray],
         final: bool,
     ):
         cont = np.concatenate(buf_cont)
@@ -181,28 +219,30 @@ class StreamingStepsDataset(IterableDataset):
         nsec = np.concatenate(buf_nsec)
         sec = np.concatenate(buf_sec)
         proc = np.concatenate(buf_proc)
+        styp = np.concatenate(buf_type)
 
         if self.shuffle:
             idx = np.random.permutation(len(cont))
             cont, cat, tgt = cont[idx], cat[idx], tgt[idx]
-            nsec, sec, proc = nsec[idx], sec[idx], proc[idx]
+            nsec, sec, proc, styp = nsec[idx], sec[idx], proc[idx], styp[idx]
 
         bs = self.batch_size
         n = len(cont)
         n_full = n // bs if not final else (n + bs - 1) // bs
         for start in range(0, n_full * bs, bs):
             end = min(start + bs, n)
-            yield (
-                torch.from_numpy(cont[start:end]).float(),
-                torch.from_numpy(cat[start:end]).long(),
-                torch.from_numpy(tgt[start:end]).float(),
-                torch.from_numpy(nsec[start:end]).long(),
-                torch.from_numpy(sec[start:end]).float(),
-                torch.from_numpy(proc[start:end]).long(),
+            yield StepBatch(
+                cond_cont=torch.from_numpy(cont[start:end]).float(),
+                cond_cat=torch.from_numpy(cat[start:end]).long(),
+                target_s1=torch.from_numpy(tgt[start:end]).float(),
+                n_sec=torch.from_numpy(nsec[start:end]).long(),
+                sec_cont=torch.from_numpy(sec[start:end]).float(),
+                proc_idx=torch.from_numpy(proc[start:end]).long(),
+                sec_type_idx=torch.from_numpy(styp[start:end]).long(),
             )
 
         if final:
-            return [], [], [], [], [], [], 0
+            return [], [], [], [], [], [], [], 0
         rem = n_full * bs
         return (
             [cont[rem:]],
@@ -211,5 +251,6 @@ class StreamingStepsDataset(IterableDataset):
             [nsec[rem:]],
             [sec[rem:]],
             [proc[rem:]],
+            [styp[rem:]],
             n - rem,
         )

giant/data/loader.py

@@ -1,3 +1,4 @@
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Iterator
 
@@ -5,6 +6,8 @@ import numpy as np
 import pandas as pd
 import pyarrow.parquet as pq
 
+from giant.constants import K_MAX
+
 # A manifest is a plain text file listing one parquet path per line, used to
 # name a curated subset of files (e.g. a train/holdout pool) without copying
 # or symlinking the underlying parquet files. Lines are resolved relative to
@@ -13,7 +16,7 @@ import pyarrow.parquet as pq
 MANIFEST_SUFFIX = ".manifest"
 
 # Each input parquet file is a separate Geant4 job converted 1:1 from its own
-# ROOT file (scripts/steps_to_parquet.py), and a job's event_id numbering
+# ROOT file (giant/tools/steps_to_parquet.py), and a job's event_id numbering
 # always restarts from 0 — so when multiple files are loaded together (a
 # directory or .manifest), raw event_id values collide across files even
 # though they refer to unrelated events. Every per-file event_id column gets
@@ -112,9 +115,7 @@ def _pad_dir_col(dx: pd.Series, dy: pd.Series, dz: pd.Series, K: int) -> np.ndar
     return out
 
 
-def _df_to_dict(df: pd.DataFrame, offset: int = 0) -> dict[str, np.ndarray]:
-    from giant.constants import K_MAX
-
+def _df_to_dict(df: pd.DataFrame, offset: int = 0, k_max: int = K_MAX) -> dict[str, np.ndarray]:
     has_sec_lists = "sec_E_list" in df.columns
 
     d: dict[str, np.ndarray] = {
@@ -133,9 +134,7 @@ def _df_to_dict(df: pd.DataFrame, offset: int = 0) -> dict[str, np.ndarray]:
         # / ProcessRouter). Guarded like has_sec_lists: older parquet
         # conversions predating this column still load fine.
         "process": (
-            df["process"].to_numpy(dtype=object)
-            if "process" in df.columns
-            else np.full(len(df), "", dtype=object)
+            df["process"].to_numpy(dtype=object) if "process" in df.columns else np.full(len(df), "", dtype=object)
         ),
         "step_length": df["step_length"].to_numpy(dtype=np.float32),
         "post_E": df["post_E"].to_numpy(dtype=np.float32),
@@ -146,17 +145,15 @@ def _df_to_dict(df: pd.DataFrame, offset: int = 0) -> dict[str, np.ndarray]:
     }
 
     if has_sec_lists:
-        d["sec_E_list"] = _pad_list_col(df["sec_E_list"], K_MAX)
-        d["sec_pdg_list"] = _pad_list_col_int(df["sec_pdg_list"], K_MAX)
-        d["sec_dir_list"] = _pad_dir_col(
-            df["sec_dx_list"], df["sec_dy_list"], df["sec_dz_list"], K_MAX
-        )
+        d["sec_E_list"] = _pad_list_col(df["sec_E_list"], k_max)
+        d["sec_pdg_list"] = _pad_list_col_int(df["sec_pdg_list"], k_max)
+        d["sec_dir_list"] = _pad_dir_col(df["sec_dx_list"], df["sec_dy_list"], df["sec_dz_list"], k_max)
 
     return d
 
 
-def load_steps(path: str | Path, offset: int = 0) -> dict[str, np.ndarray]:
-    return _df_to_dict(pd.read_parquet(path), offset=offset)
+def load_steps(path: str | Path, offset: int = 0, k_max: int = K_MAX) -> dict[str, np.ndarray]:
+    return _df_to_dict(pd.read_parquet(path), offset=offset, k_max=k_max)
 
 
 def load_event_ids(path: str | Path, offset: int = 0) -> np.ndarray:
@@ -165,13 +162,15 @@ def load_event_ids(path: str | Path, offset: int = 0) -> np.ndarray:
     return _offset_event_id(ids, offset)
 
 
-def iter_file_chunks(
-    path: str | Path, offset: int = 0
-) -> Iterator[dict[str, np.ndarray]]:
-    """Yield one parquet row-group at a time so a large file never fully loads."""
+def iter_file_chunks(path: str | Path, offset: int = 0, k_max: int = K_MAX) -> Iterator[dict[str, np.ndarray]]:
+    """Yield one parquet row-group at a time so a large file never fully loads.
+
+    `k_max` sets the padded width of the sec_*_list columns (should match
+    `stage2_model.k_max`); defaults to the
+    module constant for callers that don't care (e.g. Stage-1-only reads)."""
     pf = pq.ParquetFile(path)
     for i in range(pf.num_row_groups):
-        yield _df_to_dict(pf.read_row_group(i).to_pandas(), offset=offset)
+        yield _df_to_dict(pf.read_row_group(i).to_pandas(), offset=offset, k_max=k_max)
 
 
 _COND_COLS = [
@@ -205,15 +204,11 @@ def _cond_df_to_dict(df: pd.DataFrame, offset: int = 0) -> dict[str, np.ndarray]
     }
 
 
-def iter_cond_chunks(
-    path: str | Path, offset: int = 0
-) -> Iterator[dict[str, np.ndarray]]:
+def iter_cond_chunks(path: str | Path, offset: int = 0) -> Iterator[dict[str, np.ndarray]]:
     """Yield conditioning-only row-groups (no post-step columns read from disk)."""
     pf = pq.ParquetFile(path)
     for i in range(pf.num_row_groups):
-        yield _cond_df_to_dict(
-            pf.read_row_group(i, columns=_COND_COLS).to_pandas(), offset=offset
-        )
+        yield _cond_df_to_dict(pf.read_row_group(i, columns=_COND_COLS).to_pandas(), offset=offset)
 
 
 def build_index_maps(
@@ -243,6 +238,44 @@ def build_index_maps_from_files(
     )
 
 
+def _accumulate_value_counts(counts: dict, series: pd.Series, cast) -> None:
+    for name, count in series.value_counts().items():
+        name = cast(name)
+        counts[name] = counts.get(name, 0) + int(count)
+
+
+def _rank_by_frequency_from_files(files: list[Path], column: str, cast) -> dict:
+    """Scan `column` across `files` and return `{cast(value): total_count}`,
+    accumulated in file order (see `fingerprint_files`'s docstring on why
+    scan order — not a normalized/sorted order — is preserved: it drives
+    tie-breaking in the frequency ranking below)."""
+    counts: dict = {}
+    for path in files:
+        df = pd.read_parquet(path, columns=[column])
+        _accumulate_value_counts(counts, df[column], cast)
+    return counts
+
+
+def _topn_plus_other_map(counts: dict, n_classes: int) -> tuple[dict, dict]:
+    """Frequency-capped value->index map: the `n_classes - 1` most frequent
+    keys get their own index; every rarer key is bucketed into a shared
+    "other" index (`n_classes - 1`).
+
+    Returns `(class_map, other_members)` — `other_members` is `{key: count}`
+    for every key bucketed into "other" (the empirical within-bucket
+    distribution, for `other_policy = "sample"` at rollout).
+    """
+    ranked = sorted(counts, key=lambda k: counts[k], reverse=True)
+    keep = ranked[: max(n_classes - 1, 0)]
+    class_map = {k: i for i, k in enumerate(keep)}
+    other_idx = n_classes - 1
+    other_members: dict = {}
+    for k in ranked[len(keep) :]:
+        class_map[k] = other_idx
+        other_members[k] = counts[k]
+    return class_map, other_members
+
+
 def build_process_map_from_files(files: list[Path], n_experts: int) -> dict[str, int]:
     """Scan the `process` column and build a frequency-capped process->index map.
 
@@ -253,16 +286,66 @@ def build_process_map_from_files(files: list[Path], n_experts: int) -> dict[str,
     mirrors how `build_features` clamps the n_sec label to K_MAX for the
     fixed-width n_sec_head classifier.
     """
-    counts: dict[str, int] = {}
+    counts = _rank_by_frequency_from_files(files, "process", str)
+    class_map, _ = _topn_plus_other_map(counts, n_experts)
+    return class_map
+
+
+@dataclass
+class TopNMap:
+    """A frequency-capped value->index map for a conditioning/type axis (PDG
+    or material), plus the empirical within-bucket distribution of whatever
+    got folded into "other" — see `build_topn_map_from_files`."""
+
+    class_map: dict
+    other_members: dict
+
+
+def build_topn_map_from_files(files: list[Path], column: str, n_classes: int, cast=str) -> TopNMap:
+    """Scan `column` and build a frequency-capped value->index map, structurally
+    identical to `build_process_map_from_files` (shares its ranking core via
+    `_topn_plus_other_map`), generalized over the source column and key type.
+
+    Used for the material axis (`column="material"`, `cast=str`, matching
+    `mat_map`'s key type). The PDG axis uses
+    `build_pdg_topn_map_from_files` instead (it needs to pool two columns,
+    which this single-column form can't express). Also records
+    `other_members` (the empirical within-"other" distribution), needed
+    later for `other_policy = "sample"` at rollout — computed now since it's
+    free during this same scan.
+    """
+    counts = _rank_by_frequency_from_files(files, column, cast)
+    class_map, other_members = _topn_plus_other_map(counts, n_classes)
+    return TopNMap(class_map=class_map, other_members=other_members)
+
+
+def build_pdg_topn_map_from_files(files: list[Path], n_classes: int) -> TopNMap:
+    """PDG top-N-plus-other map, pooling counts from BOTH roles a PDG code
+    plays in this dataset: a step's own primary particle (`pdg` column) and
+    an emitted secondary's species (`sec_pdg_list`, exploded) — shared by
+    `conditioning.particle.type = "onehot"` and
+    `stage2_model.particle_type.target = "onehot"`. Pooling both is what
+    keeps a species that's common as a secondary
+    but rare as a primary (or vice versa) from being pushed into "other"
+    just because one role's count alone looks small — the meeting's failure
+    mode (zero photon secondaries, hallucinated antineutrinos) was
+    specifically about secondary-species collapse, so the map this feeds
+    needs to reflect secondary frequency, not just primary frequency.
+
+    `sec_pdg_list` is absent from parquet files predating the parent->child
+    join (see `_df_to_dict`'s `has_sec_lists` guard) — silently skipped for
+    those, same convention as elsewhere in this module.
+    """
+    counts: dict = {}
     for path in files:
-        df = pd.read_parquet(path, columns=["process"])
-        for name, count in df["process"].value_counts().items():
-            name = str(name)
-            counts[name] = counts.get(name, 0) + int(count)
-    ranked = sorted(counts, key=lambda name: counts[name], reverse=True)
-    keep = ranked[: max(n_experts - 1, 0)]
-    proc_map = {name: i for i, name in enumerate(keep)}
-    other_idx = n_experts - 1
-    for name in ranked[len(keep) :]:
-        proc_map[name] = other_idx
-    return proc_map
+        columns = ["pdg"]
+        has_sec = "sec_pdg_list" in pq.ParquetFile(path).schema_arrow.names
+        if has_sec:
+            columns.append("sec_pdg_list")
+        df = pd.read_parquet(path, columns=columns)
+        _accumulate_value_counts(counts, df["pdg"], int)
+        if has_sec:
+            exploded = df["sec_pdg_list"].explode().dropna()
+            _accumulate_value_counts(counts, exploded, int)
+    class_map, other_members = _topn_plus_other_map(counts, n_classes)
+    return TopNMap(class_map=class_map, other_members=other_members)

giant/data/setup_cache.py

@@ -23,7 +23,7 @@ import numpy as np
 
 from giant import config
 from giant.constants import COND_DIM, K_MAX, PARTICLE_PHYS_DIM, SEC_SLOT_DIM, X_DIM
-from giant.data.loader import event_id_offset, load_event_ids
+from giant.data.loader import TopNMap, event_id_offset, load_event_ids
 from giant.data.transforms import Normalizer, sorted_membership
 
 # Bump manually on a change to the data-encoding semantics (e.g. a future
@@ -96,10 +96,50 @@ def fingerprint_files(files: list[Path]) -> list[list]:
     return out
 
 
-def normalizer_key(val_fraction: float, seed: int, conditioning: str) -> str:
+def normalizer_key(
+    val_fraction: float,
+    seed: int,
+    particle_conditioning: str,
+    material_conditioning: str,
+) -> str:
     # .6g avoids float-repr drift (e.g. 0.1 vs 0.10000000000000002) causing
-    # spurious cache misses between runs with the "same" val_fraction.
-    return f"valfrac={val_fraction:.6g}_seed={seed}_cond={conditioning}"
+    # spurious cache misses between runs with the "same" val_fraction. The two
+    # conditioning axes are independent and both
+    # affect which cond_cont columns are computed for real vs. zero-filled
+    # (giant.data.transforms._physical_cond_columns), so both must be part of
+    # the key or two mixed-axis runs could collide on the same cache entry.
+    return f"valfrac={val_fraction:.6g}_seed={seed}_pcond={particle_conditioning}_mcond={material_conditioning}"
+
+
+# Top-N-map axes: "pdg" keys match pdg_map's int
+# keys (shared by conditioning.particle.type="onehot" and
+# stage2_model.particle_type.target="onehot" — one map for both), "material"
+# keys match mat_map's str keys.
+_TOPN_AXIS_CASTS = {"pdg": int, "material": str}
+
+
+def topn_key(axis: str, n_classes: int) -> str:
+    """JSON-safe key for `SetupCache.topn_maps` — N is part of the key so the
+    sidecar stays reusable across runs with different emb_dim (see the
+    dict[int, dict] precedent `proc_maps` sets, keyed by n_experts)."""
+    if axis not in _TOPN_AXIS_CASTS:
+        raise ValueError(f"unknown top-N map axis {axis!r}, expected one of {sorted(_TOPN_AXIS_CASTS)}")
+    return f"{axis}:{n_classes}"
+
+
+def topnmap_to_json(m: TopNMap) -> dict:
+    return {
+        "class_map": {str(k): v for k, v in m.class_map.items()},
+        "other_members": {str(k): v for k, v in m.other_members.items()},
+    }
+
+
+def topnmap_from_json(d: dict, axis: str) -> TopNMap:
+    cast = _TOPN_AXIS_CASTS[axis]
+    return TopNMap(
+        class_map={cast(k): v for k, v in d["class_map"].items()},
+        other_members={cast(k): v for k, v in d["other_members"].items()},
+    )
 
 
 @dataclass
@@ -119,9 +159,7 @@ class NormalizerEntry:
             "tgt_norm": self.tgt_norm.to_dict(),
             "sec_phys_norm": self.sec_phys_norm.to_dict(),
             "n_train_steps": self.n_train_steps,
-            "energy_quantiles": np.asarray(
-                self.energy_quantiles, dtype=np.float32
-            ).tolist(),
+            "energy_quantiles": np.asarray(self.energy_quantiles, dtype=np.float32).tolist(),
         }
 
     @classmethod
@@ -143,6 +181,8 @@ class SetupCache:
     event_index: tuple[np.ndarray, np.ndarray] | None = None
     proc_maps: dict[int, dict[str, int]] = field(default_factory=dict)
     normalizers: dict[str, NormalizerEntry] = field(default_factory=dict)
+    topn_maps: dict[str, TopNMap] = field(default_factory=dict)
+    """Keyed by `topn_key(axis, n_classes)`."""
 
     @classmethod
     def empty(cls, files: list[Path]) -> "SetupCache":
@@ -156,6 +196,7 @@ class SetupCache:
             "fingerprint": self.fingerprint,
             "proc_maps": {str(k): v for k, v in self.proc_maps.items()},
             "normalizers": {k: v.to_json() for k, v in self.normalizers.items()},
+            "topn_maps": {k: topnmap_to_json(v) for k, v in self.topn_maps.items()},
         }
         if self.vocab is not None:
             pdg_map, mat_map = self.vocab
@@ -185,9 +226,8 @@ class SetupCache:
                 np.array(d["event_index"]["counts"], dtype=np.int64),
             )
         proc_maps = {int(k): v for k, v in d.get("proc_maps", {}).items()}
-        normalizers = {
-            k: NormalizerEntry.from_json(v) for k, v in d.get("normalizers", {}).items()
-        }
+        normalizers = {k: NormalizerEntry.from_json(v) for k, v in d.get("normalizers", {}).items()}
+        topn_maps = {k: topnmap_from_json(v, axis=k.split(":", 1)[0]) for k, v in d.get("topn_maps", {}).items()}
         return cls(
             fingerprint=d["fingerprint"],
             git_hash=d.get("git_hash", "unknown"),
@@ -195,6 +235,7 @@ class SetupCache:
             event_index=event_index,
             proc_maps=proc_maps,
             normalizers=normalizers,
+            topn_maps=topn_maps,
         )
 
     def merge(self, other: "SetupCache") -> "SetupCache":
@@ -209,17 +250,14 @@ class SetupCache:
             fingerprint=other.fingerprint,
             git_hash=other.git_hash,
             vocab=other.vocab if other.vocab is not None else self.vocab,
-            event_index=(
-                other.event_index if other.event_index is not None else self.event_index
-            ),
+            event_index=(other.event_index if other.event_index is not None else self.event_index),
             proc_maps={**self.proc_maps, **other.proc_maps},
             normalizers={**self.normalizers, **other.normalizers},
+            topn_maps={**self.topn_maps, **other.topn_maps},
         )
 
 
-def load(
-    data: str | Path, files: list[Path], echo=lambda *a, **k: None
-) -> SetupCache | None:
+def load(data: str | Path, files: list[Path], echo=lambda *a, **k: None) -> SetupCache | None:
     """Load and validate the sidecar for `data`; `None` on any miss (never raises).
 
     A missing file, corrupt JSON, format-version mismatch, dimension-constant
@@ -243,9 +281,7 @@ def load(
             echo("setup cache: format version changed — ignoring stale cache")
             return None
         if raw.get("dims") != _DIMS:
-            echo(
-                "setup cache: model dimension constants changed — ignoring stale cache"
-            )
+            echo("setup cache: model dimension constants changed — ignoring stale cache")
             return None
         fp = fingerprint_files(files)
         if raw.get("fingerprint") != fp:
@@ -288,9 +324,7 @@ def save(
         with open(lock_path, "a") as lock_file:
             fcntl.flock(lock_file, fcntl.LOCK_EX)
             try:
-                base = load(data, files, echo=lambda *a, **k: None) or SetupCache.empty(
-                    files
-                )
+                base = load(data, files, echo=lambda *a, **k: None) or SetupCache.empty(files)
                 merged = base.merge(sections)
                 payload = json.dumps(merged.to_json(), separators=(",", ":"))
                 tmp.write_text(payload)
@@ -298,9 +332,7 @@ def save(
             finally:
                 fcntl.flock(lock_file, fcntl.LOCK_UN)
     except OSError as exc:
-        echo(
-            f"setup cache: could not write {path} ({exc}) — continuing without caching"
-        )
+        echo(f"setup cache: could not write {path} ({exc}) — continuing without caching")
         try:
             tmp.unlink(missing_ok=True)
         except OSError:
@@ -311,16 +343,12 @@ def compute_event_index_from_files(files: list[Path]) -> tuple[np.ndarray, np.nd
     """Unique event ids + per-event row (step) counts, across all `files`."""
     if not files:
         return np.empty(0, dtype=np.int64), np.empty(0, dtype=np.int64)
-    all_ids = np.concatenate(
-        [load_event_ids(f, offset=event_id_offset(i)) for i, f in enumerate(files)]
-    )
+    all_ids = np.concatenate([load_event_ids(f, offset=event_id_offset(i)) for i, f in enumerate(files)])
     unique_ids, counts = np.unique(all_ids, return_counts=True)
     return unique_ids, counts
 
 
-def n_train_steps_for_split(
-    unique_ids: np.ndarray, counts: np.ndarray, train_events_arr: np.ndarray
-) -> int:
+def n_train_steps_for_split(unique_ids: np.ndarray, counts: np.ndarray, train_events_arr: np.ndarray) -> int:
     """Row (step) count summed over whichever `unique_ids` fall in `train_events_arr`.
 
     `train_events_arr` must be ascending and duplicate-free (as produced by

giant/data/transforms.py

@@ -1,7 +1,11 @@
 import warnings
+from typing import NamedTuple
 
 import numpy as np
 
+from giant.cond_layout import CondLayout
+from giant.constants import K_MAX
+
 _EPS = 1e-8
 
 # Floor added to each energy fraction before taking log-ratios so the simplex
@@ -82,9 +86,7 @@ def energy_simplex_encode(
     return z.astype(np.float32)
 
 
-def energy_simplex_decode(
-    z: np.ndarray, pre_E: np.ndarray
-) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+def energy_simplex_decode(z: np.ndarray, pre_E: np.ndarray) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
     """Inverse of `energy_simplex_encode`: ALR coords + pre_E → physical energies.
 
     A softmax over `[z_edep, z_sec, 0]` recovers the three simplex fractions, so
@@ -116,9 +118,7 @@ def _rodrigues_axis(pre_dir: np.ndarray) -> np.ndarray:
     arbitrary second operand) on every row; profiling on a 114M-row file
     showed `np.cross` as the single hottest call inside this rotation.
     """
-    axis = np.stack(
-        [pre_dir[:, 1], -pre_dir[:, 0], np.zeros_like(pre_dir[:, 0])], axis=1
-    )
+    axis = np.stack([pre_dir[:, 1], -pre_dir[:, 0], np.zeros_like(pre_dir[:, 0])], axis=1)
     axis_norm = np.linalg.norm(axis, axis=1, keepdims=True)  # (N,1)
     # axis_norm ~ 0 happens at BOTH poles: pre_dir ~ +ẑ (forward) and
     # pre_dir ~ -ẑ (near-exact backscatter) — ‖pre_dir × ẑ‖ = sin(angle to
@@ -197,9 +197,7 @@ def local_frame_rotation(pre_dir: np.ndarray, post_dir: np.ndarray) -> np.ndarra
     kxv = _cross_with_z_axis(axis, post_dir)  # (N,3)
     kdv = (axis * post_dir).sum(axis=1, keepdims=True)  # (N,1)
 
-    return (post_dir * cos_t + kxv * sin_t + axis * kdv * (1.0 - cos_t)).astype(
-        np.float32
-    )
+    return (post_dir * cos_t + kxv * sin_t + axis * kdv * (1.0 - cos_t)).astype(np.float32)
 
 
 class Normalizer:
@@ -339,22 +337,21 @@ def sorted_membership(values: np.ndarray, sorted_arr: np.ndarray) -> np.ndarray:
     return sorted_arr[idx] == values
 
 
-def _vectorized_map_lookup(
-    values: np.ndarray, mapping: dict, strict: bool = True
-) -> np.ndarray:
+def _vectorized_map_lookup(values: np.ndarray, mapping: dict, strict: bool = True, default: int = 0) -> np.ndarray:
     """Vectorized equivalent of `np.array([mapping[v] for v in values], dtype=np.int64)`.
 
     Replaces a per-element Python dict lookup with one `searchsorted` call.
     Raises `KeyError` if any value in `values` isn't a key of `mapping`,
     matching the dict-comprehension it replaces (never silently misassigns)
-    — unless `strict=False`, in which case unmapped values get a dummy index
-    of 0 instead. Only pass `strict=False` where the caller has independently
-    verified the resulting index is never actually read (e.g.
+    — unless `strict=False`, in which case unmapped values get `default`
+    instead. Only pass `strict=False` where the caller has independently
+    verified the resulting index is either never actually read (e.g.
     `build_cond_features` under `conditioning="physical"`, where
-    `ConditionEncoder` ignores `cond_cat` entirely); it exists so a rollout
-    can be seeded with a species/material outside the training vocab without
-    a spurious `KeyError`, which is the entire point of physical-property
-    conditioning.
+    `ConditionEncoder` ignores `cond_cat` entirely) or where `default` is a
+    deliberate fallback class (e.g. a top-N map's "other" index for a raw
+    value outside the training vocab). It exists so a rollout can be seeded
+    with a species/material outside the training vocab without a spurious
+    `KeyError`, which is the entire point of physical-property conditioning.
     """
     keys = np.asarray(list(mapping.keys()))
     vals = np.asarray(list(mapping.values()), dtype=np.int64)
@@ -366,7 +363,7 @@ def _vectorized_map_lookup(
     found = keys_sorted[pos] == values
     if not found.all():
         if not strict:
-            out = np.zeros(values.shape, dtype=np.int64)
+            out = np.full(values.shape, default, dtype=np.int64)
             out[found] = vals_sorted[pos[found]]
             return out
         missing = np.unique(values[~found])
@@ -384,9 +381,7 @@ def travel_direction(pre_pos: np.ndarray, post_pos: np.ndarray) -> np.ndarray:
     disp = post_pos - pre_pos
     norm = np.linalg.norm(disp, axis=1, keepdims=True)
     safe_norm = np.where(norm < 1e-7, 1.0, norm)
-    return np.where(norm < 1e-7, np.array([[0.0, 0.0, 1.0]]), disp / safe_norm).astype(
-        np.float32
-    )
+    return np.where(norm < 1e-7, np.array([[0.0, 0.0, 1.0]]), disp / safe_norm).astype(np.float32)
 
 
 def reconstruct_post_pos(
@@ -405,9 +400,7 @@ def reconstruct_post_pos(
     return (pre_pos + step_length.reshape(-1, 1) * travel_dir_world).astype(np.float32)
 
 
-def inv_local_frame_rotation(
-    pre_dir: np.ndarray, post_dir_local: np.ndarray
-) -> np.ndarray:
+def inv_local_frame_rotation(pre_dir: np.ndarray, post_dir_local: np.ndarray) -> np.ndarray:
     """Inverse of local_frame_rotation: rotate from local frame back to world frame.
 
     Applies R^T (same axis, negative angle) to post_dir_local.
@@ -421,9 +414,7 @@ def inv_local_frame_rotation(
     kdv = (axis * post_dir_local).sum(axis=1, keepdims=True)
 
     # Negative angle: sin_t → -sin_t
-    return (post_dir_local * cos_t - kxv * sin_t + axis * kdv * (1.0 - cos_t)).astype(
-        np.float32
-    )
+    return (post_dir_local * cos_t - kxv * sin_t + axis * kdv * (1.0 - cos_t)).astype(np.float32)
 
 
 _STICK_LOGIT_CLIP = 10.0  # logit value used for the last valid secondary slot
@@ -488,14 +479,10 @@ def encode_secondaries(
                 remaining_raw = e_sec - cumsum[:, i - 1]
             shortfall_flagged |= sec_valid[:, i] & (remaining_raw < -_SHORTFALL_TOL)
             remaining = np.maximum(remaining_raw, _EPS)
-            f = np.clip(
-                sec_E_list[:, i].astype(np.float64) / remaining, _EPS, 1.0 - _EPS
-            )
+            f = np.clip(sec_E_list[:, i].astype(np.float64) / remaining, _EPS, 1.0 - _EPS)
             logit = np.log(f / (1.0 - f)).astype(np.float32)
             # Last valid slot: give it the full remaining budget
-            is_last = sec_valid[:, i] & ~(
-                sec_valid[:, i + 1] if i + 1 < K else np.zeros(N, dtype=bool)
-            )
+            is_last = sec_valid[:, i] & ~(sec_valid[:, i + 1] if i + 1 < K else np.zeros(N, dtype=bool))
             logit = np.where(is_last, _STICK_LOGIT_CLIP, logit)
             logit = np.where(
                 sec_valid[:, i],
@@ -522,9 +509,7 @@ def encode_secondaries(
             # Only rotate valid slots; leave padded slots as (0,0,1) or whatever.
             valid_mask = sec_valid[:, i]
             if valid_mask.any():
-                dir_local[valid_mask, i] = local_frame_rotation(
-                    pre_dir[valid_mask], sec_dir_list[valid_mask, i]
-                )
+                dir_local[valid_mask, i] = local_frame_rotation(pre_dir[valid_mask], sec_dir_list[valid_mask, i])
 
     if sec_pdg_list is not None:
         from giant.particles import particle_phys_array
@@ -550,42 +535,69 @@ def encode_secondaries(
     return sec_cont.astype(np.float32)
 
 
-def decode_secondaries(
+def encode_secondary_type_idx(sec_pdg_list: np.ndarray, sec_valid: np.ndarray, class_map: dict) -> np.ndarray:
+    """Per-secondary-slot class index into `class_map` — (N, K_MAX) int64.
+
+    `class_map` is either a top-N-plus-other map's `class_map`
+    (`stage2_model.particle_type.target = "onehot"`, see
+    `giant.data.loader.build_pdg_topn_map_from_files`) or the dense `pdg_map`
+    (`target = "embedding"`). Not used at all for `target = "physical"` —
+    that target keeps using `encode_secondaries`'s (log_mass, charge)
+    columns unchanged.
+
+    Padding slots get index 0 (their looked-up value is discarded downstream
+    by the `sec_valid`/`n_sec` mask regardless, so any in-vocabulary dummy
+    code works). A *real, valid* secondary whose code is missing from
+    `class_map` raises `KeyError` (`strict=True`) rather than silently
+    misassigning — for `target="onehot"` this should never actually
+    trigger, since `build_pdg_topn_map_from_files` pools both primary and
+    secondary occurrences precisely so every secondary species seen in
+    these files has a key (in "other" at worst); for `target="embedding"`
+    (which reuses the dense, primary-only `pdg_map`) it's a real signal
+    that a secondary-only species exists with no primary-role counterpart.
+    """
+    N, K = sec_pdg_list.shape
+    # An arbitrary already-present key works as the padding-slot dummy code
+    # (unlike encode_secondaries' physics-derived phys lookup, this is an
+    # index into class_map's own vocabulary, so a fixed sentinel like 22
+    # isn't guaranteed to be a key — an arbitrary present one always is).
+    dummy = next(iter(class_map))
+    safe_pdg = np.where(sec_valid, sec_pdg_list, dummy)
+    idx = _vectorized_map_lookup(safe_pdg.reshape(-1), class_map, strict=True).reshape(N, K)
+    return np.where(sec_valid, idx, 0).astype(np.int64)
+
+
+def decode_secondary_cont(
     sec_cont: np.ndarray,
     n_sec: np.ndarray,
     e_sec: np.ndarray,
     pre_dir: np.ndarray,
-    sec_phys_normalizer: "Normalizer | None" = None,
-) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
-    """Inverse of encode_secondaries: continuous targets → physical secondary attrs.
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Continuous-only half of `decode_secondaries`'s inverse: the
+    stick-breaking energy split and local->world direction — generator/
+    `particle_type.target`-independent, since every target (`"physical"`,
+    `"onehot"`, `"embedding"`) shares the same `CONT_SLOT_DIM`-wide
+    (stick_logit, dir) prefix and differs only
+    in what follows it. `decode_secondaries` (target="physical") is the
+    original all-in-one form built on top of this; `target` in `("onehot",
+    "embedding")` decodes their type slice separately via
+    `giant.particles.decode_topn_class`/`decode_embedding_nearest` and calls
+    this directly instead — see `giant/rollout.py`.
 
-    sec_cont: (N, K_MAX, 6) — [stick_logit, local_dir_x, local_dir_y,
-        local_dir_z, log_mass, charge] (log_mass/charge normalised iff
-        `sec_phys_normalizer` was applied when this was produced — e.g. a
-        raw model prediction; pass the same normalizer here to invert it)
+    sec_cont: (N, K, >=CONT_SLOT_DIM) — only columns `[:, :, :CONT_SLOT_DIM]`
+        (stick_logit, local dir) are read; a caller may pass its full
+        per-slot tensor (continuous + type) unsliced.
     n_sec: (N,) integer secondary counts
     e_sec: (N,) total secondary energy budget [MeV]
     pre_dir: (N, 3) pre-step world-frame direction
 
-    Returns (sec_E, sec_dir_world, sec_mass, sec_charge, sec_valid) each
-    shape (N, K_MAX). The valid slots' energies (`sec_E[sec_valid]`, per row)
-    always sum to exactly `e_sec` — see the rescaling below. mass/charge are
-    the model's raw predicted physical identity for each secondary, used
-    as-is (no snapping to a discrete PDG code) — see giant/particles.py for
-    the separate, reporting-only nearest-PDG lookup callers may apply on top
-    of this for display/bookkeeping purposes.
+    Returns (sec_E, sec_dir_world, sec_valid), shapes (N, K), (N, K, 3),
+    (N, K). The valid slots' energies (`sec_E[sec_valid]`, per row) always
+    sum to exactly `e_sec` — see the rescaling below.
     """
-    if sec_phys_normalizer is not None:
-        N_, K_, _ = sec_cont.shape
-        phys = sec_phys_normalizer.inverse_transform(sec_cont[:, :, 4:6].reshape(-1, 2))
-        sec_cont = sec_cont.copy()
-        sec_cont[:, :, 4:6] = phys.reshape(N_, K_, 2)
-
-    N, K, _ = sec_cont.shape
+    N, K = sec_cont.shape[0], sec_cont.shape[1]
     stick_logits = sec_cont[:, :, 0]  # (N, K)
     dir_local = sec_cont[:, :, 1:4].copy()  # (N, K, 3)
-    log_mass = sec_cont[:, :, 4]  # (N, K)
-    charge = sec_cont[:, :, 5]  # (N, K)
 
     # Flow-matching output isn't guaranteed unit norm; normalise before the
     # rotation below, which preserves magnitude rather than fixing it up.
@@ -625,9 +637,53 @@ def decode_secondaries(
     for i in range(K):
         valid = sec_valid[:, i]
         if valid.any():
-            sec_dir_world[valid, i] = inv_local_frame_rotation(
-                pre_dir[valid], dir_local[valid, i]
-            )
+            sec_dir_world[valid, i] = inv_local_frame_rotation(pre_dir[valid], dir_local[valid, i])
+
+    return sec_E, sec_dir_world, sec_valid
+
+
+def decode_secondaries(
+    sec_cont: np.ndarray,
+    n_sec: np.ndarray,
+    e_sec: np.ndarray,
+    pre_dir: np.ndarray,
+    sec_phys_normalizer: "Normalizer | None" = None,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
+    """Inverse of encode_secondaries: continuous targets → physical secondary attrs.
+
+    `particle_type.target = "physical"` only (the type slice is a raw
+    (log_mass, charge) regression target folded straight into `sec_cont`) —
+    `"onehot"`/`"embedding"` decode through `decode_secondary_cont` +
+    `giant.particles.decode_topn_class`/`decode_embedding_nearest` instead,
+    since their type slice isn't (log_mass, charge) at all. See
+    `decode_secondary_cont`'s docstring for why the two share the energy/
+    direction logic below.
+
+    sec_cont: (N, K_MAX, 6) — [stick_logit, local_dir_x, local_dir_y,
+        local_dir_z, log_mass, charge] (log_mass/charge normalised iff
+        `sec_phys_normalizer` was applied when this was produced — e.g. a
+        raw model prediction; pass the same normalizer here to invert it)
+    n_sec: (N,) integer secondary counts
+    e_sec: (N,) total secondary energy budget [MeV]
+    pre_dir: (N, 3) pre-step world-frame direction
+
+    Returns (sec_E, sec_dir_world, sec_mass, sec_charge, sec_valid) each
+    shape (N, K_MAX). mass/charge are the model's raw predicted physical
+    identity for each secondary, used as-is (no snapping to a discrete PDG
+    code) — see giant/particles.py for the separate, reporting-only
+    nearest-PDG lookup callers may apply on top of this for display/
+    bookkeeping purposes.
+    """
+    if sec_phys_normalizer is not None:
+        N_, K_, _ = sec_cont.shape
+        phys = sec_phys_normalizer.inverse_transform(sec_cont[:, :, 4:6].reshape(-1, 2))
+        sec_cont = sec_cont.copy()
+        sec_cont[:, :, 4:6] = phys.reshape(N_, K_, 2)
+
+    sec_E, sec_dir_world, sec_valid = decode_secondary_cont(sec_cont, n_sec, e_sec, pre_dir)
+
+    log_mass = sec_cont[:, :, 4]  # (N, K)
+    charge = sec_cont[:, :, 5]  # (N, K)
 
     # mass is non-negative by construction (inv_log_transform of a real
     # number is always > 0); clip to 0 for padded/invalid slots rather than
@@ -638,63 +694,74 @@ def decode_secondaries(
     return sec_E, sec_dir_world, sec_mass, sec_charge, sec_valid
 
 
-def _physical_cond_columns(
-    data: dict[str, np.ndarray], conditioning: str
-) -> np.ndarray:
+def _physical_cond_columns(data: dict[str, np.ndarray], layout: CondLayout) -> np.ndarray:
     """(N, PARTICLE_PHYS_DIM + MATERIAL_PHYS_DIM) physical conditioning columns.
 
-    "embedding" mode zero-fills (cheap, and ConditionEncoder never reads
-    these columns in that mode — so an unfilled giant.materials table can
-    never crash an "embedding"-mode run). "physical" mode computes them for
-    real: particle columns come from `data["mass"]`/`data["charge"]` when the
-    caller already knows them directly (rollout.py, for a track descended
-    from a model-predicted secondary — see giant/rollout.py's "no snapping"
-    design), else derived from `data["pdg"]` via giant.particles; material
-    columns always come from `data["material"]` via giant.materials, since
-    material is never itself a model prediction.
+    The particle and material blocks are gated independently and may mix
+    freely — e.g. material `physical` with particle `embedding` — so e.g.
+    `particle_type="embedding"` + `material_type="physical"` zero-fills only
+    the particle columns and computes the material ones for real.
+
+    "embedding"/"onehot" zero-fill their block (cheap, and ConditionEncoder
+    never reads these columns in either mode — so an unfilled
+    giant.materials table can never crash an "embedding"/"onehot"-mode run).
+    "physical" computes it for real: particle columns come from
+    `data["mass"]`/`data["charge"]` when the caller already knows them
+    directly (rollout.py, for a track descended from a model-predicted
+    secondary — see giant/rollout.py's "no snapping" design), else derived
+    from `data["pdg"]` via giant.particles; material columns always come
+    from `data["material"]` via giant.materials, since material is never
+    itself a model prediction.
     """
     from giant.constants import MATERIAL_PHYS_DIM, PARTICLE_PHYS_DIM
 
-    if conditioning == "embedding":
-        n = len(next(iter(data.values())))
-        return np.zeros((n, PARTICLE_PHYS_DIM + MATERIAL_PHYS_DIM), dtype=np.float32)
-    if conditioning != "physical":
-        raise ValueError(f"unknown conditioning mode {conditioning!r}")
+    n = len(next(iter(data.values())))
 
-    from giant.materials import material_properties_array
-    from giant.particles import particle_phys_array
+    if layout.particle_type == "physical":
+        from giant.particles import particle_phys_array
 
-    if "mass" in data and "charge" in data:
-        mass = np.asarray(data["mass"], dtype=np.float32)
-        charge = np.asarray(data["charge"], dtype=np.float32)
+        if "mass" in data and "charge" in data:
+            mass = np.asarray(data["mass"], dtype=np.float32)
+            charge = np.asarray(data["charge"], dtype=np.float32)
+        else:
+            mass, charge = particle_phys_array(data["pdg"]).T
+        particle_cols = np.column_stack([log_transform(mass), charge])
     else:
-        mass, charge = particle_phys_array(data["pdg"]).T
+        particle_cols = np.zeros((n, PARTICLE_PHYS_DIM), dtype=np.float32)
 
-    z_eff, a_eff, density, x0, lambda_int = material_properties_array(
-        data["material"]
-    ).T
+    if layout.material_type == "physical":
+        from giant.materials import material_properties_array
 
-    return np.column_stack(
-        [
-            log_transform(mass),
-            charge,
-            z_eff,
-            a_eff,
-            log_transform(density),
-            log_transform(x0),
-            log_transform(lambda_int),
-        ]
-    ).astype(np.float32)
+        z_eff, a_eff, density, x0, lambda_int = material_properties_array(data["material"]).T
+        material_cols = np.column_stack(
+            [
+                z_eff,
+                a_eff,
+                log_transform(density),
+                log_transform(x0),
+                log_transform(lambda_int),
+            ]
+        )
+    else:
+        material_cols = np.zeros((n, MATERIAL_PHYS_DIM), dtype=np.float32)
+
+    return np.column_stack([particle_cols, material_cols]).astype(np.float32)
 
 
-def build_cond_features(
+def _build_cond_arrays(
     data: dict[str, np.ndarray],
     pdg_map: dict[int, int],
     mat_map: dict[str, int],
-    cond_normalizer: "Normalizer | None" = None,
-    conditioning: str = "embedding",
+    layout: CondLayout,
+    pdg_topn_map: dict[int, int] | None,
+    mat_topn_map: dict[str, int] | None,
 ) -> tuple[np.ndarray, np.ndarray]:
-    """Build conditioning arrays only — no target, no post-step variables."""
+    """The un-normalized `(cond_cont, cond_cat)` pair, in `layout`'s column order.
+
+    Both `build_cond_features` and `build_features` go through here, so the
+    column order — and everything that depends on it — is stated once. See
+    `giant.cond_layout.CondLayout` for the layout itself.
+    """
     cond_cont = np.column_stack(
         [
             data["pre_pos"],
@@ -702,51 +769,100 @@ def build_cond_features(
             data["pre_dir"],
             data["layer_id"].astype(np.float32),
         ]
-    ).astype(np.float32)
-    cond_cont = np.column_stack(
-        [cond_cont, _physical_cond_columns(data, conditioning)]
-    ).astype(np.float32)
+    ).astype(np.float32)  # (N, COND_DIM_BASE=8)
+    cond_cont = np.column_stack([cond_cont, _physical_cond_columns(data, layout)]).astype(
+        np.float32
+    )  # (N, COND_DIM=15)
 
-    # In "physical" mode cond_cat is only a reporting/router convenience —
-    # ConditionEncoder never reads it (giant/model/network.py) — so a
-    # species/material outside the training vocab (the whole point of
-    # physical-property conditioning) gets a dummy index instead of raising.
-    # In "embedding" mode cond_cat IS the conditioning signal, so an unmapped
-    # value must still raise loudly rather than silently misassign.
-    strict = conditioning == "embedding"
-    pdg_idx = _vectorized_map_lookup(data["pdg"], pdg_map, strict=strict)
-    mat_idx = _vectorized_map_lookup(data["material"], mat_map, strict=strict)
-    cond_cat = np.column_stack([pdg_idx, mat_idx])
+    # In "physical" mode cond_cat's first two columns are only a
+    # reporting/router convenience — ConditionEncoder never reads them
+    # (giant/model/encoders.py) — so a species/material outside the training
+    # vocab (the whole point of physical-property conditioning) gets a dummy
+    # index instead of raising. In "embedding" mode those columns ARE the
+    # conditioning signal, so an unmapped value must still raise loudly
+    # rather than silently misassign. In "onehot" mode they again go unread
+    # (the topN columns below are the real signal), so they're as permissive
+    # as "physical". Each axis's strictness is independent.
+    pdg_idx = _vectorized_map_lookup(data["pdg"], pdg_map, strict=layout.particle_type == "embedding")
+    mat_idx = _vectorized_map_lookup(data["material"], mat_map, strict=layout.material_type == "embedding")
+    # Which extra columns exist is the layout's call, not "did the caller
+    # happen to pass a map" — that's what used to let the producer and
+    # ConditionEncoder disagree. A map for a non-"onehot" axis is unused.
+    cat_cols = [pdg_idx, mat_idx]
+    if layout.particle_topn_col is not None:
+        if pdg_topn_map is None:
+            raise ValueError("conditioning.particle.type='onehot' needs pdg_topn_map")
+        cat_cols.append(_vectorized_map_lookup(data["pdg"], pdg_topn_map))
+    if layout.material_topn_col is not None:
+        if mat_topn_map is None:
+            raise ValueError("conditioning.material.type='onehot' needs mat_topn_map")
+        cat_cols.append(_vectorized_map_lookup(data["material"], mat_topn_map))
+    cond_cat = np.column_stack(cat_cols)  # (N, layout.cat_dim)
+
+    return cond_cont, cond_cat
+
+
+def build_cond_features(
+    data: dict[str, np.ndarray],
+    pdg_map: dict[int, int],
+    mat_map: dict[str, int],
+    cond_normalizer: "Normalizer | None" = None,
+    particle_conditioning: str = "embedding",
+    material_conditioning: str = "embedding",
+    pdg_topn_map: dict[int, int] | None = None,
+    mat_topn_map: dict[str, int] | None = None,
+) -> tuple[np.ndarray, np.ndarray]:
+    """Build conditioning arrays only — no target, no post-step variables.
+
+    `particle_conditioning`/`material_conditioning` are independent —
+    e.g. `particle_conditioning="embedding"` +
+    `material_conditioning="physical"` is a valid mix.
+
+    `pdg_topn_map`/`mat_topn_map` (a top-N-plus-other `class_map`, see
+    `giant.data.loader.build_topn_map_from_files`) supply the extra `cond_cat`
+    columns read by `ConditionEncoder`'s `"onehot"` mode, and are required
+    whenever the corresponding axis is `"onehot"`. See
+    `giant.cond_layout.CondLayout` for which columns exist where.
+    """
+    layout = CondLayout.from_types(particle_conditioning, material_conditioning)
+    cond_cont, cond_cat = _build_cond_arrays(data, pdg_map, mat_map, layout, pdg_topn_map, mat_topn_map)
 
     if cond_normalizer is not None:
-        cond_cont = _cond_normalizer_transform(cond_cont, cond_normalizer, conditioning)
+        cond_cont = _cond_normalizer_transform(cond_cont, cond_normalizer, layout)
 
     return cond_cont, cond_cat
 
 
 def _cond_normalizer_transform(
-    cond_cont: np.ndarray, cond_normalizer: "Normalizer", conditioning: str
+    cond_cont: np.ndarray,
+    cond_normalizer: "Normalizer",
+    layout: CondLayout,
 ) -> np.ndarray:
     """Apply ``cond_normalizer``, padding a legacy narrower normalizer if needed.
 
     Checkpoints trained before physical-property conditioning (``COND_DIM``
     8->15, ``giant/constants.py``) saved a ``COND_DIM_BASE``-wide (8) cond
     normalizer, fit before ``build_cond_features`` grew the extra physical
-    columns. In "embedding" mode those columns are never read by
-    ``ConditionEncoder`` (``giant/model/network.py``), so padding the missing
+    columns. When NEITHER axis is "physical" those columns are never read by
+    ``ConditionEncoder`` (``giant/model/encoders.py``), so padding the missing
     entries with mean=0/std=1 is a safe no-op that keeps such checkpoints
-    usable under the current, always-``COND_DIM``-wide contract. In
-    "physical" mode the physical columns are load-bearing, so a mismatch
-    there is a real incompatibility, not something to paper over.
+    usable under the current, always-``COND_DIM``-wide contract. If EITHER
+    axis is "physical" its columns are load-bearing, so a mismatch there is a
+    real incompatibility, not something to paper over.
     """
     mean, std = cond_normalizer.mean, cond_normalizer.std
     assert mean is not None and std is not None, "Normalizer not fitted"
     width = cond_cont.shape[-1]
     if mean.shape[-1] < width:
-        if conditioning != "embedding":
+        physical_load_bearing = "physical" in (
+            layout.particle_type,
+            layout.material_type,
+        )
+        if physical_load_bearing:
             raise ValueError(
                 f"cond normalizer has {mean.shape[-1]} columns, expected "
-                f"{width}, and conditioning={conditioning!r} reads the "
+                f"{width}, and particle_conditioning={layout.particle_type!r}/"
+                f"material_conditioning={layout.material_type!r} reads the "
                 "physical columns directly — this checkpoint predates "
                 "physical-property conditioning and can't be safely padded; "
                 "retrain it under the current code."
@@ -757,6 +873,41 @@ def _cond_normalizer_transform(
     return ((cond_cont - mean) / std).astype(np.float32)
 
 
+class StepFeatures(NamedTuple):
+    """Output of `build_features`. Field order is load-bearing for existing
+    positional unpacking (tests, `StreamingStepsDataset`) — append only,
+    never insert or reorder.
+
+    target_s1: (N, 9) Stage-1 primary post-step target (unchanged from Phase 1)
+    n_sec: (N,) integer secondary counts (target for n_sec head)
+    sec_cont: (N, K_MAX, SEC_SLOT_DIM=6) continuous secondary targets
+              [stick_logit, dir_local, log_mass, charge] — mass/charge are
+              the secondary's real physical identity (from its ground-truth
+              PDG code), a fixed regression target, not a learned/snapped one.
+              Always computed the same way regardless of
+              `stage2_model.particle_type.target` — only actually used
+              downstream under `target = "physical"`.
+    proc_idx: (N,) integer process-class label (ProcessRouter supervision only —
+              never conditioning). Zeros when `proc_map` is None or the loaded
+              data has no "process" column (e.g. pre-conversion parquet files).
+    sec_type_idx: (N, K_MAX) integer secondary class index into
+              `sec_type_class_map`, for `stage2_model.particle_type.target`
+              in `("onehot", "embedding")` — see `encode_secondary_type_idx`.
+              Zero-filled (and unused) when `sec_type_class_map` is None
+              (i.e. `target = "physical"`).
+    """
+
+    cond_cont: np.ndarray
+    cond_cat: np.ndarray
+    target_s1: np.ndarray
+    n_sec: np.ndarray
+    sec_cont: np.ndarray
+    proc_idx: np.ndarray
+    sec_type_idx: np.ndarray
+    cond_normalizer: Normalizer | None
+    target_normalizer: Normalizer | None
+
+
 def build_features(
     data: dict[str, np.ndarray],
     pdg_map: dict[int, int],
@@ -767,29 +918,17 @@ def build_features(
     fit: bool = False,
     proc_map: dict[str, int] | None = None,
     require_secondaries: bool = False,
-    conditioning: str = "embedding",
+    particle_conditioning: str = "embedding",
+    material_conditioning: str = "embedding",
     sec_phys_only: bool = False,
-) -> tuple[
-    np.ndarray,
-    np.ndarray,
-    np.ndarray,
-    np.ndarray,
-    np.ndarray,
-    np.ndarray,
-    Normalizer | None,
-    Normalizer | None,
-]:
-    """Assemble (cond_cont, cond_cat, target_s1, n_sec, sec_cont, proc_idx) arrays.
-
-    target_s1: (N, 9) Stage-1 primary post-step target (unchanged from Phase 1)
-    n_sec: (N,) integer secondary counts (target for n_sec head)
-    sec_cont: (N, K_MAX, SEC_SLOT_DIM=6) continuous secondary targets
-              [stick_logit, dir_local, log_mass, charge] — mass/charge are
-              the secondary's real physical identity (from its ground-truth
-              PDG code), a fixed regression target, not a learned/snapped one.
-    proc_idx: (N,) integer process-class label (ProcessRouter supervision only —
-              never conditioning). Zeros when `proc_map` is None or the loaded
-              data has no "process" column (e.g. pre-conversion parquet files).
+    pdg_topn_map: dict[int, int] | None = None,
+    mat_topn_map: dict[str, int] | None = None,
+    sec_type_class_map: dict | None = None,
+    k_max: int = K_MAX,
+) -> StepFeatures:
+    """Assemble a `StepFeatures` of (cond_cont, cond_cat, target_s1, n_sec,
+    sec_cont, proc_idx, sec_type_idx, cond_normalizer, target_normalizer) —
+    see `StepFeatures` for field meanings.
 
     require_secondaries: when True, raise if any step has n_sec > 0 but the
         per-secondary list columns are absent (a mis-converted file that would
@@ -800,17 +939,27 @@ def build_features(
         the stick-breaking/direction-rotation blocks of `sec_cont` (zero-filled
         instead) for callers (normalizer fitting) that only read
         `sec_cont[:, :, 4:6]` and would otherwise discard that work.
+
+    pdg_topn_map/mat_topn_map: source of the extra `cond_cat` columns for
+        `ConditionEncoder`'s `"onehot"` mode — see `build_cond_features`.
+
+    sec_type_class_map: the map `sec_type_idx` is looked up against — a
+        top-N-plus-other map's `class_map` for `target = "onehot"`, or the
+        dense `pdg_map` for `target = "embedding"` (pass `pdg_map` itself).
+        `None` for `target = "physical"`.
+
+    k_max: should match `stage2_model.k_max` —
+        overridden internally by `data["sec_E_list"]`'s own padded width when
+        present (the loader already padded it to some k_max; that width is
+        authoritative), so this only actually matters when secondary list
+        columns are absent (Stage-1-only reads, or a pre-secondary-join
+        file), where it sets `sec_cont`/`sec_type_idx`'s zero-filled width.
     """
-    from giant.constants import K_MAX
 
     post_dir_local = local_frame_rotation(data["pre_dir"], data["post_dir"])
-    travel_dir_local = local_frame_rotation(
-        data["pre_dir"], travel_direction(data["pre_pos"], data["post_pos"])
-    )
+    travel_dir_local = local_frame_rotation(data["pre_dir"], travel_direction(data["pre_pos"], data["post_pos"]))
 
-    energy_z = energy_simplex_encode(
-        data["edep"], data["e_sec"], data["post_E"], data["pre_E"]
-    )  # (N, 2)
+    energy_z = energy_simplex_encode(data["edep"], data["e_sec"], data["post_E"], data["pre_E"])  # (N, 2)
 
     target_s1 = np.column_stack(
         [
@@ -822,39 +971,31 @@ def build_features(
     ).astype(np.float32)  # (N, 9)
 
     # Phase 2: conditioning drops n_sec and log(e_sec)
-    cond_cont = np.column_stack(
-        [
-            data["pre_pos"],
-            log_transform(data["pre_E"]),
-            data["pre_dir"],
-            data["layer_id"].astype(np.float32),
-        ]
-    ).astype(np.float32)  # (N, COND_DIM_BASE=8)
-    cond_cont = np.column_stack(
-        [cond_cont, _physical_cond_columns(data, conditioning)]
-    ).astype(np.float32)  # (N, COND_DIM=15)
+    layout = CondLayout.from_types(particle_conditioning, material_conditioning)
+    cond_cont, cond_cat = _build_cond_arrays(data, pdg_map, mat_map, layout, pdg_topn_map, mat_topn_map)
 
-    pdg_idx = _vectorized_map_lookup(data["pdg"], pdg_map)
-    mat_idx = _vectorized_map_lookup(data["material"], mat_map)
-    cond_cat = np.column_stack([pdg_idx, mat_idx])  # (N, 2)
-
-    n_sec_raw = data["n_sec"].astype(
-        np.int64
-    )  # (N,) unclamped, for the valid-slot mask
-    # Clamp the classification label to K_MAX: the head only has K_MAX+1 classes
-    # (0..K_MAX), and truncating here mirrors the K_MAX-slot truncation already
-    # applied to sec_cont by the loader's list padding. Without this, a rare
-    # high-multiplicity step (real data goes up to ~37) hands cross_entropy
-    # an out-of-range target and CUDA asserts.
-    n_sec = np.minimum(n_sec_raw, K_MAX).astype(np.int64)  # (N,)
+    n_sec_raw = data["n_sec"].astype(np.int64)  # (N,) unclamped, for the valid-slot mask
 
     # Secondary continuous targets
     sec_E_list = data.get("sec_E_list")
     sec_dir_list = data.get("sec_dir_list")
     sec_pdg_list = data.get("sec_pdg_list")
+    if sec_E_list is not None:
+        # The loader already padded sec_*_list to some k_max (see
+        # giant.data.loader.iter_file_chunks); that padded width is
+        # authoritative over whatever this call happened to pass in, so the
+        # two can never drift apart.
+        k_max = sec_E_list.shape[1]
+
+    # Clamp the classification label to k_max: the head only has k_max+1
+    # classes (0..k_max), and truncating here mirrors the k_max-slot
+    # truncation already applied to sec_cont by the loader's list padding.
+    # Without this, a rare high-multiplicity step (real data goes up to ~37)
+    # hands cross_entropy an out-of-range target and CUDA asserts.
+    n_sec = np.minimum(n_sec_raw, k_max).astype(np.int64)  # (N,)
 
     if sec_E_list is not None and sec_dir_list is not None and sec_pdg_list is not None:
-        sec_valid = np.arange(K_MAX)[None, :] < n_sec_raw[:, None]  # (N, K_MAX)
+        sec_valid = np.arange(k_max)[None, :] < n_sec_raw[:, None]  # (N, k_max)
         sec_cont = encode_secondaries(
             sec_E_list,
             sec_dir_list,
@@ -863,7 +1004,12 @@ def build_features(
             data["pre_dir"],
             sec_pdg_list=sec_pdg_list,
             phys_only=sec_phys_only,
-        )  # (N, K_MAX, 6)
+        )  # (N, k_max, 6)
+        sec_type_idx = (
+            encode_secondary_type_idx(sec_pdg_list, sec_valid, sec_type_class_map)
+            if sec_type_class_map is not None
+            else np.zeros((len(n_sec), k_max), dtype=np.int64)
+        )
     else:
         # Guard against silently training Stage 2 on zeroed targets: if any step
         # actually spawned secondaries (n_sec > 0, from child_track_ids) but the
@@ -885,14 +1031,15 @@ def build_features(
                 "require_secondaries=False for Stage-1-only use."
             )
         N = len(n_sec)
-        sec_cont = np.zeros((N, K_MAX, 6), dtype=np.float32)
+        sec_cont = np.zeros((N, k_max, 6), dtype=np.float32)
+        sec_type_idx = np.zeros((N, k_max), dtype=np.int64)
 
     if fit:
         cond_normalizer = Normalizer().fit(cond_cont)
         target_normalizer = Normalizer().fit(target_s1)
 
     if cond_normalizer is not None:
-        cond_cont = cond_normalizer.transform(cond_cont)
+        cond_cont = _cond_normalizer_transform(cond_cont, cond_normalizer, layout)
     if target_normalizer is not None:
         target_s1 = target_normalizer.transform(target_s1)
     if sec_phys_normalizer is not None:
@@ -907,13 +1054,14 @@ def build_features(
     else:
         proc_idx = np.zeros(len(cond_cat), dtype=np.int64)
 
-    return (
-        cond_cont,
-        cond_cat,
-        target_s1,
-        n_sec,
-        sec_cont,
-        proc_idx,
-        cond_normalizer,
-        target_normalizer,
+    return StepFeatures(
+        cond_cont=cond_cont,
+        cond_cat=cond_cat,
+        target_s1=target_s1,
+        n_sec=n_sec,
+        sec_cont=sec_cont,
+        proc_idx=proc_idx,
+        sec_type_idx=sec_type_idx,
+        cond_normalizer=cond_normalizer,
+        target_normalizer=target_normalizer,
     )

giant/geometry.py

@@ -31,10 +31,7 @@ import numpy as np
 import pandas as pd
 import pyarrow.parquet as pq
 
-_INSTALL_HINT = (
-    "the geometry oracle needs scikit-learn — install it with "
-    "`uv sync --extra cpu --extra geometry`"
-)
+_INSTALL_HINT = "the geometry oracle needs scikit-learn — install it with `uv sync --extra cpu --extra geometry`"
 
 
 def _require_sklearn():
@@ -63,9 +60,7 @@ class _SlabLookup:
     layer_ids: np.ndarray  # (n_segments,) int64, layer_id of each segment
     radius_max: float  # largest transverse radius seen in training data
 
-    def query(
-        self, pos: np.ndarray, margin: float
-    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    def query(self, pos: np.ndarray, margin: float) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
         other = [i for i in range(3) if i != self.axis]
         z = pos[:, self.axis]
         radius = np.sqrt(pos[:, other[0]] ** 2 + pos[:, other[1]] ** 2)
@@ -75,11 +70,7 @@ class _SlabLookup:
         material = self.materials[idx]
         layer_id = self.layer_ids[idx]
 
-        escaped = (
-            (z < self.z_edges[0] - margin)
-            | (z > self.z_edges[-1] + margin)
-            | (radius > self.radius_max + margin)
-        )
+        escaped = (z < self.z_edges[0] - margin) | (z > self.z_edges[-1] + margin) | (radius > self.radius_max + margin)
         return material, layer_id, escaped
 
 
@@ -300,16 +291,12 @@ def _fit_slab_lookup(
     """
     other = [i for i in range(3) if i != axis]
     z = pos[:, axis].astype(np.float64)
-    radius = np.sqrt(
-        pos[:, other[0]].astype(np.float64) ** 2
-        + pos[:, other[1]].astype(np.float64) ** 2
-    )
+    radius = np.sqrt(pos[:, other[0]].astype(np.float64) ** 2 + pos[:, other[1]].astype(np.float64) ** 2)
 
     z_min, z_max = float(z.min()), float(z.max())
     if z_min == z_max:
         raise ValueError(
-            "all points share the same depth-axis coordinate — pick a "
-            "different `depth_axis` or use method='knn'/'svm'"
+            "all points share the same depth-axis coordinate — pick a different `depth_axis` or use method='knn'/'svm'"
         )
     edges = np.linspace(z_min, z_max, n_bins + 1)
     bin_idx = np.clip(np.searchsorted(edges, z, side="right") - 1, 0, n_bins - 1)
@@ -344,12 +331,7 @@ def _fit_slab_lookup(
     bin_layer = bin_layer[fill_from]
 
     # Run-length-encode consecutive bins sharing a label into segments.
-    changed = (
-        np.flatnonzero(
-            (bin_material[1:] != bin_material[:-1]) | (bin_layer[1:] != bin_layer[:-1])
-        )
-        + 1
-    )
+    changed = np.flatnonzero((bin_material[1:] != bin_material[:-1]) | (bin_layer[1:] != bin_layer[:-1])) + 1
     seg_starts = np.concatenate([[0], changed])
     z_edges = np.concatenate([edges[seg_starts], edges[-1:]])
     materials = bin_material[seg_starts]
@@ -460,11 +442,7 @@ def build_geometry_oracle(
     # Escape threshold from the reference point spacing. Sample a subset for the
     # median 2-NN distance (the 1st neighbour of a training point is itself).
     nn = NearestNeighbors(n_neighbors=2).fit(X)
-    probe = (
-        X
-        if len(X) <= 20_000
-        else X[np.random.default_rng(seed).choice(len(X), 20_000, replace=False)]
-    )
+    probe = X if len(X) <= 20_000 else X[np.random.default_rng(seed).choice(len(X), 20_000, replace=False)]
     d2, _ = nn.kneighbors(probe, n_neighbors=2)
     median_nn = float(np.median(d2[:, 1]))
     escape_threshold = escape_factor * median_nn

giant/materials.py

@@ -64,18 +64,10 @@ MATERIAL_PROPERTIES: dict[str, MaterialProperties] = {
     "G4_CESIUM_IODIDE": MaterialProperties(
         z_eff=54.0, a_eff=129.904539, density=4.51, x0=1.860288, lambda_int=39.305990
     ),
-    "G4_Pb": MaterialProperties(
-        z_eff=82.0, a_eff=207.216962, density=11.35, x0=0.561253, lambda_int=18.247950
-    ),
-    "G4_W": MaterialProperties(
-        z_eff=74.0, a_eff=183.841648, density=19.30, x0=0.350418, lambda_int=10.311580
-    ),
-    "G4_Cu": MaterialProperties(
-        z_eff=29.0, a_eff=63.545648, density=8.96, x0=1.435578, lambda_int=15.587940
-    ),
-    "G4_Fe": MaterialProperties(
-        z_eff=26.0, a_eff=55.845113, density=7.874, x0=1.757493, lambda_int=16.990300
-    ),
+    "G4_Pb": MaterialProperties(z_eff=82.0, a_eff=207.216962, density=11.35, x0=0.561253, lambda_int=18.247950),
+    "G4_W": MaterialProperties(z_eff=74.0, a_eff=183.841648, density=19.30, x0=0.350418, lambda_int=10.311580),
+    "G4_Cu": MaterialProperties(z_eff=29.0, a_eff=63.545648, density=8.96, x0=1.435578, lambda_int=15.587940),
+    "G4_Fe": MaterialProperties(z_eff=26.0, a_eff=55.845113, density=7.874, x0=1.757493, lambda_int=16.990300),
     "G4_BRASS": MaterialProperties(
         z_eff=30.939130,
         a_eff=68.500857,
@@ -83,9 +75,7 @@ MATERIAL_PROPERTIES: dict[str, MaterialProperties] = {
         x0=1.367465,
         lambda_int=16.947420,
     ),
-    "G4_POLYSTYRENE": MaterialProperties(
-        z_eff=3.5, a_eff=6.509339, density=1.06, x0=41.312510, lambda_int=68.749880
-    ),
+    "G4_POLYSTYRENE": MaterialProperties(z_eff=3.5, a_eff=6.509339, density=1.06, x0=41.312510, lambda_int=68.749880),
     "G4_PLASTIC_SC_VINYLTOLUENE": MaterialProperties(
         z_eff=3.368421,
         a_eff=6.219791,
@@ -108,20 +98,15 @@ MATERIAL_PROPERTIES: dict[str, MaterialProperties] = {
         x0=30392.070000,
         lambda_int=71009.500000,
     ),
-    "G4_lAr": MaterialProperties(
-        z_eff=18.0, a_eff=39.947692, density=1.396, x0=14.003440, lambda_int=85.706400
-    ),
+    "G4_lAr": MaterialProperties(z_eff=18.0, a_eff=39.947692, density=1.396, x0=14.003440, lambda_int=85.706400),
 }
 
 
-def get_material_properties(
-    name: str, table: dict[str, MaterialProperties] | None = None
-) -> MaterialProperties:
+def get_material_properties(name: str, table: dict[str, MaterialProperties] | None = None) -> MaterialProperties:
     t = MATERIAL_PROPERTIES if table is None else table
     if name not in t:
         raise UnknownMaterialError(
-            f"material {name!r} is not in giant.materials.MATERIAL_PROPERTIES "
-            f"-- add it (known: {sorted(t)})"
+            f"material {name!r} is not in giant.materials.MATERIAL_PROPERTIES -- add it (known: {sorted(t)})"
         )
     props = t[name]
     if any(v is None for v in props):
@@ -134,9 +119,7 @@ def get_material_properties(
     return props
 
 
-def material_properties_array(
-    names: np.ndarray, table: dict[str, MaterialProperties] | None = None
-) -> np.ndarray:
+def material_properties_array(names: np.ndarray, table: dict[str, MaterialProperties] | None = None) -> np.ndarray:
     """(N,) str material names -> (N, 5) float32 [z_eff, a_eff, density, x0, lambda_int]."""
     out = np.array(
         [get_material_properties(str(m), table) for m in np.asarray(names)],

giant/model/_legacy.py

@@ -0,0 +1,114 @@
+"""v0.2 -> v0.3 checkpoint migration: translates a v0.2 checkpoint's flat
+`model_config`/state dicts into the current nested shape (issues.md Issue 8;
+see also `giant._migration` and `giant.config.migrate_config`, the sibling
+config.toml migration surface — issues.md Issue 6)."""
+
+from giant._migration import V02_FIXED_FACTS, reject_legacy_router_expert_sizing
+from giant.constants import EMB_DIM, K_MAX
+
+
+def _migrate_legacy_model_config(model_config: dict) -> dict:
+    """Translate a v0.2 checkpoint's flat `model_config` (giant/pipeline.py's
+    old shape: `hidden_dim`/`n_blocks`/`emb_dim`/`dropout`/`conditioning`/
+    `router`/`mode`/... all at one level) into the nested
+    `{"pdg_vocab", "mat_vocab", "conditioning", "stage1_model",
+    "stage2_model"}` shape `build_models` expects.
+
+    Sets `stage2_model.n_sec.owner = "stage1"` so the n_sec_head weights a v0.2
+    checkpoint carries on its Stage-1 module keep loading there instead of the new
+    default location (`Stage2OneShot`) — the n_sec head was trained against Stage 1's
+    own `ConditionEncoder` output, so it has to stay attached to Stage 1's module, not
+    just be labeled as such.
+
+    Only the monolithic (non-routed) trunk shape is exercised by the step-2
+    migration test; a routed v0.2 checkpoint still builds correctly here
+    (the router config passes through), but its
+    state dict isn't covered by `migrate_legacy_state_dict` below.
+    """
+    m = model_config
+    conditioning_mode = m.get("conditioning", "embedding")
+    generator = m.get("mode", "flow")
+    hidden_dim = m.get("hidden_dim", 256)
+    n_blocks = m.get("n_blocks", 6)
+    emb_dim = m.get("emb_dim", EMB_DIM)
+    dropout = m.get("dropout", 0.1)
+    k_max = m.get("k_max", K_MAX)
+    noise_dim = m.get("noise_dim", 64)
+    router_cfg = dict(m.get("router") or {})
+    reject_legacy_router_expert_sizing(router_cfg, source="this checkpoint's model_config.router")
+    router_cfg.setdefault("enabled", False)
+
+    F = V02_FIXED_FACTS
+    cond_n_layers = F["conditioning.particle.n_layers"]  # same fact for both axes
+    return {
+        "pdg_vocab": m["pdg_vocab"],
+        "mat_vocab": m["mat_vocab"],
+        "conditioning": {
+            "out_dim": F["conditioning.out_dim"],
+            "share_stages": False,
+            "particle": {"type": conditioning_mode, "emb_dim": emb_dim, "n_layers": cond_n_layers},
+            "material": {"type": conditioning_mode, "emb_dim": emb_dim, "n_layers": cond_n_layers},
+        },
+        "stage1_model": {
+            "active": F["stage1_model.active"],
+            "generator": generator,
+            "hidden_dim": hidden_dim,
+            "n_res_blocks": n_blocks,
+            "dropout": dropout,
+            "flow": {"time_dim": F["stage1_model.flow.time_dim"]},
+            "ddpm": {"time_dim": F["stage1_model.ddpm.time_dim"]},
+            "wgan": {"noise_dim": noise_dim},
+            "router": dict(router_cfg),
+        },
+        "stage2_model": {
+            "active": F["stage2_model.active"],
+            "decoder": F["stage2_model.decoder"],
+            "generator": generator,
+            "hidden_dim": hidden_dim,
+            "n_res_blocks": n_blocks,
+            "dropout": dropout,
+            "k_max": k_max,
+            "context_dim": F["stage2_model.context_dim"],
+            "n_sec": {"mode": "head", "owner": "stage1"},
+            "particle_type": {"target": F["stage2_model.particle_type.target"]},
+            "flow": {"time_dim": F["stage2_model.flow.time_dim"]},
+            "ddpm": {"time_dim": F["stage2_model.ddpm.time_dim"]},
+            "wgan": {"noise_dim": noise_dim},
+            "router": {**router_cfg, "tie_to_stage1": False},
+        },
+    }
+
+
+def migrate_legacy_state_dict(old_stage1_sd: dict, old_stage2_sd: dict) -> tuple[dict, dict]:
+    """Remap a v0.2 checkpoint's (`DenoisingMLP`-or-`WGANGenerator`,
+    `SecondaryDecoder`-or-`WGANSecondaryGenerator`) state dicts onto the new
+    `(Stage1Model, Stage2OneShot)` module structure produced by
+    `build_models(_migrate_legacy_model_config(model_config))`.
+
+    Only the monolithic (non-routed) trunk shape is handled.
+    """
+
+    def _trunk_prefix(k: str) -> str:
+        if k.startswith(("input_proj.", "blocks.", "out_proj.")):
+            return f"trunk.{k}"
+        return k
+
+    new_stage1 = {}
+    for k, v in old_stage1_sd.items():
+        if k.startswith("n_sec_head."):
+            new_stage1[k] = v  # stays top-level (n_sec.owner="stage1")
+        else:
+            new_stage1[_trunk_prefix(k)] = v
+
+    new_stage2 = {}
+    for k, v in old_stage2_sd.items():
+        if k.startswith("cond_enc.base."):
+            new_stage2["cond_enc." + k[len("cond_enc.base.") :]] = v
+        elif k.startswith("cond_enc.stage1_proj."):
+            new_stage2["context_adapter.proj." + k[len("cond_enc.stage1_proj.") :]] = v
+        elif k.startswith("cond_enc.fuse."):
+            new_stage2["fuse." + k[len("cond_enc.fuse.") :]] = v
+        else:
+            new_stage2[_trunk_prefix(k)] = v
+
+    return new_stage1, new_stage2

giant/model/builders.py

@@ -0,0 +1,226 @@
+"""Factories: `build_models`/`build_critics` assemble the top-level stage
+models from a config dict (issues.md Issue 8)."""
+
+import torch.nn as nn
+
+from giant.config import ConditioningConfig, Stage1ModelConfig, Stage2ModelConfig
+from giant.constants import X_DIM
+from giant.model._legacy import _migrate_legacy_model_config
+from giant.model.encoders import ConditionEncoder
+from giant.model.models import (
+    CriticModel,
+    Stage1Model,
+    Stage2Autoregressive,
+    Stage2OneShot,
+    resolve_type_n_classes,
+    stage2_trunk_sec_dim,
+)
+from giant.model.objectives import build_objective
+from giant.model.routers import Router, _build_router_from_cfg
+
+# ---------------------------------------------------------------------------
+# Factories
+# ---------------------------------------------------------------------------
+
+
+def build_models(model_config: dict) -> dict[str, nn.Module | None]:
+    """Construct `{"stage1": ..., "stage2": ...}` from a config dict — either
+    the new nested shape (has a `"stage1_model"` key, plus `"pdg_vocab"`/
+    `"mat_vocab"`/`"conditioning"` at the top level) or a v0.2 checkpoint's
+    flat `model_config`, auto-migrated via `_migrate_legacy_model_config`.
+
+    A stage is `None` in the result when that stage's `active = False`.
+    `stage2_model.router.tie_to_stage1` shares stage 1's literal `Router`
+    instance rather than building a second, independently-parameterized one
+    (v0.2's actual — probably accidental — behaviour: two routers built from
+    one config with no semantic relationship between them).
+
+    `conditioning.share_stages = true` builds one `ConditionEncoder`
+    instance here and passes it to both stages (`Stage1Model`/`Stage2OneShot`/
+    `Stage2Autoregressive`'s `cond_enc` param), instead of each stage
+    building its own — halving the conditioning parameter count and forcing a
+    common representation. `false` (default) keeps v0.2 behaviour:
+    independent instances with identical config but independent weights.
+    """
+    cfg = model_config if "stage1_model" in model_config else _migrate_legacy_model_config(model_config)
+    pdg_vocab = cfg["pdg_vocab"]
+    mat_vocab = cfg["mat_vocab"]
+    conditioning_cfg = ConditioningConfig.from_dict(cfg["conditioning"])
+    particle_cfg = conditioning_cfg.particle
+    material_cfg = conditioning_cfg.material
+    particle_conditioning = particle_cfg.type
+    s1_spec = Stage1ModelConfig.from_dict(cfg["stage1_model"])
+    s2_spec = Stage2ModelConfig.from_dict(cfg["stage2_model"])
+    cond_out_dim = conditioning_cfg.out_dim
+    shared_cond_enc: ConditionEncoder | None = None
+    if conditioning_cfg.share_stages:
+        shared_cond_enc = ConditionEncoder(pdg_vocab, mat_vocab, particle_cfg, material_cfg, out_dim=cond_out_dim)
+
+    result: dict[str, nn.Module | None] = {"stage1": None, "stage2": None}
+
+    stage1_router: Router | None = None
+    if s1_spec.active:
+        router_cfg = cfg["stage1_model"].get("router") or {}
+        if s1_spec.router.enabled:
+            stage1_router = _build_router_from_cfg(router_cfg, pdg_vocab, mat_vocab, particle_conditioning)
+        generator = s1_spec.generator
+        objective = build_objective(generator)
+        # wgan has no time_dim concept (no diffusion/flow time variable) —
+        # matches the pre-dataclass .get("time_dim", 64) fallback, which
+        # always hit its default for a wgan sub-block too.
+        time_dim = getattr(s1_spec, generator).time_dim if objective.needs_time else 64
+        n_sec_owner = s2_spec.n_sec.owner
+        n_sec_head_k_max = s2_spec.k_max if n_sec_owner == "stage1" else None
+        result["stage1"] = Stage1Model(
+            pdg_vocab=pdg_vocab,
+            mat_vocab=mat_vocab,
+            particle_cfg=particle_cfg,
+            material_cfg=material_cfg,
+            hidden_dim=s1_spec.hidden_dim,
+            n_res_blocks=s1_spec.n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            dropout=s1_spec.dropout,
+            generator=generator,
+            time_dim=time_dim,
+            noise_dim=s1_spec.wgan.noise_dim,
+            router=stage1_router,
+            trunk_type=s1_spec.trunk.type,
+            block_conditioning=s1_spec.trunk.block_conditioning,
+            n_sec_head_k_max=n_sec_head_k_max,
+            cond_enc=shared_cond_enc,
+            n_sec_head_cfg=s1_spec.heads.n_sec.to_dict(),
+        )
+
+    if s2_spec.active:
+        decoder = s2_spec.decoder
+        router_cfg = cfg["stage2_model"].get("router") or {}
+        stage2_router: Router | None = None
+        if s2_spec.router.enabled:
+            if s2_spec.router.tie_to_stage1 and stage1_router is not None:
+                stage2_router = stage1_router
+            else:
+                stage2_router = _build_router_from_cfg(router_cfg, pdg_vocab, mat_vocab, particle_conditioning)
+        generator = s2_spec.generator
+        objective = build_objective(generator)
+        # wgan has no time_dim concept — see the matching comment in stage 1
+        # above.
+        time_dim = getattr(s2_spec, generator).time_dim if objective.needs_time else 64
+        n_sec_owner = s2_spec.n_sec.owner
+        k_max = s2_spec.k_max
+        particle_type_cfg = s2_spec.particle_type
+
+        if decoder == "autoregressive":
+            ar_cfg = s2_spec.autoregressive
+            result["stage2"] = Stage2Autoregressive(
+                pdg_vocab=pdg_vocab,
+                mat_vocab=mat_vocab,
+                particle_cfg=particle_cfg,
+                material_cfg=material_cfg,
+                hidden_dim=s2_spec.hidden_dim,
+                n_res_blocks=s2_spec.n_res_blocks,
+                cond_out_dim=cond_out_dim,
+                context_dim=s2_spec.context_dim,
+                dropout=s2_spec.dropout,
+                generator=generator,
+                time_dim=time_dim,
+                noise_dim=s2_spec.wgan.noise_dim,
+                k_max=k_max,
+                router=stage2_router,
+                trunk_type=s2_spec.trunk.type,
+                block_conditioning=s2_spec.trunk.block_conditioning,
+                build_n_sec_head=n_sec_owner != "stage1",
+                particle_type_cfg=particle_type_cfg,
+                history=ar_cfg.history,
+                attn_n_heads=ar_cfg.attn_n_heads,
+                attn_n_layers=ar_cfg.attn_n_layers,
+                cond_enc=shared_cond_enc,
+                n_sec_head_cfg=s2_spec.heads.n_sec.to_dict(),
+                type_head_cfg=s2_spec.heads.type.to_dict(),
+            )
+        else:
+            sec_dim = stage2_trunk_sec_dim(
+                particle_type_cfg, generator, k_max, resolve_type_n_classes(particle_type_cfg, particle_cfg.emb_dim)
+            )
+            result["stage2"] = Stage2OneShot(
+                pdg_vocab=pdg_vocab,
+                mat_vocab=mat_vocab,
+                particle_cfg=particle_cfg,
+                material_cfg=material_cfg,
+                hidden_dim=s2_spec.hidden_dim,
+                n_res_blocks=s2_spec.n_res_blocks,
+                cond_out_dim=cond_out_dim,
+                context_dim=s2_spec.context_dim,
+                sec_dim=sec_dim,
+                dropout=s2_spec.dropout,
+                generator=generator,
+                time_dim=time_dim,
+                noise_dim=s2_spec.wgan.noise_dim,
+                k_max=k_max,
+                router=stage2_router,
+                trunk_type=s2_spec.trunk.type,
+                block_conditioning=s2_spec.trunk.block_conditioning,
+                build_n_sec_head=n_sec_owner != "stage1",
+                particle_type_cfg=particle_type_cfg,
+                cond_enc=shared_cond_enc,
+                n_sec_head_cfg=s2_spec.heads.n_sec.to_dict(),
+                type_head_cfg=s2_spec.heads.type.to_dict(),
+            )
+
+    return result
+
+
+def build_critics(model_config: dict) -> dict[str, nn.Module | None]:
+    """Construct `{"stage1": ..., "stage2": ...}` critics for `generator =
+    "wgan"` training. Training-only — never persisted for inference the way
+    `build_models`'s pair is. `None` for a stage that's inactive or not
+    WGAN."""
+    cfg = model_config if "stage1_model" in model_config else _migrate_legacy_model_config(model_config)
+    pdg_vocab = cfg["pdg_vocab"]
+    mat_vocab = cfg["mat_vocab"]
+    conditioning_cfg = ConditioningConfig.from_dict(cfg["conditioning"])
+    particle_cfg = conditioning_cfg.particle
+    material_cfg = conditioning_cfg.material
+    cond_out_dim = conditioning_cfg.out_dim
+    s1_spec = Stage1ModelConfig.from_dict(cfg["stage1_model"])
+    s2_spec = Stage2ModelConfig.from_dict(cfg["stage2_model"])
+
+    result: dict[str, nn.Module | None] = {"stage1": None, "stage2": None}
+
+    if s1_spec.active and build_objective(s1_spec.generator).is_adversarial:
+        result["stage1"] = CriticModel(
+            pdg_vocab=pdg_vocab,
+            mat_vocab=mat_vocab,
+            particle_cfg=particle_cfg,
+            material_cfg=material_cfg,
+            in_dim=X_DIM,
+            hidden_dim=s1_spec.wgan.critic_hidden_dim or s1_spec.hidden_dim,
+            n_res_blocks=s1_spec.wgan.critic_n_res_blocks or s1_spec.n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            dropout=s1_spec.dropout,
+            stage="stage1",
+        )
+
+    if s2_spec.active and build_objective(s2_spec.generator).is_adversarial:
+        k_max = s2_spec.k_max
+        particle_type_cfg = s2_spec.particle_type
+        in_dim = stage2_trunk_sec_dim(
+            particle_type_cfg,
+            s2_spec.generator,
+            k_max,
+            resolve_type_n_classes(particle_type_cfg, particle_cfg.emb_dim),
+        )
+        result["stage2"] = CriticModel(
+            pdg_vocab=pdg_vocab,
+            mat_vocab=mat_vocab,
+            particle_cfg=particle_cfg,
+            material_cfg=material_cfg,
+            in_dim=in_dim,
+            hidden_dim=s2_spec.wgan.critic_hidden_dim or s2_spec.hidden_dim,
+            n_res_blocks=s2_spec.wgan.critic_n_res_blocks or s2_spec.n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            dropout=s2_spec.dropout,
+            stage="stage2",
+            context_dim=s2_spec.context_dim,
+        )
+
+    return result

giant/model/encoders.py

@@ -0,0 +1,101 @@
+"""Conditioning encoder — fuses continuous conditioning with particle/material
+identity (issues.md Issue 8)."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from giant.cond_layout import CondLayout
+from giant.config import ConditioningAxisConfig
+from giant.constants import COND_DIM, COND_DIM_BASE, MATERIAL_PHYS_DIM, PARTICLE_PHYS_DIM
+from giant.model.layers import _make_axis_mlp
+
+
+class ConditionEncoder(nn.Module):
+    """Fuses continuous conditioning with particle/material identity.
+
+    The particle and material axes are configured independently
+    (`particle_cfg`/`material_cfg`, each a `ConditioningAxisConfig`) and may
+    mix freely, e.g. material "physical" with particle "embedding". Three
+    modes per axis:
+      - "embedding": a learned `nn.Embedding` lookup, indexed by `cond_cat`'s
+        dense training-vocab index. Memorizes the training menu.
+      - "physical": an `n_layers`-deep MLP over the axis's raw physical
+        properties (already present in `cond_cont`'s physical block — see
+        giant.data.transforms.build_features), computable for any PDG code /
+        material name rather than only ones seen in training.
+      - "onehot": a fixed, unlearned one-hot vector over a top-N-plus-other
+        class map (`giant.data.loader.build_topn_map_from_files`/
+        `build_pdg_topn_map_from_files`), read from `cond_cat`'s extra
+        top-N-index column(s).
+
+    Every column index/slice comes from `self.layout`
+    (`giant.cond_layout.CondLayout`), the same object the feature builders
+    lay the arrays out with, so the two sides cannot drift apart.
+    """
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        cont_dim: int = COND_DIM,
+        out_dim: int = 128,
+    ) -> None:
+        super().__init__()
+        self.particle_cfg = particle_cfg
+        self.material_cfg = material_cfg
+        # Also validates both axis types — an unknown one raises here.
+        self.layout = CondLayout.from_types(particle_cfg.type, material_cfg.type)
+
+        p_type = particle_cfg.type
+        p_emb_dim = particle_cfg.emb_dim
+        if p_type == "embedding":
+            self.pdg_emb = nn.Embedding(pdg_vocab, p_emb_dim)
+        elif p_type == "physical":
+            self.particle_mlp = _make_axis_mlp(PARTICLE_PHYS_DIM, p_emb_dim, particle_cfg.n_layers)
+
+        m_type = material_cfg.type
+        m_emb_dim = material_cfg.emb_dim
+        if m_type == "embedding":
+            self.mat_emb = nn.Embedding(mat_vocab, m_emb_dim)
+        elif m_type == "physical":
+            self.material_mlp = _make_axis_mlp(MATERIAL_PHYS_DIM, m_emb_dim, material_cfg.n_layers)
+
+        in_dim = COND_DIM_BASE + p_emb_dim + m_emb_dim
+        self.mlp = nn.Sequential(
+            nn.Linear(in_dim, out_dim),
+            nn.SiLU(),
+            nn.Linear(out_dim, out_dim),
+        )
+
+    def _particle_embed(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor):
+        p_type = self.particle_cfg.type
+        if p_type == "embedding":
+            return self.pdg_emb(cond_cat[:, self.layout.PDG_COL])
+        if p_type == "physical":
+            return self.particle_mlp(cond_cont[:, self.layout.particle_phys])
+        assert self.layout.particle_topn_col is not None
+        return F.one_hot(
+            cond_cat[:, self.layout.particle_topn_col],
+            num_classes=self.particle_cfg.emb_dim,
+        ).float()
+
+    def _material_embed(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor):
+        m_type = self.material_cfg.type
+        if m_type == "embedding":
+            return self.mat_emb(cond_cat[:, self.layout.MAT_COL])
+        if m_type == "physical":
+            return self.material_mlp(cond_cont[:, self.layout.material_phys])
+        assert self.layout.material_topn_col is not None
+        return F.one_hot(
+            cond_cat[:, self.layout.material_topn_col],
+            num_classes=self.material_cfg.emb_dim,
+        ).float()
+
+    def forward(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        pdg_e = self._particle_embed(cond_cont, cond_cat)
+        mat_e = self._material_embed(cond_cont, cond_cat)
+        x = torch.cat([cond_cont[:, self.layout.base], pdg_e, mat_e], dim=-1)
+        return self.mlp(x)

giant/model/history.py

@@ -0,0 +1,200 @@
+"""History encoders — stage-2 autoregressive only. Self-contained, no
+dependency on any other `giant.model` submodule (issues.md Issue 8), except
+for the `HISTORY_REGISTRY`/`build_history` factory, which mirrors
+`giant.model.routers`'s `Router`/`ROUTER_REGISTRY` pattern (gitea #35)."""
+
+import inspect
+
+import torch
+import torch.nn as nn
+
+
+class HistoryEncoder(nn.Module):
+    """Interface for stage-2 autoregressive per-token history summaries:
+    `forward(feat, has_prev) -> (B, K, out_dim)`, a single parallel pass over
+    a full (teacher-forced) token sequence — used by training. `MarkovHistory`
+    and `AttentionHistory` are the two registered implementations (see
+    `HISTORY_REGISTRY`/`build_history`). Inference (`giant/sample.py`)
+    generates one token at a time and cannot afford `forward`'s per-step cost
+    to be O(K) (attention would then be O(K^2) over a rollout's k_max loop),
+    so this interface also declares `init_cache`/`step` for that incremental
+    path, with working O(1) defaults here (`init_cache` -> `None`, `step` ->
+    one `forward` call ignoring `cache`) — correct for any encoder whose
+    per-step cost is already O(1) (i.e. it only ever looks at the previous
+    token, not the full prefix), which is what `MarkovHistory` relies on.
+    `AttentionHistory` overrides both with real incremental-cache versions,
+    since its `forward` genuinely needs the full prefix."""
+
+    def forward(self, feat: torch.Tensor, has_prev: torch.Tensor) -> torch.Tensor:
+        raise NotImplementedError
+
+    def init_cache(self) -> object:
+        return None
+
+    def step(self, feat: torch.Tensor, has_prev: torch.Tensor, cache: object) -> tuple[torch.Tensor, object]:
+        return self.forward(feat, has_prev), cache
+
+
+HISTORY_REGISTRY: dict[str, type[HistoryEncoder]] = {}
+
+
+def register_history(name: str):
+    def decorator(cls: type[HistoryEncoder]) -> type[HistoryEncoder]:
+        HISTORY_REGISTRY[name] = cls
+        return cls
+
+    return decorator
+
+
+def build_history(name: str, in_dim: int, out_dim: int, **kwargs) -> HistoryEncoder:
+    """Factory: look up a `HistoryEncoder` subclass by name from the registry.
+
+    Every registered history type is fed the same `stage2_model.autoregressive`
+    kwargs; kwargs not declared by that type's constructor are silently
+    dropped, so per-type hyperparameters (e.g. `AttentionHistory`'s
+    `n_heads`/`n_layers`) can coexist in one config without special-casing —
+    mirrors `giant.model.routers.build_router`.
+    """
+    if name not in HISTORY_REGISTRY:
+        raise ValueError(f"unknown history type {name!r}; available: {sorted(HISTORY_REGISTRY)}")
+    cls = HISTORY_REGISTRY[name]
+    accepted = set(inspect.signature(cls.__init__).parameters) - {"self", "in_dim", "out_dim"}
+    filtered = {k: v for k, v in kwargs.items() if k in accepted}
+    return cls(in_dim, out_dim, **filtered)
+
+
+@register_history("markov")
+class MarkovHistory(HistoryEncoder):
+    """Summarizes the previous secondary's own `(energy_fraction, direction,
+    type_representation)` through one small MLP — the "markov" history:
+    token i+1 only ever sees token i plus the running scalars
+    (`remaining_frac`/`slot_idx`, fused in separately by
+    `Stage2Autoregressive._token_cond`), not the full prefix.
+
+    At slot 0 (`has_prev` False) substitutes a learned start vector rather
+    than zeros — a reasonable default.
+    """
+
+    def __init__(self, in_dim: int, out_dim: int) -> None:
+        super().__init__()
+        self.start = nn.Parameter(torch.zeros(in_dim))
+        self.mlp = nn.Sequential(nn.Linear(in_dim, out_dim), nn.SiLU())
+
+    def forward(self, feat: torch.Tensor, has_prev: torch.Tensor) -> torch.Tensor:
+        start = self.start.view(1, 1, -1).expand_as(feat)
+        x = torch.where(has_prev.unsqueeze(-1), feat, start)
+        return self.mlp(x)
+
+
+class _CausalAttnBlock(nn.Module):
+    """One pre-norm causal self-attention block for `AttentionHistory`.
+
+    Exposes two forward paths that must agree (see
+    `test_attention_history_step_matches_forward` in `tests/test_network.py`):
+    `forward` — the full-sequence, causally-masked pass used for training;
+    `step` — an incremental pass for inference, given the *pre-attention*
+    normalized hidden states of every earlier position (`kv_cache`, i.e.
+    `norm1(x)` for positions `< t`, not `x` itself). Caching `norm1(x)` rather
+    than raw `x` is what makes `step` correct: this block's attention needs
+    exactly that quantity as keys/values, and `LayerNorm` has no cross-position
+    interaction, so recomputing it per position instead of caching it would
+    still be correct but pointlessly repeat work. The *next* block's cache is
+    built from a different sequence (this block's output), so each block owns
+    an independent cache entry.
+    """
+
+    def __init__(self, dim: int, n_heads: int, dropout: float = 0.0) -> None:
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim)
+        self.attn = nn.MultiheadAttention(dim, n_heads, dropout=dropout, batch_first=True)
+        self.norm2 = nn.LayerNorm(dim)
+        self.mlp = nn.Sequential(nn.Linear(dim, 4 * dim), nn.GELU(), nn.Linear(4 * dim, dim))
+
+    def forward(self, x: torch.Tensor, causal_mask: torch.Tensor) -> torch.Tensor:
+        h = self.norm1(x)
+        attn_out, _ = self.attn(h, h, h, attn_mask=causal_mask, need_weights=False)
+        x = x + attn_out
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+    def step(self, x_new: torch.Tensor, kv_cache: torch.Tensor | None) -> tuple[torch.Tensor, torch.Tensor]:
+        """`x_new`: `(B, 1, dim)`, this position's input. `kv_cache`: `None`
+        (first position) or `(B, T, dim)` — `norm1(x)` of every earlier
+        position at this same block. Returns `(out, new_kv_cache)`, `out`
+        being this position's block output (`(B, 1, dim)`, to feed the next
+        block's `step`), `new_kv_cache` the same cache extended by this
+        position (to reuse at this block's *next* `step` call)."""
+        h_new = self.norm1(x_new)
+        kv = h_new if kv_cache is None else torch.cat([kv_cache, h_new], dim=1)
+        attn_out, _ = self.attn(h_new, kv, kv, need_weights=False)
+        x = x_new + attn_out
+        x = x + self.mlp(self.norm2(x))
+        return x, kv
+
+
+@register_history("attention")
+class AttentionHistory(HistoryEncoder):
+    """Causal self-attention over the emitted-token prefix — the more
+    expressive alternative to `MarkovHistory`'s fixed previous-token-only
+    summary. `feat`/`has_prev`
+    follow the same shifted-by-one convention `MarkovHistory` and
+    `Stage2Autoregressive._token_cond` use: `feat[:, i]` is token `i - 1`'s
+    own `(energy_fraction, direction, type_representation)`, with a learned
+    start vector substituted at `has_prev == False` positions (only slot 0 in
+    practice — see `giant.training.stage2_inputs._ar_has_prev`). Causal masking then makes
+    position `i`'s output a function of `feat[:, 1:i+1]` — i.e. tokens
+    `0..i-1` — exactly the prefix available when predicting token `i`.
+
+    `forward` is the parallel training path (one pass over the whole
+    teacher-forced sequence); `init_cache`/`step` are the incremental
+    inference path `giant/sample.py` uses, one new token per call, to avoid
+    re-encoding the whole prefix from scratch every slot — `step` must be
+    called exactly once per slot (its cache-extension is not idempotent),
+    so a slot's output must be reused for
+    every model call within that slot (`forward`'s ODE substeps, or a separate
+    `predict_type` call) rather than re-derived — see
+    `Stage2Autoregressive.history_step`.
+    """
+
+    def __init__(self, in_dim: int, out_dim: int, n_heads: int = 4, n_layers: int = 2) -> None:
+        super().__init__()
+        self.start = nn.Parameter(torch.zeros(in_dim))
+        self.in_proj = nn.Linear(in_dim, out_dim)
+        self.blocks = nn.ModuleList([_CausalAttnBlock(out_dim, n_heads) for _ in range(n_layers)])
+
+    def _embed(self, feat: torch.Tensor, has_prev: torch.Tensor) -> torch.Tensor:
+        start = self.start.view(1, 1, -1).expand_as(feat)
+        x = torch.where(has_prev.unsqueeze(-1), feat, start)
+        return self.in_proj(x)
+
+    def forward(self, feat: torch.Tensor, has_prev: torch.Tensor) -> torch.Tensor:
+        B, K, _ = feat.shape
+        x = self._embed(feat, has_prev)
+        mask = nn.Transformer.generate_square_subsequent_mask(K, device=feat.device)
+        for block in self.blocks:
+            x = block(x, mask)
+        return x
+
+    def init_cache(self) -> list[torch.Tensor | None]:
+        return [None for _ in self.blocks]
+
+    def step(
+        self,
+        feat: torch.Tensor,
+        has_prev: torch.Tensor,
+        cache: object,
+    ) -> tuple[torch.Tensor, object]:
+        """`feat`/`has_prev`: `(B, 1, in_dim)`/`(B, 1)` — the newest token's
+        own features (what would be `feat[:, k]` in `forward`). `cache`: the
+        `list[Tensor | None]` from `init_cache`/a previous `step` call (typed
+        `object` here to match `HistoryEncoder.step`'s base signature).
+        Advances every block's cache by this position and returns this
+        position's output (`(B, 1, out_dim)`, the correct history summary for
+        the NEXT slot) plus the updated cache."""
+        assert isinstance(cache, list)
+        x = self._embed(feat, has_prev)
+        new_cache: list[torch.Tensor | None] = []
+        for block, kv in zip(self.blocks, cache):
+            x, kv_new = block.step(x, kv)
+            new_cache.append(kv_new)
+        return x, new_cache

giant/model/layers.py

@@ -0,0 +1,184 @@
+"""Small stateless-ish building blocks shared across encoders/trunks/models —
+no dependency on any other `giant.model` submodule (issues.md Issue 8)."""
+
+import math
+
+import torch
+import torch.nn as nn
+
+
+class SinusoidalEmbedding(nn.Module):
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        assert dim % 2 == 0, "dim must be even"
+        half = dim // 2
+        freqs = torch.exp(-math.log(10000) * torch.arange(half, dtype=torch.float32) / max(half - 1, 1))
+        self.register_buffer("freqs", freqs)
+
+    def forward(self, t: torch.Tensor) -> torch.Tensor:
+        t = t.reshape(-1, 1).float()
+        args = t * self.freqs.unsqueeze(0)  # (B, half)
+        return torch.cat([args.sin(), args.cos()], dim=-1)  # (B, dim)
+
+
+def _make_axis_mlp(in_dim: int, emb_dim: int, n_layers: int) -> nn.Sequential:
+    """`n_layers`-deep MLP producing an `emb_dim`-wide vector from `in_dim`
+    physical properties (`conditioning.{particle,material}.n_layers`).
+
+    `n_layers=1` (the v0.3.0 default): a single `Linear`, no hidden
+    activation. `n_layers=2` reproduces v0.2's hardcoded depth exactly —
+    `Linear -> SiLU -> Linear` — which is why `migrate_config` back-fills
+    `n_layers=2` for migrated configs rather than the v0.3 default of 1 (see
+    its docstring).
+    """
+    if n_layers < 1:
+        raise ValueError(f"n_layers must be >= 1, got {n_layers}")
+    if n_layers == 1:
+        return nn.Sequential(nn.Linear(in_dim, emb_dim))
+    layers: list[nn.Module] = [nn.Linear(in_dim, emb_dim), nn.SiLU()]
+    for _ in range(n_layers - 2):
+        layers += [nn.Linear(emb_dim, emb_dim), nn.SiLU()]
+    layers.append(nn.Linear(emb_dim, emb_dim))
+    return nn.Sequential(*layers)
+
+
+def build_mlp_head(
+    in_dim: int, out_dim: int, hidden: int, depth: int = 2, act: type[nn.Module] = nn.SiLU
+) -> nn.Sequential:
+    """`depth`-layer MLP head (gitea #36) — factors out the n_sec_head/
+    type_head pattern duplicated five times across `giant.model.models`.
+
+    `depth=1` is a bare `Linear(in_dim, out_dim)` (no hidden layer/
+    activation); `depth>=2` is `Linear(in_dim, hidden) -> act -> [Linear
+    (hidden, hidden) -> act] * (depth-2) -> Linear(hidden, out_dim)` —
+    `depth=2` reproduces every pre-#36 n_sec_head/type_head exactly when
+    `hidden == hidden_dim // 2`. Mirrors `_make_axis_mlp`'s depth
+    convention above, but takes `hidden` and `out_dim` as independent
+    widths (n_sec_head/type_head's hidden width is not their output width,
+    unlike the particle/material axis MLPs)."""
+    if depth < 1:
+        raise ValueError(f"depth must be >= 1, got {depth}")
+    if depth == 1:
+        return nn.Sequential(nn.Linear(in_dim, out_dim))
+    layers: list[nn.Module] = [nn.Linear(in_dim, hidden), act()]
+    for _ in range(depth - 2):
+        layers += [nn.Linear(hidden, hidden), act()]
+    layers.append(nn.Linear(hidden, out_dim))
+    return nn.Sequential(*layers)
+
+
+class ContextAdapter(nn.Module):
+    """Projects a stage's outcome (e.g. Stage 1's 9D target) down to a
+    fixed-width context vector for a downstream stage's conditioning —
+    `stage2_model.context_dim`. Was `SecondaryConditionEncoder.stage1_proj`
+    (+ its `tanh`) in v0.2; pulled out as its own module in v0.3.0 since
+    `SecondaryConditionEncoder` as a wrapper class disappears."""
+
+    def __init__(self, in_dim: int, context_dim: int) -> None:
+        super().__init__()
+        self.proj = nn.Linear(in_dim, context_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.tanh(self.proj(x))
+
+
+BLOCK_REGISTRY: dict[str, type[nn.Module]] = {}
+
+
+def register_block(name: str):
+    def decorator(cls: type[nn.Module]) -> type[nn.Module]:
+        BLOCK_REGISTRY[name] = cls
+        return cls
+
+    return decorator
+
+
+def build_block(name: str, dim: int, cond_dim: int, dropout: float = 0.0) -> nn.Module:
+    """Factory: look up a registered conditioning-injection block by name and
+    construct one instance — `trunk.block_conditioning` (gitea #34)."""
+    if name not in BLOCK_REGISTRY:
+        raise ValueError(f"unknown block conditioning type {name!r}; available: {sorted(BLOCK_REGISTRY)}")
+    return BLOCK_REGISTRY[name](dim, cond_dim, dropout)
+
+
+@register_block("add")
+class ResBlock(nn.Module):
+    def __init__(self, dim: int, cond_dim: int, dropout: float = 0.0) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.linear1 = nn.Linear(dim, dim)
+        self.cond_proj = nn.Linear(cond_dim, dim, bias=False)
+        self.act = nn.SiLU()
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim, dim)
+
+    def forward(self, x: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        h = self.norm(x)
+        h = self.linear1(h) + self.cond_proj(cond)
+        h = self.act(h)
+        h = self.dropout(h)
+        h = self.linear2(h)
+        return x + h
+
+
+@register_block("film")
+class FilmResBlock(nn.Module):
+    """FiLM conditioning (Perez et al. 2018): a per-channel scale+shift
+    modulates the normalized features, on top of the norm's own affine —
+    an *additional* modulation, unlike `AdaLNResBlock` below, which replaces
+    the norm's affine outright. `film_proj` is zero-initialized so
+    `gamma=beta=0` at construction — conditioning has no effect on the
+    output until training moves it, a stable starting point (though not a
+    literal identity block, since `linear1`/`linear2` aren't zero-init)."""
+
+    def __init__(self, dim: int, cond_dim: int, dropout: float = 0.0) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.linear1 = nn.Linear(dim, dim)
+        self.film_proj = nn.Linear(cond_dim, 2 * dim)
+        nn.init.zeros_(self.film_proj.weight)
+        nn.init.zeros_(self.film_proj.bias)
+        self.act = nn.SiLU()
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim, dim)
+
+    def forward(self, x: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        h = self.norm(x)
+        gamma, beta = self.film_proj(cond).chunk(2, dim=-1)
+        h = h * (1 + gamma) + beta
+        h = self.linear1(h)
+        h = self.act(h)
+        h = self.dropout(h)
+        h = self.linear2(h)
+        return x + h
+
+
+@register_block("adaln")
+class AdaLNResBlock(nn.Module):
+    """AdaLN-Zero conditioning (DiT, Peebles & Xie 2022): the norm's own
+    affine is replaced by a conditioning-derived scale/shift, and the
+    residual branch is scaled by a conditioning-derived gate. `adaln_proj`
+    is zero-initialized, so `scale=shift=gate=0` at construction — the block
+    is the exact identity function at init (`x + 0 * h' == x`), regardless
+    of `x`/`cond`."""
+
+    def __init__(self, dim: int, cond_dim: int, dropout: float = 0.0) -> None:
+        super().__init__()
+        self.norm = nn.LayerNorm(dim, elementwise_affine=False)
+        self.linear1 = nn.Linear(dim, dim)
+        self.adaln_proj = nn.Linear(cond_dim, 3 * dim)
+        nn.init.zeros_(self.adaln_proj.weight)
+        nn.init.zeros_(self.adaln_proj.bias)
+        self.act = nn.SiLU()
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim, dim)
+
+    def forward(self, x: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        h = self.norm(x)
+        scale, shift, gate = self.adaln_proj(cond).chunk(3, dim=-1)
+        h = h * (1 + scale) + shift
+        h = self.linear1(h)
+        h = self.act(h)
+        h = self.dropout(h)
+        h = self.linear2(h)
+        return x + gate * h

giant/model/models.py

@@ -0,0 +1,695 @@
+"""Top-level stage models: `Stage1Model`, `Stage2OneShot`, `Stage2Autoregressive`,
+`CriticModel` — composed from encoders/trunks/history (issues.md Issue 8)."""
+
+import torch
+import torch.nn as nn
+
+from giant.config import ConditioningAxisConfig, HeadConfig, ParticleTypeConfig
+from giant.constants import CONT_SLOT_DIM, K_MAX, PARTICLE_PHYS_DIM, SEC_DIM, SEC_SLOT_DIM, X_DIM
+from giant.model.encoders import ConditionEncoder
+from giant.model.history import HistoryEncoder, build_history
+from giant.model.layers import ContextAdapter, ResBlock, SinusoidalEmbedding, build_mlp_head
+from giant.model.objectives import build_objective
+from giant.model.routers import Router
+from giant.model.trunks import build_trunk
+
+# ---------------------------------------------------------------------------
+# Stage models
+# ---------------------------------------------------------------------------
+
+
+def resolve_type_n_classes(particle_type_cfg: ParticleTypeConfig, particle_emb_dim: int) -> int:
+    """Effective width fed to `stage2_type_dim`/`stage2_trunk_sec_dim` in
+    place of a bare `conditioning.particle.emb_dim` read. Under
+    `target = "onehot"` this is `stage2_model.particle_type.n_classes` (0 =
+    inherit `conditioning.particle.emb_dim`) — see gitea #29, which decoupled
+    the secondary-species vocabulary size from the unrelated
+    physical-conditioning MLP's output width. Under `target = "embedding"`
+    (or `"physical"`, which ignores this value entirely) `n_classes` doesn't
+    apply — the width stays `conditioning.particle.emb_dim`, the embedding
+    table's own dimensionality (`validate_config` requires
+    `conditioning.particle.type = "embedding"` here)."""
+    if particle_type_cfg.target == "onehot":
+        return particle_type_cfg.n_classes or particle_emb_dim
+    return particle_emb_dim
+
+
+def stage2_type_dim(particle_type_cfg: ParticleTypeConfig, emb_dim: int) -> int:
+    """Width of a single secondary slot's type slice —
+    `PARTICLE_PHYS_DIM` (log_mass, charge) for `target = "physical"`, else
+    `emb_dim` (both `"onehot"` class logits and `"embedding"` vectors are
+    this many classes/dims wide — callers resolve `emb_dim` via
+    `resolve_type_n_classes` first)."""
+    return PARTICLE_PHYS_DIM if particle_type_cfg.target == "physical" else emb_dim
+
+
+def stage2_trunk_sec_dim(particle_type_cfg: ParticleTypeConfig, generator: str, k_max: int, emb_dim: int) -> int:
+    """`Stage2OneShot`'s trunk output width.
+
+    `target = "physical"` is untouched from v0.2/today:
+    `k_max * SEC_SLOT_DIM`, the type slice folded into the same
+    flow-matched/WGAN vector as the continuous stick/dir slots.
+
+    `target` in `("onehot", "embedding")`: under an objective with
+    `folds_type_slice` (currently just wgan) the type slice is still folded
+    in (adversarial for onehot via ST-Gumbel, already-continuous for
+    embedding), just `emb_dim` wide instead of `PARTICLE_PHYS_DIM` wide:
+    `k_max * (CONT_SLOT_DIM + emb_dim)`. Otherwise (flow/ddpm) the type slice
+    isn't part of this vector at all — it's `Stage2OneShot.type_head`'s job
+    instead — so the trunk only covers `k_max * CONT_SLOT_DIM`.
+    """
+    if particle_type_cfg.target == "physical":
+        return k_max * SEC_SLOT_DIM
+    if build_objective(generator).folds_type_slice:
+        return k_max * (CONT_SLOT_DIM + emb_dim)
+    return k_max * CONT_SLOT_DIM
+
+
+class StageModel(nn.Module):
+    """Base owning the scaffolding common to `Stage1Model`, `Stage2OneShot`,
+    `Stage2Autoregressive` (gitea #39): build-or-share `cond_enc`,
+    `particle_type_cfg` normalisation, and — via `_build_trunk_and_heads`,
+    called by each subclass's `__init__` once its own conditioning-assembly
+    modules exist — the objective/time-embedding/trunk construction and the
+    `n_sec_head`/`type_head` classifier heads. A subclass supplies only its
+    own conditioning assembly (`Stage1Model` uses `cond_enc` directly;
+    `Stage2OneShot`/`Stage2Autoregressive` add a context-fusion path) and its
+    trunk's output width.
+
+    `cond_enc`, if given, is used in place of building a fresh
+    `ConditionEncoder` — `conditioning.share_stages = true`: `build_models`
+    constructs one shared instance and passes it to both stages, halving the
+    conditioning parameter count and forcing a common representation."""
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        cond_out_dim: int,
+        generator: str,
+        noise_dim: int,
+        k_max: int | None = None,
+        particle_type_cfg: ParticleTypeConfig | None = None,
+        cond_enc: ConditionEncoder | None = None,
+    ) -> None:
+        super().__init__()
+        self.generator_kind = generator
+        self.noise_dim = noise_dim
+        self.k_max = k_max
+        # `ParticleTypeConfig()`'s own dataclass default is target="onehot"
+        # (the config.toml default when [stage2_model.particle_type] is
+        # omitted) — a different question from "nobody passed anything to
+        # this constructor", which direct/test construction relies on
+        # defaulting to "physical" (build_models/build_critics always pass
+        # particle_type_cfg explicitly, so this sentinel is never hit there).
+        self.particle_type_cfg = (
+            particle_type_cfg if particle_type_cfg is not None else ParticleTypeConfig(target="physical")
+        )
+        self.type_dim = stage2_type_dim(
+            self.particle_type_cfg, resolve_type_n_classes(self.particle_type_cfg, particle_cfg.emb_dim)
+        )
+        self.cond_enc = (
+            cond_enc
+            if cond_enc is not None
+            else ConditionEncoder(pdg_vocab, mat_vocab, particle_cfg, material_cfg, out_dim=cond_out_dim)
+        )
+
+    def _build_trunk_and_heads(
+        self,
+        *,
+        trunk_out_dim: int,
+        hidden_dim: int,
+        n_res_blocks: int,
+        cond_out_dim: int,
+        time_dim: int,
+        router: Router | None,
+        trunk_type: str,
+        block_conditioning: str,
+        dropout: float,
+        n_sec_head_k_max: int | None,
+        n_sec_head_cfg: dict | None,
+        type_head_out_dim: int | None,
+        type_head_cfg: dict | None,
+    ) -> None:
+        """Builds `self.time_emb`, `self.trunk`, `self.n_sec_head`,
+        `self.type_head`. Called by a subclass's `__init__` after it has set
+        up its own conditioning-assembly modules — `merged_cond_dim` below
+        must match the width that assembly (`_cond_embed`/`_base_cond`/
+        `_token_cond`, or plain `cond_enc` for `Stage1Model`) actually
+        produces.
+
+        `n_sec_head` is built iff `n_sec_head_k_max is not None` (output
+        width `n_sec_head_k_max + 1`) — `Stage1Model` passes this only for a
+        migrated v0.2 checkpoint, `Stage2OneShot`/`Stage2Autoregressive` pass
+        it whenever `build_n_sec_head=True`. `type_head` is built iff
+        `type_head_out_dim is not None` (the caller — only the two Stage2
+        classes — passes `None` exactly when `particle_type_cfg.target ==
+        "physical"`) *and* the objective doesn't fold the type slice into its
+        own trunk output (checked here, since `objective` is already needed
+        for the trunk itself).
+        """
+        objective = build_objective(self.generator_kind)
+        has_time = objective.needs_time
+        self.time_emb = SinusoidalEmbedding(time_dim) if has_time else None
+        merged_cond_dim = (time_dim if has_time else 0) + cond_out_dim
+        in_dim = objective.trunk_in_dim(trunk_out_dim, self.noise_dim)
+        self.trunk = build_trunk(
+            router,
+            trunk_type,
+            in_dim,
+            trunk_out_dim,
+            hidden_dim,
+            n_res_blocks,
+            merged_cond_dim,
+            dropout,
+            block_conditioning,
+        )
+        self.n_sec_head = None
+        if n_sec_head_k_max is not None:
+            head_cfg = HeadConfig.from_dict(n_sec_head_cfg)
+            hidden = max(1, round(hidden_dim * head_cfg.hidden_ratio))
+            self.n_sec_head = build_mlp_head(cond_out_dim, n_sec_head_k_max + 1, hidden, head_cfg.depth)
+        self.type_head = None
+        if type_head_out_dim is not None and not objective.folds_type_slice:
+            head_cfg = HeadConfig.from_dict(type_head_cfg)
+            hidden = max(1, round(hidden_dim * head_cfg.hidden_ratio))
+            self.type_head = build_mlp_head(cond_out_dim, type_head_out_dim, hidden, head_cfg.depth)
+
+    def _require_n_sec_head(self) -> None:
+        if self.n_sec_head is None:
+            raise RuntimeError(
+                f"this {type(self).__name__} has no n_sec_head — it belongs to "
+                "a migrated v0.2 checkpoint (n_sec.owner='stage1'); call "
+                "stage1.predict_n_sec(cond_cont, cond_cat) instead"
+            )
+
+    def _require_type_head(self) -> None:
+        if self.type_head is None:
+            raise RuntimeError(
+                f"this {type(self).__name__} has no type_head — either "
+                "particle_type.target='physical' (the type slice is part of "
+                "forward()'s own output) or generator='wgan' (the WGAN "
+                "trainer reads the type slice out of forward()'s output "
+                "directly instead)"
+            )
+
+
+class Stage1Model(StageModel):
+    """Predicts the 9D primary post-step vector. No `n_sec_head` — fresh runs
+    move it to stage 2, except for a migrated v0.2 checkpoint
+    (`n_sec_head_k_max` given), where it stays attached here
+    since that's where its weights live and what conditioning it was trained
+    against (see `_migrate_legacy_model_config`)."""
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        hidden_dim: int = 256,
+        n_res_blocks: int = 6,
+        cond_out_dim: int = 128,
+        x_dim: int = X_DIM,
+        dropout: float = 0.0,
+        generator: str = "flow",
+        time_dim: int = 64,
+        noise_dim: int = 64,
+        router: Router | None = None,
+        trunk_type: str = "resmlp",
+        block_conditioning: str = "add",
+        n_sec_head_k_max: int | None = None,
+        cond_enc: ConditionEncoder | None = None,
+        n_sec_head_cfg: dict | None = None,
+    ) -> None:
+        super().__init__(
+            pdg_vocab,
+            mat_vocab,
+            particle_cfg,
+            material_cfg,
+            cond_out_dim=cond_out_dim,
+            generator=generator,
+            noise_dim=noise_dim,
+            cond_enc=cond_enc,
+        )
+        self._build_trunk_and_heads(
+            trunk_out_dim=x_dim,
+            hidden_dim=hidden_dim,
+            n_res_blocks=n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            time_dim=time_dim,
+            router=router,
+            trunk_type=trunk_type,
+            block_conditioning=block_conditioning,
+            dropout=dropout,
+            n_sec_head_k_max=n_sec_head_k_max,
+            n_sec_head_cfg=n_sec_head_cfg,
+            type_head_out_dim=None,
+            type_head_cfg=None,
+        )
+
+    def forward(
+        self,
+        x_t: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        t: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        c_emb = self.cond_enc(cond_cont, cond_cat)
+        cond = torch.cat([self.time_emb(t), c_emb], dim=-1) if self.time_emb is not None else c_emb
+        return self.trunk(x_t, cond, cond_cont, cond_cat)
+
+    def _require_n_sec_head(self) -> None:
+        """Overrides `StageModel`'s guard — a `Stage1Model` with no
+        `n_sec_head` points the caller to stage 2 (n_sec's default owner),
+        not to `stage1` as the base's message would."""
+        if self.n_sec_head is None:
+            raise RuntimeError(
+                "this Stage1Model has no n_sec_head — n_sec now lives on "
+                "stage 2 by default; this method only exists "
+                "for a migrated v0.2 checkpoint (n_sec.owner='stage1')"
+            )
+
+    def predict_n_sec(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """Return n_sec logits (B, K_MAX+1) from conditioning alone. Only
+        valid on a migrated v0.2 checkpoint's Stage1Model — fresh v0.3.0
+        configs predict n_sec from Stage2OneShot instead."""
+        self._require_n_sec_head()
+        assert self.n_sec_head is not None
+        c_emb = self.cond_enc(cond_cont, cond_cat)
+        return self.n_sec_head(c_emb)
+
+
+class Stage2OneShot(StageModel):
+    """Predicts all `k_max` secondary slots simultaneously — v0.2 behaviour,
+    reproduced exactly (`decoder = "autoregressive"` is `Stage2Autoregressive`,
+    step 4/5, not implemented yet).
+
+    Owns `n_sec_head` by default unless `build_n_sec_head=False`
+    (a migrated v0.2 checkpoint, whose n_sec_head instead attaches to
+    Stage1Model — see `_migrate_legacy_model_config`).
+
+    `particle_type_cfg.target` (default `"physical"`) selects the
+    secondary-type mechanism: `"physical"` keeps the type slice folded into
+    the trunk's own
+    flow-matched/WGAN output, unchanged from v0.2 (`sec_dim` — computed by
+    the caller via `stage2_trunk_sec_dim` — already reflects this). Under
+    `"onehot"`/`"embedding"` with an objective (`giant.model.objectives`) that
+    doesn't fold the type slice (flow/ddpm), the type
+    slice is predicted by a separate `type_head` instead (same shape pattern
+    as `n_sec_head`) — `sec_dim` then covers only the continuous
+    stick/dir slots, `type_head` covers `k_max * emb_dim` type logits/vectors.
+    Under a folding objective (wgan) the type slice stays folded into `sec_dim`
+    (just `emb_dim` instead of `PARTICLE_PHYS_DIM` wide) and `type_head` is
+    unused (`None`) — the WGAN trainer handles the ST-Gumbel relaxation.
+    """
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        hidden_dim: int = 256,
+        n_res_blocks: int = 6,
+        cond_out_dim: int = 128,
+        context_dim: int = 64,
+        sec_dim: int = SEC_DIM,
+        x_dim: int = X_DIM,
+        dropout: float = 0.0,
+        generator: str = "wgan",
+        time_dim: int = 64,
+        noise_dim: int = 64,
+        k_max: int = K_MAX,
+        router: Router | None = None,
+        trunk_type: str = "resmlp",
+        block_conditioning: str = "add",
+        build_n_sec_head: bool = True,
+        particle_type_cfg: ParticleTypeConfig | None = None,
+        cond_enc: ConditionEncoder | None = None,
+        n_sec_head_cfg: dict | None = None,
+        type_head_cfg: dict | None = None,
+    ) -> None:
+        super().__init__(
+            pdg_vocab,
+            mat_vocab,
+            particle_cfg,
+            material_cfg,
+            cond_out_dim=cond_out_dim,
+            generator=generator,
+            noise_dim=noise_dim,
+            k_max=k_max,
+            particle_type_cfg=particle_type_cfg,
+            cond_enc=cond_enc,
+        )
+        self.context_adapter = ContextAdapter(x_dim, context_dim)
+        self.fuse = nn.Sequential(
+            nn.Linear(cond_out_dim + context_dim, cond_out_dim),
+            nn.SiLU(),
+        )
+        target = self.particle_type_cfg.target
+        type_head_out_dim = None if target == "physical" else k_max * self.type_dim
+        self._build_trunk_and_heads(
+            trunk_out_dim=sec_dim,
+            hidden_dim=hidden_dim,
+            n_res_blocks=n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            time_dim=time_dim,
+            router=router,
+            trunk_type=trunk_type,
+            block_conditioning=block_conditioning,
+            dropout=dropout,
+            n_sec_head_k_max=k_max if build_n_sec_head else None,
+            n_sec_head_cfg=n_sec_head_cfg,
+            type_head_out_dim=type_head_out_dim,
+            type_head_cfg=type_head_cfg,
+        )
+
+    def _cond_embed(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, stage1_out: torch.Tensor) -> torch.Tensor:
+        base = self.cond_enc(cond_cont, cond_cat)
+        ctx = self.context_adapter(stage1_out)
+        return self.fuse(torch.cat([base, ctx], dim=-1))
+
+    def forward(
+        self,
+        x_t: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+        t: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        c_emb = self._cond_embed(cond_cont, cond_cat, stage1_out)
+        cond = torch.cat([self.time_emb(t), c_emb], dim=-1) if self.time_emb is not None else c_emb
+        return self.trunk(x_t, cond, cond_cont, cond_cat)
+
+    def predict_n_sec(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+    ) -> torch.Tensor:
+        self._require_n_sec_head()
+        assert self.n_sec_head is not None
+        c_emb = self._cond_embed(cond_cont, cond_cat, stage1_out)
+        return self.n_sec_head(c_emb)
+
+    def predict_type(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """`(B, k_max, emb_dim)` per-slot type logits (`target="onehot"`) or
+        vectors (`target="embedding"`) — only under `generator in ("flow",
+        "ddpm")`; `generator == "wgan"` folds the type slice into `forward`'s
+        own output instead (see class docstring)."""
+        self._require_type_head()
+        assert self.type_head is not None
+        c_emb = self._cond_embed(cond_cont, cond_cat, stage1_out)
+        return self.type_head(c_emb).view(-1, self.k_max, self.type_dim)
+
+
+class Stage2Autoregressive(StageModel):
+    """Emits secondaries one at a time in descending-energy order, instead
+    of `Stage2OneShot`'s simultaneous
+    k_max-slot prediction. `history` selects `MarkovHistory` or
+    `AttentionHistory` (`attn_n_heads`/`attn_n_layers`, attention only).
+    `teacher_forcing` handling lives entirely in the trainer
+    (`giant/train.py`), since it only affects how training inputs are
+    assembled, not this module's architecture.
+
+    Under teacher forcing every token's conditioning is built from ground
+    truth, so a whole K-token sequence trains in one parallel batched pass:
+    `forward` accepts `(B, K, ...)` tensors for an arbitrary K (not hardcoded
+    to `k_max`) — this also means a future one-token-at-a-time inference loop
+    (`K=1` per call, step 6) needs no interface change here.
+
+    Two independent conditioning paths, mirroring `Stage2OneShot`'s
+    `_cond_embed` but split in two: `_base_cond` (`cond_enc` +
+    `context_adapter` only) feeds `predict_n_sec`, since n_sec doesn't depend
+    on token position; `_token_cond` additionally fuses in the history
+    encoding and two running scalars (remaining energy-budget fraction,
+    normalized slot index), and feeds `forward`/`predict_type`/the trunk.
+    """
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        hidden_dim: int = 256,
+        n_res_blocks: int = 6,
+        cond_out_dim: int = 128,
+        context_dim: int = 64,
+        x_dim: int = X_DIM,
+        dropout: float = 0.0,
+        generator: str = "wgan",
+        time_dim: int = 64,
+        noise_dim: int = 64,
+        k_max: int = K_MAX,
+        router: Router | None = None,
+        trunk_type: str = "resmlp",
+        block_conditioning: str = "add",
+        build_n_sec_head: bool = True,
+        particle_type_cfg: ParticleTypeConfig | None = None,
+        history: str = "markov",
+        attn_n_heads: int = 4,
+        attn_n_layers: int = 2,
+        cond_enc: ConditionEncoder | None = None,
+        n_sec_head_cfg: dict | None = None,
+        type_head_cfg: dict | None = None,
+    ) -> None:
+        super().__init__(
+            pdg_vocab,
+            mat_vocab,
+            particle_cfg,
+            material_cfg,
+            cond_out_dim=cond_out_dim,
+            generator=generator,
+            noise_dim=noise_dim,
+            k_max=k_max,
+            particle_type_cfg=particle_type_cfg,
+            cond_enc=cond_enc,
+        )
+        self.history_kind = history
+        self.context_adapter = ContextAdapter(x_dim, context_dim)
+        self.base_fuse = nn.Sequential(
+            nn.Linear(cond_out_dim + context_dim, cond_out_dim),
+            nn.SiLU(),
+        )
+
+        # Reuses conditioning.out_dim for the history encoder's own output
+        # width — there's no dedicated stage2_model.autoregressive key for
+        # this, a reasonable default rather than a design-doc-specified value.
+        history_dim = cond_out_dim
+        hist_in_dim = CONT_SLOT_DIM + self.type_dim
+        self.history_encoder: HistoryEncoder = build_history(
+            history, hist_in_dim, history_dim, n_heads=attn_n_heads, n_layers=attn_n_layers
+        )
+        token_fuse_in = cond_out_dim + context_dim + history_dim + 2  # +2: remaining_frac, slot_idx
+        self.token_fuse = nn.Sequential(
+            nn.Linear(token_fuse_in, cond_out_dim),
+            nn.SiLU(),
+        )
+
+        # `self.type_dim` (set by StageModel.__init__) doubles as the raw
+        # `emb_dim` `stage2_trunk_sec_dim` wants: for a non-"physical" target
+        # `stage2_type_dim` already resolved `type_dim` to exactly that value;
+        # for "physical" the emb_dim argument goes unused anyway.
+        token_dim = stage2_trunk_sec_dim(self.particle_type_cfg, generator, 1, self.type_dim)
+        target = self.particle_type_cfg.target
+        type_head_out_dim = None if target == "physical" else self.type_dim
+        self._build_trunk_and_heads(
+            trunk_out_dim=token_dim,
+            hidden_dim=hidden_dim,
+            n_res_blocks=n_res_blocks,
+            cond_out_dim=cond_out_dim,
+            time_dim=time_dim,
+            router=router,
+            trunk_type=trunk_type,
+            block_conditioning=block_conditioning,
+            dropout=dropout,
+            n_sec_head_k_max=k_max if build_n_sec_head else None,
+            n_sec_head_cfg=n_sec_head_cfg,
+            type_head_out_dim=type_head_out_dim,
+            type_head_cfg=type_head_cfg,
+        )
+
+    def _base_cond(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, stage1_out: torch.Tensor) -> torch.Tensor:
+        base = self.cond_enc(cond_cont, cond_cat)
+        ctx = self.context_adapter(stage1_out)
+        return self.base_fuse(torch.cat([base, ctx], dim=-1))
+
+    def _token_cond(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+        history_feat: torch.Tensor,
+        has_prev: torch.Tensor,
+        remaining_frac: torch.Tensor,
+        slot_idx: torch.Tensor,
+        hist: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        """`hist`, if given, overrides recomputing `self.history_encoder`
+        from `history_feat`/`has_prev` — the inference-time KV-cache path
+        (`Stage2Autoregressive.history_step`) precomputes it once per slot and
+        passes it in here so a slot's (possibly several) model calls — an ODE
+        loop's substeps, or a separate `predict_type` call — read the same
+        cached history instead of each re-deriving (and, under attention,
+        re-appending to the cache — see `AttentionHistory.step`'s docstring)."""
+        K = history_feat.size(1)
+        base = self.cond_enc(cond_cont, cond_cat).unsqueeze(1).expand(-1, K, -1)
+        ctx = self.context_adapter(stage1_out).unsqueeze(1).expand(-1, K, -1)
+        if hist is None:
+            hist = self.history_encoder(history_feat, has_prev)
+        scalars = torch.stack([remaining_frac, slot_idx], dim=-1)
+        return self.token_fuse(torch.cat([base, ctx, hist, scalars], dim=-1))
+
+    def init_history_cache(self):
+        """Inference-only incremental-decoding state for `self.history_encoder`
+        (`giant/sample.py`'s AR loop) — whatever `self.history_encoder.init_cache()`
+        returns for the configured `history` type: `None` under `history="markov"`
+        (its per-step cost is already O(1) — see `HistoryEncoder`'s docstring),
+        or `AttentionHistory.init_cache()`'s real per-block KV cache under
+        `history="attention"`."""
+        return self.history_encoder.init_cache()
+
+    def history_step(self, token_feat: torch.Tensor, has_prev: torch.Tensor, cache) -> tuple[torch.Tensor, object]:
+        """One inference slot's worth of history encoding: advances `cache`
+        (from `init_history_cache`, or a previous `history_step` call) by
+        `token_feat`/`has_prev` (`(B, 1, ...)` — the just-emitted previous
+        token, same convention `giant.sample.sample_secondaries_ar` already
+        threads as `prev_repr`), and returns `(hist, new_cache)` — `hist` is
+        this slot's history summary (pass it as `_token_cond`'s `hist=` to
+        every model call made for this slot), `new_cache` is what to pass into
+        the *next* slot's `history_step`. Must be called exactly once per
+        slot — see `AttentionHistory.step`'s docstring."""
+        return self.history_encoder.step(token_feat, has_prev, cache)
+
+    def forward(
+        self,
+        x_t: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+        history_feat: torch.Tensor,
+        has_prev: torch.Tensor,
+        remaining_frac: torch.Tensor,
+        slot_idx: torch.Tensor,
+        t: torch.Tensor | None = None,
+        hist: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        B, K = x_t.shape[0], x_t.shape[1]
+        c_emb = self._token_cond(
+            cond_cont,
+            cond_cat,
+            stage1_out,
+            history_feat,
+            has_prev,
+            remaining_frac,
+            slot_idx,
+            hist=hist,
+        )
+        if self.time_emb is not None:
+            assert t is not None
+            t_emb = self.time_emb(t.reshape(-1)).view(B, K, -1)
+            cond = torch.cat([t_emb, c_emb], dim=-1)
+        else:
+            cond = c_emb
+        x_flat = x_t.reshape(B * K, -1)
+        cond_flat = cond.reshape(B * K, -1)
+        cond_cont_flat = cond_cont.unsqueeze(1).expand(-1, K, -1).reshape(B * K, -1)
+        cond_cat_flat = cond_cat.unsqueeze(1).expand(-1, K, -1).reshape(B * K, -1)
+        out = self.trunk(x_flat, cond_flat, cond_cont_flat, cond_cat_flat)
+        return out.view(B, K, -1)
+
+    def predict_n_sec(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, stage1_out: torch.Tensor) -> torch.Tensor:
+        self._require_n_sec_head()
+        assert self.n_sec_head is not None
+        return self.n_sec_head(self._base_cond(cond_cont, cond_cat, stage1_out))
+
+    def predict_type(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor,
+        history_feat: torch.Tensor,
+        has_prev: torch.Tensor,
+        remaining_frac: torch.Tensor,
+        slot_idx: torch.Tensor,
+        hist: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        self._require_type_head()
+        assert self.type_head is not None
+        c_emb = self._token_cond(
+            cond_cont,
+            cond_cat,
+            stage1_out,
+            history_feat,
+            has_prev,
+            remaining_frac,
+            slot_idx,
+            hist=hist,
+        )
+        B, K, _ = c_emb.shape
+        return self.type_head(c_emb.reshape(B * K, -1)).view(B, K, self.type_dim)
+
+
+class CriticModel(nn.Module):
+    """Generator-agnostic WGAN-GP critic body: a scalar realism score, for
+    either stage (`stage="stage1"` mirrors v0.2 `Critic`; `stage="stage2"`
+    mirrors v0.2 `SecondaryCritic`, adding the same context-fusion path as
+    `Stage2OneShot`). Used only when that stage's `generator == "wgan"`."""
+
+    def __init__(
+        self,
+        pdg_vocab: int,
+        mat_vocab: int,
+        particle_cfg: ConditioningAxisConfig,
+        material_cfg: ConditioningAxisConfig,
+        in_dim: int,
+        hidden_dim: int = 256,
+        n_res_blocks: int = 6,
+        cond_out_dim: int = 128,
+        dropout: float = 0.0,
+        stage: str = "stage1",
+        context_dim: int = 64,
+        context_in_dim: int = X_DIM,
+    ) -> None:
+        super().__init__()
+        if stage not in ("stage1", "stage2"):
+            raise ValueError(f"stage must be 'stage1' or 'stage2', got {stage!r}")
+        self.stage = stage
+        self.cond_enc = ConditionEncoder(pdg_vocab, mat_vocab, particle_cfg, material_cfg, out_dim=cond_out_dim)
+        if stage == "stage2":
+            self.context_adapter = ContextAdapter(context_in_dim, context_dim)
+            self.fuse = nn.Sequential(
+                nn.Linear(cond_out_dim + context_dim, cond_out_dim),
+                nn.SiLU(),
+            )
+        self.input_proj = nn.Linear(in_dim, hidden_dim)
+        self.blocks = nn.ModuleList([ResBlock(hidden_dim, cond_out_dim, dropout=dropout) for _ in range(n_res_blocks)])
+        self.out_norm = nn.LayerNorm(hidden_dim)
+        self.out_proj = nn.Linear(hidden_dim, 1)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_out: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        base = self.cond_enc(cond_cont, cond_cat)
+        if self.stage == "stage2":
+            ctx = self.context_adapter(stage1_out)
+            cond = self.fuse(torch.cat([base, ctx], dim=-1))
+        else:
+            cond = base
+        h = self.input_proj(x)
+        for block in self.blocks:
+            h = block(h, cond)
+        return self.out_proj(self.out_norm(h)).squeeze(-1)

giant/model/objectives.py

@@ -0,0 +1,202 @@
+"""Generative objectives (flow/ddpm/wgan): `Objective` base + registry,
+mirroring `giant.model.routers`'s `Router` pattern (gitea #32). Each objective
+answers, in one place, the handful of questions every stage model/sampler/
+trainer used to re-derive independently from a bare `generator` string: does
+this stage need a time embedding, is it adversarial, does it fold the
+secondary type slice into its own trunk output, what does the trunk take as
+input, which stage-1/stage-2 loss does it train against.
+
+Self-contained (no dependency on `giant.model.models`, unlike `Router` which
+`giant.model.trunks` depends on) — `Objective` never needs to construct a
+stage model or critic itself, only describe one. This also sidesteps a
+`models.py` <-> `objectives.py` import cycle, since `models.py` calls
+`build_objective`.
+"""
+
+import inspect
+
+import torch
+
+from giant.model.schedule import (
+    CosineSchedule,
+    flow_matching_loss,
+    flow_matching_loss_secondary,
+    flow_matching_loss_secondary_ar,
+)
+
+# ---------------------------------------------------------------------------
+# Objective contract
+# ---------------------------------------------------------------------------
+
+
+class Objective:
+    """Contract for a pluggable generative objective. Not an `nn.Module` —
+    unlike `Router`, no objective owns learnable parameters, so a plain
+    strategy object is the honest fit.
+
+    `needs_time`/`is_adversarial`/`folds_type_slice`/`supports_stage2_decoder`
+    are set by each concrete subclass (no defaults here — a new objective
+    should have to state all four, not silently inherit one that happens to
+    be wrong for it). See `FlowObjective`/`DdpmObjective`/`WganObjective`.
+    """
+
+    needs_time: bool
+    is_adversarial: bool
+    folds_type_slice: bool
+    supports_stage2_decoder: bool = True
+
+    def trunk_in_dim(self, out_dim: int, noise_dim: int) -> int:
+        """Width of the trunk's own input — `out_dim` (denoising/flow-matching
+        a same-shape vector) for every non-adversarial objective;
+        `WganObjective` overrides to `noise_dim` (a single-pass noise-to-output
+        generator)."""
+        return out_dim
+
+    def build_schedule(self, n_steps: int, device: torch.device) -> CosineSchedule | None:
+        """Objective-owned auxiliary state a stage trainer must build once
+        and hold onto (device-placed) across its training loop. `None` for
+        every objective except `DdpmObjective` (its noise schedule)."""
+        return None
+
+    def stage1_loss(
+        self,
+        model: torch.nn.Module,
+        x1: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        *,
+        schedule: object | None = None,
+    ) -> torch.Tensor:
+        """Stage-1 training loss. Only implemented by non-adversarial
+        objectives — `WganObjective` is unused here, `WGANStageTrainer` has
+        its own G/D step instead."""
+        raise NotImplementedError(f"{type(self).__name__} has no stage1_loss")
+
+    def stage2_loss(
+        self,
+        model: torch.nn.Module,
+        x1_s2: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_ctx: torch.Tensor,
+        sec_mask: torch.Tensor,
+        *,
+        type_dim: int | None,
+        ar_inputs: dict[str, torch.Tensor] | None = None,
+    ) -> torch.Tensor:
+        """Stage-2 secondary-decoder training loss, one-shot or
+        autoregressive depending on whether `ar_inputs` is given. Same
+        adversarial caveat as `stage1_loss`."""
+        raise NotImplementedError(f"{type(self).__name__} has no stage2_loss")
+
+
+OBJECTIVE_REGISTRY: dict[str, type[Objective]] = {}
+
+
+def register_objective(name: str):
+    def decorator(cls: type[Objective]) -> type[Objective]:
+        OBJECTIVE_REGISTRY[name] = cls
+        return cls
+
+    return decorator
+
+
+def build_objective(name: str, **kwargs) -> Objective:
+    """Factory: look up an `Objective` subclass by name (a `generator`
+    config value) from the registry.
+
+    Every registered objective is fed the same kwargs; kwargs not declared by
+    that type's constructor are silently dropped, so per-type hyperparameters
+    (e.g. `DdpmObjective`'s `n_steps`) can coexist in one call without
+    special-casing — same convention as `giant.model.routers.build_router`.
+    """
+    if name not in OBJECTIVE_REGISTRY:
+        raise ValueError(f"unknown generator/objective {name!r}; available: {sorted(OBJECTIVE_REGISTRY)}")
+    cls = OBJECTIVE_REGISTRY[name]
+    accepted = set(inspect.signature(cls.__init__).parameters) - {"self"}
+    filtered = {k: v for k, v in kwargs.items() if k in accepted}
+    return cls(**filtered)
+
+
+# ---------------------------------------------------------------------------
+# Concrete objectives
+# ---------------------------------------------------------------------------
+
+
+@register_objective("flow")
+class FlowObjective(Objective):
+    """Conditional flow matching (Lipman et al. 2022) — the primary
+    objective. ~10 ODE steps at inference (`giant.sample.sample_flow`)."""
+
+    needs_time = True
+    is_adversarial = False
+    folds_type_slice = False
+
+    def stage1_loss(self, model, x1, cond_cont, cond_cat, *, schedule=None) -> torch.Tensor:
+        return flow_matching_loss(model, x1, cond_cont, cond_cat)
+
+    def stage2_loss(
+        self,
+        model,
+        x1_s2,
+        cond_cont,
+        cond_cat,
+        stage1_ctx,
+        sec_mask,
+        *,
+        type_dim=None,
+        ar_inputs=None,
+    ) -> torch.Tensor:
+        if ar_inputs is not None:
+            return flow_matching_loss_secondary_ar(
+                model,
+                x1_s2,
+                cond_cont,
+                cond_cat,
+                stage1_ctx,
+                ar_inputs["history_feat"],
+                ar_inputs["has_prev"],
+                ar_inputs["remaining_frac"],
+                ar_inputs["slot_idx"],
+                sec_mask,
+                type_dim=type_dim,
+            )
+        return flow_matching_loss_secondary(model, x1_s2, cond_cont, cond_cat, stage1_ctx, sec_mask, type_dim=type_dim)
+
+
+@register_objective("ddpm")
+class DdpmObjective(Objective):
+    """Full DDPM ancestral sampling (Nichol & Dhariwal 2021 cosine schedule)
+    — the throwaway baseline. Stage-1 only: no `Stage2*` class has ever been
+    trained with `generator="ddpm"` in practice, so there's no stage-2 ddpm
+    loss to dispatch to (matches `FlowDDPMStageTrainer`'s pre-existing
+    stage-2 guard)."""
+
+    needs_time = True
+    is_adversarial = False
+    folds_type_slice = False
+    supports_stage2_decoder = False
+
+    def __init__(self, n_steps: int = 1000) -> None:
+        self.n_steps = n_steps
+
+    def build_schedule(self, n_steps: int, device: torch.device) -> CosineSchedule:
+        return CosineSchedule(T=n_steps).to(device)
+
+    def stage1_loss(self, model, x1, cond_cont, cond_cat, *, schedule=None) -> torch.Tensor:
+        assert schedule is not None, "DdpmObjective.stage1_loss needs a schedule (see build_schedule)"
+        return schedule.loss(model, x1, cond_cont, cond_cat)
+
+
+@register_objective("wgan")
+class WganObjective(Objective):
+    """WGAN-GP (Gulrajani et al. 2017) — single forward pass instead of an
+    ODE loop. `stage1_loss`/`stage2_loss` are unused: `WGANStageTrainer` owns
+    its own dual generator/critic step instead of a single scalar loss."""
+
+    needs_time = False
+    is_adversarial = True
+    folds_type_slice = True
+
+    def trunk_in_dim(self, out_dim: int, noise_dim: int) -> int:
+        return noise_dim

giant/model/routers.py

@@ -0,0 +1,376 @@
+"""Mixture-of-experts routing: `Router` base + registry, the four concrete
+router types, and composed/config-driven construction — self-contained, no
+dependency on any other `giant.model` submodule (issues.md Issue 8)."""
+
+import inspect
+import math
+import re
+from collections.abc import Sequence
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from giant.cond_layout import CondLayout
+from giant.constants import COND_DIM
+
+# ---------------------------------------------------------------------------
+# Routers — carried over unchanged from v0.2
+# ---------------------------------------------------------------------------
+
+
+class Router(nn.Module):
+    """Contract for a pluggable mixture-of-experts routing axis.
+
+    Subclasses implement `gate` (soft partition-of-unity weights over
+    experts, used in train mode for a fully differentiable mixture);
+    `top1` and `balance_loss` have working defaults so a new routing axis
+    is usually a one-method add. See `ROUTER_REGISTRY` / `build_router`.
+    """
+
+    def __init__(self, n_experts: int) -> None:
+        super().__init__()
+        self.n_experts = n_experts
+        self.gumbel = False
+        self.gumbel_tau = 1.0
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """(B, n_experts) soft weights, rows summing to 1."""
+        raise NotImplementedError
+
+    def combine_weights(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """(B, n_experts) train-time expert-combination weights.
+
+        Default (`gumbel=False`): identical to `gate()`. Opt-in
+        straight-through Gumbel-softmax (`gumbel=True`, train mode only):
+        hardens the forward pass to a one-hot sample (matching eval-time
+        top-1 dispatch) while keeping the soft sample's gradient on backward.
+        """
+        probs = self.gate(cond_cont, cond_cat)
+        if not (self.gumbel and self.training):
+            return probs
+        log_probs = torch.log(probs.clamp_min(1e-8))
+        return F.gumbel_softmax(log_probs, tau=self.gumbel_tau, hard=True, dim=-1)
+
+    def top1(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """(B,) hard expert index, used for eval-time grouped dispatch."""
+        return self.gate(cond_cont, cond_cat).argmax(dim=-1)
+
+    def balance_loss(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """Importance CV^2 load-balancing auxiliary loss (Shazeer et al. 2017)."""
+        importance = self.gate(cond_cont, cond_cat).sum(dim=0)  # (n_experts,)
+        return (importance.std() / (importance.mean() + 1e-8)) ** 2
+
+    def classify_loss(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+        """Optional supervised auxiliary loss shaping the router's own belief.
+
+        Default: none (a scalar 0). Routers gating on an unobservable
+        pre-step quantity (e.g. ProcessRouter) override this.
+        """
+        return torch.zeros((), device=cond_cont.device)
+
+    def entropy_loss(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        """Optional auxiliary loss rewarding sharper (lower-entropy) routing."""
+        norm_entropy, _ = self.gate_stats(cond_cont, cond_cat)
+        return norm_entropy
+
+    def gate_stats(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        """Diagnostics: `(norm_entropy, importance)` — see v0.2 docstring for
+        the full explanation, unchanged in v0.3.0."""
+        gate = self.gate(cond_cont, cond_cat)  # (B, n_experts)
+        row_entropy = -(gate * (gate + 1e-8).log()).sum(dim=-1)  # (B,)
+        norm_entropy = row_entropy.mean() / math.log(self.n_experts)
+        importance = gate.sum(dim=0)  # (n_experts,)
+        return norm_entropy, importance
+
+
+ROUTER_REGISTRY: dict[str, type[Router]] = {}
+
+
+def register_router(name: str):
+    def decorator(cls: type[Router]) -> type[Router]:
+        ROUTER_REGISTRY[name] = cls
+        return cls
+
+    return decorator
+
+
+def build_router(name: str, n_experts: int, **kwargs) -> Router:
+    """Factory: look up a `Router` subclass by name from the registry.
+
+    Every registered router type is fed the same `router` config dict;
+    kwargs not declared by that type's constructor are silently dropped, so
+    per-type hyperparameters (e.g. EnergyRouter's `temperature`) can coexist
+    in one config without special-casing.
+    """
+    if name not in ROUTER_REGISTRY:
+        raise ValueError(f"unknown router type {name!r}; available: {sorted(ROUTER_REGISTRY)}")
+    cls = ROUTER_REGISTRY[name]
+    accepted = set(inspect.signature(cls.__init__).parameters) - {"self", "n_experts"}
+    filtered = {k: v for k, v in kwargs.items() if k in accepted}
+    return cls(n_experts=n_experts, **filtered)
+
+
+def _bounded_interp(raw: torch.Tensor, lo: float, hi: float) -> torch.Tensor:
+    """Sigmoid interpolation into `[lo, hi]` — smooth, always-positive-gradient
+    bound used for EnergyRouter's `learn_width`/`learn_temperature` modes."""
+    return lo + (hi - lo) * torch.sigmoid(raw)
+
+
+def _inverse_bounded_interp(value: float, lo: float, hi: float) -> float:
+    """Inverse of `_bounded_interp`, used once at construction to warm-start
+    `raw` so the initial effective width/temperature exactly equals `value`."""
+    p = min(max((value - lo) / (hi - lo), 1e-6), 1 - 1e-6)
+    return math.log(p / (1 - p))
+
+
+@register_router("energy")
+class EnergyRouter(Router):
+    """Soft turn-on gate over normalized pre-step log-energy.
+
+    Reads `cond_cont[:, energy_idx]` (ignores cond_cat). `gate(e) =
+    softmax_i(-(e - c_i)^2 / tau)`; as tau -> 0 this hardens to
+    nearest-center (Voronoi) selection, exactly what `top1` uses at eval.
+    """
+
+    def __init__(
+        self,
+        n_experts: int = 4,
+        temperature: float = 0.5,
+        learn_centers: bool = True,
+        energy_idx: int = 3,
+        centers_init: Sequence[float] | None = None,
+        learn_width: bool = False,
+        learn_temperature: bool = False,
+        width_min_ratio: float = 0.1,
+        width_max_ratio: float = 10.0,
+    ) -> None:
+        super().__init__(n_experts)
+        if learn_width and learn_temperature:
+            raise ValueError("learn_width and learn_temperature are mutually exclusive")
+        self.temperature = temperature
+        self.energy_idx = energy_idx
+        self.learn_width = learn_width
+        self.learn_temperature = learn_temperature
+        if learn_width or learn_temperature:
+            if not (width_min_ratio < 1.0 < width_max_ratio):
+                raise ValueError(
+                    f"width_min_ratio ({width_min_ratio}) and width_max_ratio ({width_max_ratio}) must bracket 1.0"
+                )
+            self._width_lo = width_min_ratio * temperature
+            self._width_hi = width_max_ratio * temperature
+            raw0 = _inverse_bounded_interp(temperature, self._width_lo, self._width_hi)
+            if learn_width:
+                self.raw_width = nn.Parameter(torch.full((n_experts,), raw0))
+            else:
+                self.raw_temperature = nn.Parameter(torch.tensor(raw0))
+        if centers_init is None:
+            centers = torch.linspace(-2.0, 2.0, n_experts)
+        else:
+            if len(centers_init) != n_experts:
+                raise ValueError(f"centers_init has {len(centers_init)} values, expected n_experts={n_experts}")
+            centers = torch.tensor(list(centers_init), dtype=torch.float32)
+        if learn_centers:
+            self.centers = nn.Parameter(centers)
+        else:
+            self.register_buffer("centers", centers)
+
+    def effective_width(self) -> torch.Tensor | float:
+        if self.learn_width:
+            return _bounded_interp(self.raw_width, self._width_lo, self._width_hi)
+        if self.learn_temperature:
+            return _bounded_interp(self.raw_temperature, self._width_lo, self._width_hi)
+        return self.temperature
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        e = cond_cont[:, self.energy_idx].unsqueeze(-1)  # (B, 1)
+        d2 = (e - self.centers.unsqueeze(0)) ** 2  # (B, n_experts)
+        return torch.softmax(-d2 / self.effective_width(), dim=-1)
+
+
+@register_router("pdg")
+class PdgRouter(Router):
+    """Soft turn-on gate over a learned PDG embedding (own table, separate
+    from the trunk's `ConditionEncoder`). No supervision needed — PDG code
+    is already known at pre-step time."""
+
+    def __init__(
+        self,
+        n_experts: int,
+        pdg_vocab: int,
+        emb_dim: int = 8,
+        temperature: float = 0.5,
+        learn_centers: bool = True,
+    ) -> None:
+        super().__init__(n_experts)
+        self.temperature = temperature
+        self.pdg_emb = nn.Embedding(pdg_vocab, emb_dim)
+        centers = torch.randn(n_experts, emb_dim) * 0.1
+        if learn_centers:
+            self.centers = nn.Parameter(centers)
+        else:
+            self.register_buffer("centers", centers)
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        e = self.pdg_emb(cond_cat[:, CondLayout.PDG_COL])  # (B, emb_dim)
+        d2 = ((e.unsqueeze(1) - self.centers.unsqueeze(0)) ** 2).sum(-1)  # (B, n_experts)
+        return torch.softmax(-d2 / self.temperature, dim=-1)
+
+
+@register_router("process")
+class ProcessRouter(Router):
+    """Routes on the physics process expected to end the step — a post-step
+    outcome, so a small classifier over pre-step conditioning predicts it
+    (own pdg/material embeddings, separate from the trunk's ConditionEncoder).
+    `n_experts` doubles as the number of process classes. Supervised via
+    `classify_loss` against the true `process` label at train time only;
+    `gate`/`top1` never see it."""
+
+    def __init__(
+        self,
+        n_experts: int,
+        pdg_vocab: int,
+        mat_vocab: int,
+        emb_dim: int = 8,
+        hidden_dim: int = 64,
+    ) -> None:
+        super().__init__(n_experts)
+        self.pdg_emb = nn.Embedding(pdg_vocab, emb_dim)
+        self.mat_emb = nn.Embedding(mat_vocab, emb_dim)
+        self.classifier = nn.Sequential(
+            nn.Linear(COND_DIM + 2 * emb_dim, hidden_dim),
+            nn.SiLU(),
+            nn.Linear(hidden_dim, n_experts),
+        )
+
+    def logits(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        pdg_e = self.pdg_emb(cond_cat[:, CondLayout.PDG_COL])
+        mat_e = self.mat_emb(cond_cat[:, CondLayout.MAT_COL])
+        h = torch.cat([cond_cont, pdg_e, mat_e], dim=-1)
+        return self.classifier(h)
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        return torch.softmax(self.logits(cond_cont, cond_cat), dim=-1)
+
+    def classify_loss(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+        return F.cross_entropy(self.logits(cond_cont, cond_cat), labels)
+
+
+class ComposedRouter(Router):
+    """Joint router over independent axes (e.g. energy x pdg), outer-product
+    gated. Not registered in `ROUTER_REGISTRY`; use `build_composed_router`."""
+
+    def __init__(self, routers: list[Router]) -> None:
+        if not routers:
+            raise ValueError("ComposedRouter needs at least one sub-router")
+        n_experts = 1
+        for r in routers:
+            n_experts *= r.n_experts
+        super().__init__(n_experts)
+        self.routers = nn.ModuleList(routers)
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        joint = self.routers[0].gate(cond_cont, cond_cat)  # (B, n_0)
+        for router in self.routers[1:]:
+            g = router.gate(cond_cont, cond_cat)  # (B, n_i)
+            joint = (joint.unsqueeze(-1) * g.unsqueeze(1)).flatten(1)  # (B, prod so far)
+        return joint
+
+    def classify_loss(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
+        total = torch.zeros((), device=cond_cont.device)
+        for router in self.routers:
+            total = total + router.classify_loss(cond_cont, cond_cat, labels)
+        return total
+
+
+def build_composed_router(specs: list[dict], **shared_kwargs) -> ComposedRouter:
+    """Build a `ComposedRouter` from a list of per-axis router specs — see
+    `_parse_composed_axes`."""
+    routers = [
+        build_router(
+            spec["type"],
+            spec["n_experts"],
+            **{
+                **shared_kwargs,
+                **{k: v for k, v in spec.items() if k not in ("type", "n_experts")},
+            },
+        )
+        for spec in specs
+    ]
+    return ComposedRouter(routers)
+
+
+_AXIS_KEY_RE = re.compile(r"^axis(\d+)_(.+)$")
+
+
+def _parse_composed_axes(router_cfg: dict) -> list[dict]:
+    """Regroup `axis{i}_{field}` flat keys into a list of per-axis spec dicts.
+
+    e.g. `axis0_type = "energy"`, `axis0_n_experts = 4`, `axis1_type = "pdg"`,
+    `axis1_n_experts = 3`, `axis1_emb_dim = 8`. Axis indices must be
+    contiguous from 0.
+    """
+    axes: dict[int, dict] = {}
+    for key, value in router_cfg.items():
+        m = _AXIS_KEY_RE.match(key)
+        if m is None:
+            continue
+        idx, field = int(m.group(1)), m.group(2)
+        axes.setdefault(idx, {})[field] = value
+    missing = set(range(len(axes))) - axes.keys()
+    if missing:
+        raise ValueError(f"composed router config has gaps at axis indices {missing}")
+    return [axes[i] for i in range(len(axes))]
+
+
+# Router types that read cond_cat's pdg index through their own
+# nn.Embedding(pdg_vocab, ...), regardless of the trunk's particle
+# conditioning mode — see _check_router_conditioning_compat.
+_VOCAB_SCOPED_ROUTER_TYPES = ("pdg", "process")
+
+
+def _check_router_conditioning_compat(router_types: list[str], particle_conditioning: str) -> None:
+    """Reject a router axis that reintroduces a training-vocab PDG lookup
+    under `conditioning.particle.type = "physical"`.
+
+    `PdgRouter`/`ProcessRouter` always build their own dataset-scoped
+    `nn.Embedding(pdg_vocab, ...)`, independent of `ConditionEncoder`'s
+    particle mode. Pairing either with `"physical"` would silently
+    reintroduce a training-menu-scoped lookup at the routing layer,
+    defeating the point of physical-property conditioning. Raised loudly at
+    model-build time.
+    """
+    bad = sorted(set(router_types) & set(_VOCAB_SCOPED_ROUTER_TYPES))
+    if bad and particle_conditioning == "physical":
+        raise ValueError(
+            f"router type(s) {bad} always use a training-vocab PDG embedding, "
+            "which is incompatible with conditioning.particle.type='physical' "
+            "(whose whole point is generalizing beyond that vocab) — pick a "
+            "different router type (e.g. 'energy') or use "
+            "conditioning.particle.type='embedding'."
+        )
+
+
+def _build_router_from_cfg(
+    router_cfg: dict,
+    pdg_vocab: int,
+    mat_vocab: int,
+    particle_conditioning: str = "embedding",
+) -> Router:
+    """Resolve one stage's `router` config into a `Router`, single-axis or
+    composed. `gumbel` is set as a post-construction attribute (shared by
+    every router type, not a per-type constructor kwarg)."""
+    shared_vocab = dict(pdg_vocab=pdg_vocab, mat_vocab=mat_vocab)
+    if router_cfg["type"] == "composed":
+        axes = _parse_composed_axes(router_cfg)
+        _check_router_conditioning_compat([a["type"] for a in axes], particle_conditioning)
+        router = build_composed_router(axes, **shared_vocab)
+        router.gumbel = bool(router_cfg.get("gumbel", False))
+        return router
+    _check_router_conditioning_compat([router_cfg["type"]], particle_conditioning)
+    router_kwargs = {k: v for k, v in router_cfg.items() if k not in ("enabled", "type", "n_experts")}
+    router_kwargs.setdefault("pdg_vocab", pdg_vocab)
+    router_kwargs.setdefault("mat_vocab", mat_vocab)
+    router = build_router(router_cfg["type"], router_cfg["n_experts"], **router_kwargs)
+    router.gumbel = bool(router_cfg.get("gumbel", False))
+    return router

giant/model/schedule.py

@@ -11,9 +11,7 @@ class CosineSchedule:
         steps = np.arange(T + 1, dtype=np.float64)
         f = np.cos(((steps / T + s) / (1.0 + s)) * np.pi / 2.0) ** 2
         alpha_bars = (f / f[0]).astype(np.float32)
-        betas = np.clip(1.0 - alpha_bars[1:] / alpha_bars[:-1], 0.0, 0.999).astype(
-            np.float32
-        )
+        betas = np.clip(1.0 - alpha_bars[1:] / alpha_bars[:-1], 0.0, 0.999).astype(np.float32)
 
         self.betas = torch.from_numpy(betas)
         self.alphas = torch.from_numpy((1.0 - betas))
@@ -48,7 +46,7 @@ class CosineSchedule:
         noise = torch.randn_like(x0)
         x_t = self.q_sample(x0, t, noise)
         t_norm = t.float() / self.T
-        pred = model(x_t, t_norm, cond_cont, cond_cat)
+        pred = model(x_t, cond_cont, cond_cat, t=t_norm)
         return F.mse_loss(pred, noise)
 
 
@@ -67,7 +65,7 @@ def flow_matching_loss(
     x0 = torch.randn_like(x1)
     x_t = (1.0 - t.view(-1, 1)) * x0 + t.view(-1, 1) * x1
     u_t = x1 - x0
-    v_t = model(x_t, t, cond_cont, cond_cat)
+    v_t = model(x_t, cond_cont, cond_cat, t=t)
     return F.mse_loss(v_t, u_t)
 
 
@@ -78,39 +76,125 @@ def flow_matching_loss_secondary(
     cond_cat: torch.Tensor,
     stage1_out: torch.Tensor,
     sec_mask: torch.Tensor,
+    type_dim: int | None = None,
 ) -> torch.Tensor:
     """Flow matching loss for the secondary decoder with per-slot masking.
 
-    x1: (B, SEC_DIM) — flattened secondary target (stick_logit, dir, log_mass, charge)
+    x1: (B, K_MAX * (CONT_SLOT_DIM + type_dim)) — flattened secondary target
+        (stick_logit, dir, then a `type_dim`-wide type slice)
     sec_mask: (B, K_MAX) bool — True for valid secondary slots
+    type_dim: width of the per-slot type slice folded into `x1` — defaults to
+        `PARTICLE_PHYS_DIM` (log_mass, charge), `particle_type.target =
+        "physical"`'s width and the only case this function handled before
+        v0.3.0 step 4. `0` means no type slice is in `x1` at all (`target`
+        in `("onehot", "embedding")` under `generator in ("flow", "ddpm")` —
+        `Stage2OneShot.type_head` handles the type loss separately in that
+        case).
 
     Only valid-slot dimensions contribute to the loss; padded slots are zeroed
     before averaging, so the loss is not diluted by empty slots.
 
     Each slot packs CONT_SLOT_DIM continuous dims (stick_logit, dir) followed
-    by PARTICLE_PHYS_DIM physical-identity dims (log_mass, charge) — the
-    secondary's predicted physical identity, a fixed regression target (see
-    giant.data.transforms.encode_secondaries). Even though the two blocks are
-    the same order of magnitude now (unlike the 16-wide learned embedding
-    block this replaced), they're still on different physical scales, so
-    they're each averaged over their own width first and then combined with
-    equal weight — this stays correct if PARTICLE_PHYS_DIM/CONT_SLOT_DIM change.
+    by the `type_dim`-wide type slice — under `target = "physical"` (the
+    default) that's the secondary's predicted physical identity, a fixed
+    regression target (see giant.data.transforms.encode_secondaries); under
+    `target = "embedding"` (folded in only for `generator = "wgan"`, so
+    `type_dim > 0` here only ever means "physical") it would be the detached
+    embedding-table row. Even though the two blocks are the same order of
+    magnitude now (unlike the 16-wide learned embedding block "physical"
+    replaced), they're still on different physical scales, so they're each
+    averaged over their own width first and then combined with equal weight
+    — this stays correct if type_dim/CONT_SLOT_DIM change.
     """
-    from giant.constants import CONT_SLOT_DIM, K_MAX, PARTICLE_PHYS_DIM, SEC_SLOT_DIM
+    from giant.constants import CONT_SLOT_DIM, PARTICLE_PHYS_DIM
+
+    if type_dim is None:
+        type_dim = PARTICLE_PHYS_DIM
 
     B = x1.size(0)
+    k_max = sec_mask.size(1)
     t = torch.rand(B, device=x1.device)
     x0 = torch.randn_like(x1)
     x_t = (1.0 - t.view(-1, 1)) * x0 + t.view(-1, 1) * x1
     u_t = x1 - x0
-    v_t = model(x_t, t, cond_cont, cond_cat, stage1_out)
+    v_t = model(x_t, cond_cont, cond_cat, stage1_out, t=t)
 
-    err = ((v_t - u_t) ** 2).view(B, K_MAX, SEC_SLOT_DIM)
-    cont_err = err[..., :CONT_SLOT_DIM].mean(dim=-1)  # (B, K_MAX)
-    phys_err = err[..., CONT_SLOT_DIM : CONT_SLOT_DIM + PARTICLE_PHYS_DIM].mean(dim=-1)
+    slot_dim = CONT_SLOT_DIM + type_dim
+    err = ((v_t - u_t) ** 2).view(B, k_max, slot_dim)
+    cont_err = err[..., :CONT_SLOT_DIM].mean(dim=-1)  # (B, k_max)
 
     mask = sec_mask.float()
     denom = mask.sum().clamp(min=1)
     cont_loss = (cont_err * mask).sum() / denom
+    if type_dim == 0:
+        return cont_loss
+    phys_err = err[..., CONT_SLOT_DIM : CONT_SLOT_DIM + type_dim].mean(dim=-1)
+    phys_loss = (phys_err * mask).sum() / denom
+    return cont_loss + phys_loss
+
+
+def flow_matching_loss_secondary_ar(
+    model: torch.nn.Module,
+    x1: torch.Tensor,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    stage1_out: torch.Tensor,
+    history_feat: torch.Tensor,
+    has_prev: torch.Tensor,
+    remaining_frac: torch.Tensor,
+    slot_idx: torch.Tensor,
+    sec_mask: torch.Tensor,
+    type_dim: int | None = None,
+) -> torch.Tensor:
+    """`Stage2Autoregressive` analogue of `flow_matching_loss_secondary`, same
+    masked, per-block (continuous vs. type) loss recipe — but native to
+    `Stage2Autoregressive`'s `(B, K_MAX, token_dim)` I/O and its extra
+    per-token conditioning args, rather than a flattened `(B, K_MAX*token_dim)`
+    vector. Kept as a sibling rather than unified with the flat version: the
+    model call signature differs enough (four extra per-token conditioning
+    tensors) that merging would need an awkward shape-flag + closure.
+
+    Under teacher forcing this is still a
+    single parallel pass over all K_MAX tokens — `x1`/`history_feat`/etc. are
+    already built from ground truth for every slot by the caller
+    (`giant.training.stage2_inputs._assemble_stage2_ar_inputs`/`_assemble_stage2_ar_target`).
+
+    x1: (B, K_MAX, CONT_SLOT_DIM + type_dim) — per-token flattened target
+        (stick_logit, dir, then a `type_dim`-wide type slice)
+    sec_mask: (B, K_MAX) bool — True for valid secondary slots
+    type_dim: as `flow_matching_loss_secondary` — defaults to
+        `PARTICLE_PHYS_DIM`, `0` means no type slice is in `x1` at all.
+    """
+    from giant.constants import CONT_SLOT_DIM, PARTICLE_PHYS_DIM
+
+    if type_dim is None:
+        type_dim = PARTICLE_PHYS_DIM
+
+    B, K, _ = x1.shape
+    t = torch.rand(B, K, device=x1.device)
+    x0 = torch.randn_like(x1)
+    x_t = (1.0 - t.unsqueeze(-1)) * x0 + t.unsqueeze(-1) * x1
+    u_t = x1 - x0
+    v_t = model(
+        x_t,
+        cond_cont,
+        cond_cat,
+        stage1_out,
+        history_feat,
+        has_prev,
+        remaining_frac,
+        slot_idx,
+        t=t,
+    )
+
+    err = (v_t - u_t) ** 2
+    cont_err = err[..., :CONT_SLOT_DIM].mean(dim=-1)  # (B, K_MAX)
+
+    mask = sec_mask.float()
+    denom = mask.sum().clamp(min=1)
+    cont_loss = (cont_err * mask).sum() / denom
+    if type_dim == 0:
+        return cont_loss
+    phys_err = err[..., CONT_SLOT_DIM : CONT_SLOT_DIM + type_dim].mean(dim=-1)
     phys_loss = (phys_err * mask).sum() / denom
     return cont_loss + phys_loss

giant/model/trunks.py

@@ -0,0 +1,214 @@
+"""Trunks: everything downstream of the fused conditioning vector — a
+registrable expert *body* architecture (`TRUNK_REGISTRY`/`register_trunk`),
+used standalone or mixed by a `Router` (issues.md Issue 8; trunk-selectability
+gitea #33).
+
+Whether a body is mixed is orthogonal to which body it is: `RoutedTrunk`
+builds `router.n_experts` instances of whichever body `trunk_type` names, so
+a future body (e.g. a transformer) automatically gets a mixture variant for
+free — no separate "routed transformer trunk" class needed.
+"""
+
+import torch
+import torch.nn as nn
+
+from giant.model.layers import build_block
+from giant.model.routers import Router
+
+TRUNK_REGISTRY: dict[str, type[nn.Module]] = {}
+
+
+def register_trunk(name: str):
+    def decorator(cls: type[nn.Module]) -> type[nn.Module]:
+        TRUNK_REGISTRY[name] = cls
+        return cls
+
+    return decorator
+
+
+def build_expert_body(
+    name: str,
+    in_dim: int,
+    out_dim: int,
+    hidden_dim: int,
+    n_blocks: int,
+    cond_dim: int,
+    dropout: float = 0.0,
+    block_conditioning: str = "add",
+) -> nn.Module:
+    """Factory: look up a registered trunk body by name and construct one
+    instance of it — used both for a standalone (unrouted) trunk and for each
+    expert inside a `RoutedTrunk`. `block_conditioning` selects the
+    `BLOCK_REGISTRY` entry each body's internal `ResBlock`-family blocks use
+    (`trunk.block_conditioning`, gitea #34) — an optional trailing kwarg a
+    future non-`ResBlock`-based body can simply ignore, same idiom as
+    `Trunk.forward`'s accept-and-ignore `cond_cont`/`cond_cat`."""
+    if name not in TRUNK_REGISTRY:
+        raise ValueError(f"unknown trunk type {name!r}; available: {sorted(TRUNK_REGISTRY)}")
+    cls = TRUNK_REGISTRY[name]
+    return cls(in_dim, out_dim, hidden_dim, n_blocks, cond_dim, dropout, block_conditioning=block_conditioning)
+
+
+@register_trunk("resmlp")
+class ExpertTrunk(nn.Module):
+    """`input_proj -> ResBlock stack -> out_proj` — the registered `"resmlp"`
+    trunk body. Used both standalone (no router: `forward`'s `cond_cont`/
+    `cond_cat` are accepted and ignored, satisfying the `Trunk` interface
+    directly with no wrapper class) and as one expert inside a `RoutedTrunk`
+    (`_route_forward` calls it with just `(x, cond)`).
+
+    Unlike v0.2, `out_dim` is independent of `in_dim` — needed by stage-2 AR
+    tokens later (`noise_dim` in, `4 + type_dim` out), even though every
+    step-2/3 caller still has `in_dim == out_dim`.
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        hidden_dim: int,
+        n_blocks: int,
+        cond_dim: int,
+        dropout: float = 0.0,
+        block_conditioning: str = "add",
+    ) -> None:
+        super().__init__()
+        self.out_dim = out_dim
+        self.input_proj = nn.Linear(in_dim, hidden_dim)
+        self.blocks = nn.ModuleList(
+            [build_block(block_conditioning, hidden_dim, cond_dim, dropout) for _ in range(n_blocks)]
+        )
+        self.out_proj = nn.Linear(hidden_dim, out_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond: torch.Tensor,
+        cond_cont: torch.Tensor | None = None,
+        cond_cat: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        x = self.input_proj(x)
+        for block in self.blocks:
+            x = block(x, cond)
+        return self.out_proj(x)
+
+
+def _route_forward(
+    experts: nn.ModuleList,
+    router: Router,
+    x: torch.Tensor,
+    cond: torch.Tensor,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    training: bool,
+) -> torch.Tensor:
+    """Shared dispatch for `RoutedTrunk`.
+
+    Train mode: full mixture `sum_i weight_i * expert_i(x)` — always
+    N-expert dense compute, fully differentiable (`weight` is
+    `router.combine_weights`). Eval mode: grouped top-1 dispatch — each row
+    runs exactly one expert, the actual source of the per-call speedup.
+    """
+    if training:
+        weights = router.combine_weights(cond_cont, cond_cat)  # (B, n_experts)
+        out = torch.zeros(x.shape[0], experts[0].out_dim, device=x.device)
+        for i, expert in enumerate(experts):
+            out = out + weights[:, i : i + 1] * expert(x, cond)
+        return out
+
+    idx = router.top1(cond_cont, cond_cat)  # (B,)
+    out_dim = experts[0].out_dim
+    out = torch.zeros(x.shape[0], out_dim, device=x.device)
+    for i, expert in enumerate(experts):
+        mask = idx == i
+        if mask.any():
+            out[mask] = expert(x[mask], cond[mask])
+    return out
+
+
+class Trunk(nn.Module):
+    """Interface implemented by a standalone trunk body (any `TRUNK_REGISTRY`
+    entry, e.g. `ExpertTrunk`) and by `RoutedTrunk`: everything downstream of
+    the fused conditioning vector, i.e. the actual generative trunk of a
+    stage."""
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+
+class RoutedTrunk(Trunk):
+    def __init__(
+        self,
+        router: Router,
+        trunk_type: str,
+        in_dim: int,
+        out_dim: int,
+        hidden_dim: int,
+        n_res_blocks: int,
+        cond_dim: int,
+        dropout: float = 0.0,
+        block_conditioning: str = "add",
+    ) -> None:
+        super().__init__()
+        self.router = router
+        self.experts = nn.ModuleList(
+            [
+                build_expert_body(
+                    trunk_type,
+                    in_dim,
+                    out_dim,
+                    hidden_dim,
+                    n_res_blocks,
+                    cond_dim,
+                    dropout,
+                    block_conditioning=block_conditioning,
+                )
+                for _ in range(router.n_experts)
+            ]
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cond: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+    ) -> torch.Tensor:
+        return _route_forward(self.experts, self.router, x, cond, cond_cont, cond_cat, self.training)
+
+
+def build_trunk(
+    router: Router | None,
+    trunk_type: str,
+    in_dim: int,
+    out_dim: int,
+    hidden_dim: int,
+    n_res_blocks: int,
+    cond_dim: int,
+    dropout: float = 0.0,
+    block_conditioning: str = "add",
+) -> nn.Module:
+    """Build a stage's trunk: `trunk_type` (a `TRUNK_REGISTRY` key, e.g.
+    `"resmlp"`) selects the expert body architecture; `router`, if given,
+    wraps `router.n_experts` instances of that body in a `RoutedTrunk`
+    mixture — otherwise a single body is returned directly (no wrapper
+    class), which is what makes an unrouted trunk's state-dict keys land
+    directly under `trunk.*` instead of `trunk.experts.0.*` (see
+    `giant.model._legacy.migrate_legacy_state_dict`, which assumes exactly
+    this flat layout for a v0.2 monolithic checkpoint). `block_conditioning`
+    (a `BLOCK_REGISTRY` key, e.g. `"add"`/`"film"`/`"adaln"`) selects each
+    body's conditioning-injection mechanism (gitea #34).
+    """
+    if router is not None:
+        return RoutedTrunk(
+            router, trunk_type, in_dim, out_dim, hidden_dim, n_res_blocks, cond_dim, dropout, block_conditioning
+        )
+    return build_expert_body(
+        trunk_type, in_dim, out_dim, hidden_dim, n_res_blocks, cond_dim, dropout, block_conditioning
+    )

giant/particles.py

@@ -7,16 +7,28 @@ table involved) for isomer/excited nuclear codes the package's ground-state-only
 nuclide table doesn't cover — confirmed necessary for ~32% of the nuclear codes
 actually present in the multi-material dataset
 (`0932fb02-f2ce-43ca-a4ef-60a2b1221bbc.parquet`).
+
+Also holds the v0.3.0 stage-2 categorical-type rollout decode:
+`decode_topn_class`/`decode_embedding_nearest` turn
+`Stage2Autoregressive`/`Stage2OneShot`'s `"onehot"`/`"embedding"` type
+predictions back into concrete PDG codes, the one place a secondary's
+categorical/continuous type representation is ever discretized (its
+free-running history representation stays unsnapped — see
+`giant/sample.py`'s AR loop).
 """
 
 from __future__ import annotations
 
 from functools import lru_cache
+from typing import TYPE_CHECKING
 
 import numpy as np
 from particle import InvalidParticle, Particle, ParticleNotFound
 from particle import pdgid as _pdgid
 
+if TYPE_CHECKING:
+    from giant.data.loader import TopNMap
+
 # First-pass nuclear mass approximation (A * atomic mass unit); no
 # binding-energy correction. Only used for codes missing from `particle`'s
 # ground-state nuclide table -- ground-state codes get the package's real
@@ -46,9 +58,7 @@ def particle_mass_charge(pdg: int) -> tuple[float, float]:
         if _pdgid.is_nucleus(pdg):
             z, a = _pdgid.Z(pdg), _pdgid.A(pdg)
             if z is None or a is None:
-                raise ValueError(
-                    f"PDG {pdg}: is_nucleus but Z/A decode failed"
-                ) from None
+                raise ValueError(f"PDG {pdg}: is_nucleus but Z/A decode failed") from None
             return float(a) * _AMU_MEV, float(z)
         raise ValueError(
             f"PDG code {pdg} could not be resolved via the `particle` package "
@@ -97,9 +107,7 @@ def nearest_known_pdg(mass: np.ndarray, charge: np.ndarray, candidates) -> np.nd
     if len(resolved) == 0:
         raise ValueError("nearest_known_pdg: no resolvable candidates")
     codes = np.array([r[0] for r in resolved], dtype=np.int64)
-    table_log_mass = np.log(
-        np.array([r[1] for r in resolved], dtype=np.float64) + _LOG_EPS
-    )
+    table_log_mass = np.log(np.array([r[1] for r in resolved], dtype=np.float64) + _LOG_EPS)
     table_charge = np.array([r[2] for r in resolved], dtype=np.float64)
 
     mass = np.asarray(mass, dtype=np.float64)
@@ -111,3 +119,112 @@ def nearest_known_pdg(mass: np.ndarray, charge: np.ndarray, candidates) -> np.nd
     ) ** 2
     idx = d2.argmin(axis=1)
     return codes[idx]
+
+
+def invert_dense_map(m: dict[int, int]) -> dict[int, int]:
+    """index -> key, inverting a dense, bijective value->index map (`pdg_map`,
+    or a `TopNMap.class_map`'s non-"other" entries — see `decode_topn_class`,
+    which needs a *partial* inverse, not this general one, because its "other"
+    index isn't unique-preimage). `pdg_map` itself is always a true bijection
+    (`giant.data.loader.build_index_maps_from_files` enumerates the vocab), so
+    a plain dict-comprehension inversion is exact here — used for
+    `stage2_model.particle_type.target = "embedding"` decode, whose vocabulary
+    is the full dense `pdg_map`, not a top-N-plus-other map."""
+    return {v: k for k, v in m.items()}
+
+
+def decode_topn_class(
+    class_idx: np.ndarray,
+    topn_map: "TopNMap",
+    n_classes: int,
+    other_policy: str = "sample",
+    rng: np.random.Generator | None = None,
+) -> np.ndarray:
+    """`stage2_model.particle_type.target = "onehot"` inference decode:
+    per-row top-N class index -> concrete secondary-species PDG code.
+
+    class_idx: int array, any shape, values in `[0, n_classes)`.
+    topn_map: the `TopNMap` (`giant.data.loader.build_pdg_topn_map_from_files`)
+        this class index was built from — `class_map` (PDG -> class, injective
+        except at the shared "other" index) plus `other_members` (the
+        empirical within-"other" distribution, needed for `other_policy =
+        "sample"`/`"modal"`).
+    n_classes: the resolved secondary-species class count
+        (`giant.model.models.resolve_type_n_classes` —
+        `stage2_model.particle_type.n_classes`, 0 = inherit
+        `conditioning.particle.emb_dim`; see gitea #29); the "other" bucket
+        is index `n_classes - 1` by construction
+        (`giant.data.loader._topn_plus_other_map`).
+    other_policy: `"sample"` draws from `other_members`' empirical frequency;
+        `"modal"` always the single most common "other" member; `"drop"`
+        returns PDG `0` for those rows (not a valid PDG code — the caller
+        must treat it as "no secondary", the same convention as
+        `TERM_UNKNOWN_PDG` elsewhere in the rollout driver).
+
+    Every non-"other" class index has a unique inverse (the top `n_classes -
+    1` keys each got their own index in `_topn_plus_other_map`), so those
+    rows decode exactly; only "other" rows need `other_policy`.
+    """
+    other_idx = n_classes - 1
+    inv = np.zeros(n_classes, dtype=np.int64)
+    for pdg, idx in topn_map.class_map.items():
+        if idx != other_idx:
+            inv[idx] = pdg
+
+    flat = np.asarray(class_idx, dtype=np.int64).reshape(-1)
+    out = inv[np.clip(flat, 0, n_classes - 1)]
+
+    other_mask = flat == other_idx
+    n_other = int(other_mask.sum())
+    if n_other:
+        if not topn_map.other_members:
+            raise ValueError("decode_topn_class: 'other' class predicted but topn_map.other_members is empty")
+        members = np.array(list(topn_map.other_members.keys()), dtype=np.int64)
+        counts = np.array(list(topn_map.other_members.values()), dtype=np.float64)
+        if other_policy == "drop":
+            out[other_mask] = 0
+        elif other_policy == "modal":
+            out[other_mask] = members[counts.argmax()]
+        elif other_policy == "sample":
+            rng = rng if rng is not None else np.random.default_rng()
+            probs = counts / counts.sum()
+            out[other_mask] = rng.choice(members, size=n_other, p=probs)
+        else:
+            raise ValueError(f"unknown other_policy {other_policy!r}")
+
+    return out.reshape(np.asarray(class_idx).shape)
+
+
+def decode_embedding_nearest(
+    vectors: np.ndarray,
+    emb_weight: np.ndarray,
+    idx_to_pdg: dict[int, int],
+) -> tuple[np.ndarray, np.ndarray]:
+    """`stage2_model.particle_type.target = "embedding"` inference decode:
+    L1-nearest row of the conditioning's own particle embedding table, since
+    a generative model's continuous output
+    essentially never lands within float tolerance of a table row (the exact-
+    match form is only valid as a round-trip test assertion, never here).
+
+    vectors: `(..., emb_dim)` raw predicted vectors, any leading shape.
+    emb_weight: `(vocab, emb_dim)` — `ConditionEncoder.pdg_emb.weight`,
+        detached and moved to numpy by the caller. This is the SAME table
+        `particle_type.target = "embedding"` was regressed against
+        (`validate_config` requires `conditioning.particle.type =
+        "embedding"` whenever this target is used — one table, not two).
+    idx_to_pdg: `invert_dense_map(pdg_map)` — embedding row index -> PDG.
+
+    Returns `(pdg, l1_dist)`, both shaped like `vectors.shape[:-1]`. `l1_dist`
+    is a diagnostic: a heavy tail means the decoder is emitting vectors off
+    the embedding manifold, the direct analogue of the species-collapse
+    symptom this redesign exists to fix.
+    """
+    emb_dim = vectors.shape[-1]
+    flat = np.asarray(vectors, dtype=np.float64).reshape(-1, emb_dim)
+    table = np.asarray(emb_weight, dtype=np.float64)
+    d = np.abs(flat[:, None, :] - table[None, :, :]).sum(axis=-1)  # (N, vocab)
+    nearest = d.argmin(axis=1)
+    dist = d[np.arange(len(nearest)), nearest]
+    pdg = np.array([idx_to_pdg[int(i)] for i in nearest], dtype=np.int64)
+    lead_shape = vectors.shape[:-1]
+    return pdg.reshape(lead_shape), dist.reshape(lead_shape).astype(np.float32)

giant/pipeline.py

@@ -9,19 +9,19 @@ from torch.utils.data import DataLoader
 from giant import config
 from giant.constants import (
     COND_DIM,
-    EMB_DIM,
-    K_MAX,
     PARTICLE_PHYS_DIM,
-    SEC_SLOT_DIM,
     X_DIM,
 )
 from giant.data import setup_cache
 from giant.data.loader import (
+    TopNMap,
     event_id_offset,
     find_parquet_files,
     iter_file_chunks,
     build_index_maps_from_files,
+    build_pdg_topn_map_from_files,
     build_process_map_from_files,
+    build_topn_map_from_files,
 )
 from giant.data.transforms import (
     Normalizer,
@@ -31,8 +31,8 @@ from giant.data.transforms import (
     sorted_membership,
 )
 from giant.data.dataset import make_event_split, StreamingStepsDataset
-from giant.model.network import build_models, build_critics
-from giant.train import train as run_training
+from giant.model.network import build_models, build_critics, resolve_type_n_classes
+from giant.training import train as run_training
 
 
 @dataclass
@@ -48,6 +48,9 @@ class SetupStageResult:
     pdg_map: dict[int, int]
     mat_map: dict[str, int]
     proc_map: dict[str, int] | None
+    pdg_topn_map: TopNMap | None
+    sec_type_topn_map: TopNMap | None
+    mat_topn_map: TopNMap | None
     cond_norm: Normalizer
     tgt_norm: Normalizer
     sec_phys_norm: Normalizer
@@ -56,26 +59,62 @@ class SetupStageResult:
     n_train_steps: int
 
 
+def _seed_energy_router(
+    router_cfg: dict,
+    cond_norm: Normalizer,
+    energy_quantiles: np.ndarray,
+    energy_idx: int,
+    echo,
+) -> None:
+    """Mutate `router_cfg["centers_init"]` in place from real data quantiles,
+    when this stage's router is an enabled EnergyRouter. Shared by both
+    stages' router configs — each seeded independently, since v0.3.0 stages
+    may have entirely different router configs."""
+    active = router_cfg.get("enabled") and router_cfg.get("type") == "energy"
+    if not active:
+        return
+    if energy_quantiles.size == 0:
+        echo("  warning: no energy samples collected — EnergyRouter falls back to default centers")
+        return
+    assert cond_norm.mean is not None and cond_norm.std is not None
+    levels = np.linspace(0.0, 1.0, router_cfg["n_experts"])
+    raw_centers = setup_cache.energy_quantile_at(energy_quantiles, levels)
+    centers_init = (raw_centers - cond_norm.mean[energy_idx]) / cond_norm.std[energy_idx]
+    router_cfg["centers_init"] = centers_init.astype(np.float32).tolist()
+    echo(f"  seeded EnergyRouter centers from data quantiles: {router_cfg['centers_init']}")
+
+
 def run_setup_stage(
     data: str | Path,
     val_fraction: float,
     seed: int,
-    conditioning: str,
-    router_cfg: dict,
+    cfg: dict,
     cache_setup: bool = True,
     rebuild_setup_cache: bool = False,
     echo=print,
 ) -> SetupStageResult:
     """Scan `data` for everything training needs before the epoch loop: the
     train/val event split, pdg/material vocab maps, an optional process map
-    (`router_cfg.type == "process"`), and the Stage-1/Stage-2 normalizers.
+    (needed if either stage's router is type="process"), and the Stage-1/
+    Stage-2 normalizers.
+
+    `cfg` is the full merged v0.3 config (`conditioning`/`stage1_model`/
+    `stage2_model`), already passed through `giant.config.validate_config`.
+    `conditioning.particle.type` and `conditioning.material.type` are
+    independent and may differ.
 
     Reads from and writes to the `giant.data.setup_cache` sidecar when
     `cache_setup` is set (`rebuild_setup_cache` ignores — but still
-    refreshes — any existing sidecar content). `router_cfg` may be mutated
-    in place: an active `EnergyRouter` (`router_cfg["type"] == "energy"`)
+    refreshes — any existing sidecar content). Each stage's `router` config
+    is mutated in place: an active `EnergyRouter` (`router.type == "energy"`)
     gets its `centers_init` seeded from real data quantiles here.
     """
+    particle_conditioning = cfg["conditioning"]["particle"]["type"]
+    material_conditioning = cfg["conditioning"]["material"]["type"]
+    k_max = cfg["stage2_model"]["k_max"]
+    stage1_router = cfg["stage1_model"].get("router") or {}
+    stage2_router = cfg["stage2_model"].get("router") or {}
+
     files = find_parquet_files(data)
     echo(f"found {len(files)} parquet file(s)")
 
@@ -97,23 +136,14 @@ def run_setup_stage(
         if cache is not None:
             cache.event_index = (unique_ids, counts)
 
-    train_events, val_events = make_event_split(
-        unique_ids, val_fraction=val_fraction, seed=seed
-    )
+    train_events, val_events = make_event_split(unique_ids, val_fraction=val_fraction, seed=seed)
     events_arr = np.array(sorted(train_events))
     n_train_steps = setup_cache.n_train_steps_for_split(unique_ids, counts, events_arr)
-    echo(
-        f"  {int(counts.sum()):,} steps  |  "
-        f"{len(train_events)} train events  |  "
-        f"{len(val_events)} val events"
-    )
+    echo(f"  {int(counts.sum()):,} steps  |  {len(train_events)} train events  |  {len(val_events)} val events")
 
     if cache is not None and cache.vocab is not None:
         pdg_map, mat_map = cache.vocab
-        echo(
-            f"vocabulary maps: cache hit ({len(pdg_map)} PDG codes, "
-            f"{len(mat_map)} materials)"
-        )
+        echo(f"vocabulary maps: cache hit ({len(pdg_map)} PDG codes, {len(mat_map)} materials)")
     else:
         echo("building vocabulary maps …")
         pdg_map, mat_map = build_index_maps_from_files(files)
@@ -121,16 +151,23 @@ def run_setup_stage(
         if cache is not None:
             cache.vocab = (pdg_map, mat_map)
 
+    # A process map is needed if either stage's router reads the physics
+    # process label (type="process"). Only one map is built even if both
+    # stages want one — see the module-level note in giant/cli.py's
+    # _router_total_experts for why composed-router n_experts isn't a plain
+    # int; process routers are never composed in practice, so this doesn't
+    # need that generality.
     proc_map: dict[str, int] | None = None
-    if router_cfg.get("enabled") and router_cfg.get("type") == "process":
-        n_experts = router_cfg["n_experts"]
+    process_router_cfg = next(
+        (r for r in (stage1_router, stage2_router) if r.get("enabled") and r.get("type") == "process"),
+        None,
+    )
+    if process_router_cfg is not None:
+        n_experts = process_router_cfg["n_experts"]
         cached_proc_map = cache.proc_maps.get(n_experts) if cache is not None else None
         if cached_proc_map is not None:
             proc_map = cached_proc_map
-            echo(
-                f"process vocabulary: cache hit ({len(proc_map)} labels, "
-                f"{n_experts} experts)"
-            )
+            echo(f"process vocabulary: cache hit ({len(proc_map)} labels, {n_experts} experts)")
         else:
             echo("building process vocabulary …")
             proc_map = build_process_map_from_files(files, n_experts=n_experts)
@@ -138,11 +175,62 @@ def run_setup_stage(
             if cache is not None:
                 cache.proc_maps[n_experts] = proc_map
 
-    energy_router_active = (
-        router_cfg.get("enabled") and router_cfg.get("type") == "energy"
-    )
-    energy_idx = router_cfg.get("energy_idx", 3)
-    norm_key = setup_cache.normalizer_key(val_fraction, seed, conditioning)
+    # Top-N-plus-other maps for onehot conditioning/type axes.
+    # The PDG axis is used independently by conditioning.particle.type="onehot"
+    # (cond_cat's onehot feature) and stage2_model.particle_type.target="onehot"
+    # (secondary-species decode) — their class counts can now differ (gitea
+    # #29: stage2_model.particle_type.n_classes, 0 = inherit
+    # conditioning.particle.emb_dim), so each is resolved and built
+    # independently via _pdg_topn below. cache.topn_maps is keyed by
+    # (axis, n_classes) (setup_cache.topn_key), so when the two resolve to
+    # the same N the second call is a cache hit against the first — no extra
+    # scan in the common case where they still match. The material axis is
+    # independent of both.
+    particle_cfg = cfg["conditioning"]["particle"]
+    material_cfg = cfg["conditioning"]["material"]
+    particle_type_cfg = config.ParticleTypeConfig.from_dict(cfg["stage2_model"].get("particle_type"))
+    particle_type_target = particle_type_cfg.target
+
+    def _pdg_topn(n_classes: int) -> TopNMap:
+        cache_key = setup_cache.topn_key("pdg", n_classes)
+        cached = cache.topn_maps.get(cache_key) if cache is not None else None
+        if cached is not None:
+            echo(f"pdg top-N map: cache hit ({len(cached.class_map)} codes, {n_classes} classes)")
+            return cached
+        echo("building pdg top-N map …")
+        topn_map = build_pdg_topn_map_from_files(files, n_classes=n_classes)
+        echo(f"  {len(topn_map.class_map)} pdg codes mapped to {n_classes} classes")
+        if cache is not None:
+            cache.topn_maps[cache_key] = topn_map
+        return topn_map
+
+    pdg_topn_map: TopNMap | None = None
+    if particle_cfg["type"] == "onehot":
+        pdg_topn_map = _pdg_topn(particle_cfg["emb_dim"])
+
+    sec_type_topn_map: TopNMap | None = None
+    if particle_type_target == "onehot":
+        sec_type_n_classes = resolve_type_n_classes(particle_type_cfg, particle_cfg["emb_dim"])
+        sec_type_topn_map = _pdg_topn(sec_type_n_classes)
+
+    mat_topn_map: TopNMap | None = None
+    if material_cfg["type"] == "onehot":
+        n_classes = material_cfg["emb_dim"]
+        cache_key = setup_cache.topn_key("material", n_classes)
+        cached = cache.topn_maps.get(cache_key) if cache is not None else None
+        if cached is not None:
+            mat_topn_map = cached
+            echo(f"material top-N map: cache hit ({len(mat_topn_map.class_map)} materials, {n_classes} classes)")
+        else:
+            echo("building material top-N map …")
+            mat_topn_map = build_topn_map_from_files(files, "material", n_classes=n_classes, cast=str)
+            echo(f"  {len(mat_topn_map.class_map)} materials mapped to {n_classes} classes")
+            if cache is not None:
+                cache.topn_maps[cache_key] = mat_topn_map
+
+    energy_router_active = any(r.get("enabled") and r.get("type") == "energy" for r in (stage1_router, stage2_router))
+    energy_idx = 3
+    norm_key = setup_cache.normalizer_key(val_fraction, seed, particle_conditioning, material_conditioning)
     entry = cache.normalizers.get(norm_key) if cache is not None else None
 
     if entry is not None:
@@ -163,33 +251,44 @@ def run_setup_stage(
         # grid (setup_cache.energy_quantiles_from_sample) so centers can
         # instead be seeded from actual data quantiles below. Collected
         # whenever the setup cache is being populated, not only when *this*
-        # run's router is energy-typed, so a later run enabling
-        # --router-type energy against this same (val_fraction, seed,
-        # conditioning) key never needs to rescan just to seed centers.
+        # run's router is energy-typed, so a later run enabling an energy
+        # router against this same (val_fraction, seed, conditioning) key
+        # never needs to rescan just to seed centers.
         collect_energy_sample = energy_router_active or cache is not None
-        energy_sampler = (
-            _ReservoirSampler(capacity=100_000) if collect_energy_sample else None
-        )
+        energy_sampler = _ReservoirSampler(capacity=100_000) if collect_energy_sample else None
         for i, path in enumerate(files):
-            for chunk in iter_file_chunks(path, offset=event_id_offset(i)):
+            for chunk in iter_file_chunks(path, offset=event_id_offset(i), k_max=k_max):
                 mask = sorted_membership(chunk["event_id"], events_arr)
                 if not mask.any():
                     continue
                 chunk_tr = {k: v[mask] for k, v in chunk.items()}
-                cond_cont, _, target_s1, n_sec, sec_cont, _proc, _, _ = build_features(
+                feats = build_features(
                     chunk_tr,
                     pdg_map,
                     mat_map,
                     proc_map=proc_map,
                     require_secondaries=True,
-                    conditioning=conditioning,
+                    particle_conditioning=particle_conditioning,
+                    material_conditioning=material_conditioning,
                     sec_phys_only=True,
+                    # This pass reads only cond_cont/sec_cont, never cond_cat —
+                    # but cond_cat's width is the conditioning modes' call
+                    # (giant.cond_layout.CondLayout), so an "onehot" axis still
+                    # has to be handed its map rather than silently yielding a
+                    # narrower array.
+                    pdg_topn_map=pdg_topn_map.class_map if pdg_topn_map is not None else None,
+                    mat_topn_map=mat_topn_map.class_map if mat_topn_map is not None else None,
+                    k_max=k_max,
                 )
+                cond_cont = feats.cond_cont
+                target_s1 = feats.target_s1
+                n_sec = feats.n_sec
+                sec_cont = feats.sec_cont
                 cond_acc.update(cond_cont)
                 tgt_acc.update(target_s1)
                 if energy_sampler is not None:
                     energy_sampler.update(cond_cont[:, energy_idx])
-                sec_valid = np.arange(K_MAX)[None, :] < n_sec[:, None]
+                sec_valid = np.arange(sec_cont.shape[1])[None, :] < n_sec[:, None]
                 sec_phys = sec_cont[:, :, 4:6][sec_valid]
                 if len(sec_phys) > 0:
                     sec_phys_acc.update(sec_phys)
@@ -206,22 +305,8 @@ def run_setup_stage(
                 cond_norm, tgt_norm, sec_phys_norm, n_train_steps, energy_quantiles
             )
 
-    if energy_router_active and energy_quantiles.size > 0:
-        assert cond_norm.mean is not None and cond_norm.std is not None
-        levels = np.linspace(0.0, 1.0, router_cfg["n_experts"])
-        raw_centers = setup_cache.energy_quantile_at(energy_quantiles, levels)
-        centers_init = (raw_centers - cond_norm.mean[energy_idx]) / cond_norm.std[
-            energy_idx
-        ]
-        router_cfg["centers_init"] = centers_init.astype(np.float32).tolist()
-        echo(
-            f"  seeded EnergyRouter centers from data quantiles: {router_cfg['centers_init']}"
-        )
-    elif energy_router_active:
-        echo(
-            "  warning: no energy samples collected — EnergyRouter falls back to "
-            "default centers"
-        )
+    _seed_energy_router(stage1_router, cond_norm, energy_quantiles, energy_idx, echo)
+    _seed_energy_router(stage2_router, cond_norm, energy_quantiles, energy_idx, echo)
 
     if cache is not None:
         setup_cache.save(data, files, cache, echo=echo)
@@ -231,6 +316,9 @@ def run_setup_stage(
         pdg_map=pdg_map,
         mat_map=mat_map,
         proc_map=proc_map,
+        pdg_topn_map=pdg_topn_map,
+        sec_type_topn_map=sec_type_topn_map,
+        mat_topn_map=mat_topn_map,
         cond_norm=cond_norm,
         tgt_norm=tgt_norm,
         sec_phys_norm=sec_phys_norm,
@@ -252,7 +340,7 @@ def run_train_job(
     rebuild_setup_cache: bool = False,
     echo=print,
 ) -> None:
-    t, m = cfg["train"], cfg["model"]
+    t = cfg["train"]
     config.seed_everything(t["seed"])
 
     out_dir = Path(out_dir)
@@ -271,20 +359,15 @@ def run_train_job(
             "section)"
         )
 
-    router_cfg = m["router"]
-    if t["mode"] == "wgan" and router_cfg.get("enabled"):
-        raise ValueError(
-            "--mode wgan does not support --router (no routed WGAN generator/"
-            "critic exists) — disable one or the other"
-        )
-
-    conditioning = m["conditioning"]
+    config.validate_config(cfg)
+    particle_conditioning = cfg["conditioning"]["particle"]["type"]
+    material_conditioning = cfg["conditioning"]["material"]["type"]
+    k_max = cfg["stage2_model"]["k_max"]
     setup = run_setup_stage(
         data,
         val_fraction=t["val_fraction"],
         seed=t["seed"],
-        conditioning=conditioning,
-        router_cfg=router_cfg,
+        cfg=cfg,
         cache_setup=cache_setup,
         rebuild_setup_cache=rebuild_setup_cache,
         echo=echo,
@@ -302,6 +385,35 @@ def run_train_job(
         setup.n_train_steps,
     )
 
+    # cond_cat's onehot columns are present per-axis, independently, under
+    # that axis's own conditioning.{particle,material}.type == "onehot"
+    # (the two axes may mix freely). run_setup_stage builds each map
+    # whenever its own axis is "onehot" (see its own
+    # particle_cfg["type"]/material_cfg["type"]
+    # checks), so they're guaranteed non-None here — asserted, not just
+    # assumed, so a future wiring bug fails loudly instead of silently
+    # dropping the onehot columns.
+    cond_pdg_topn = None
+    cond_mat_topn = None
+    if particle_conditioning == "onehot":
+        assert setup.pdg_topn_map is not None
+        cond_pdg_topn = setup.pdg_topn_map.class_map
+    if material_conditioning == "onehot":
+        assert setup.mat_topn_map is not None
+        cond_mat_topn = setup.mat_topn_map.class_map
+
+    # The secondary type-index map depends on stage2_model.particle_type.target,
+    # independently of conditioning's own onehot/embedding choice above
+    # (physical stays untouched/None).
+    particle_type_target = config.ParticleTypeConfig.from_dict(cfg["stage2_model"].get("particle_type")).target
+    if particle_type_target == "onehot":
+        assert setup.sec_type_topn_map is not None
+        sec_type_class_map = setup.sec_type_topn_map.class_map
+    elif particle_type_target == "embedding":
+        sec_type_class_map = pdg_map
+    else:
+        sec_type_class_map = None
+
     total_train_batches = n_train_steps // t["batch_size"]
     echo(f"  ~{n_train_steps:,} train steps, ~{total_train_batches:,} batches")
 
@@ -316,8 +428,13 @@ def run_train_job(
         shuffle_buffer=shuffle_buffer,
         shuffle=True,
         proc_map=proc_map,
-        conditioning=conditioning,
+        particle_conditioning=particle_conditioning,
+        material_conditioning=material_conditioning,
         sec_phys_normalizer=sec_phys_norm,
+        pdg_topn_map=cond_pdg_topn,
+        mat_topn_map=cond_mat_topn,
+        sec_type_class_map=sec_type_class_map,
+        k_max=k_max,
     )
     val_ds = StreamingStepsDataset(
         files=files,
@@ -329,8 +446,13 @@ def run_train_job(
         batch_size=t["batch_size"],
         shuffle=False,
         proc_map=proc_map,
-        conditioning=conditioning,
+        particle_conditioning=particle_conditioning,
+        material_conditioning=material_conditioning,
         sec_phys_normalizer=sec_phys_norm,
+        pdg_topn_map=cond_pdg_topn,
+        mat_topn_map=cond_mat_topn,
+        sec_type_class_map=sec_type_class_map,
+        k_max=k_max,
     )
 
     pin = device.type == "cuda"
@@ -347,60 +469,19 @@ def run_train_job(
         pin_memory=pin,
     )
 
-    emb_dim = m.get("emb_dim", EMB_DIM)
-    expert_hidden_dim, expert_n_blocks = config.resolve_expert_dims(
-        router_cfg, m["hidden_dim"], m["n_blocks"]
-    )
-    if router_cfg.get("enabled") and (expert_hidden_dim, expert_n_blocks) != (
-        m["hidden_dim"],
-        m["n_blocks"],
-    ):
-        # Only reachable via an explicit router.expert_hidden_dim/n_blocks
-        # override (the 0/"unset" sentinel always resolves to m["hidden_dim"]/
-        # ["n_blocks"] — see resolve_expert_dims), so this is never a false
-        # positive from inheritance, only a deliberate narrow/wide-experts
-        # config the checkpoint dir name (_h{hidden_dim}_b{n_blocks}) won't
-        # reflect.
-        echo(
-            f"  warning: experts are {expert_hidden_dim}x{expert_n_blocks}, "
-            f"different from model.hidden_dim/n_blocks ({m['hidden_dim']}x"
-            f"{m['n_blocks']}) — the checkpoint dir name reflects the latter, "
-            "not the experts actually being trained"
-        )
-
     model_config = {
         "pdg_vocab": len(pdg_map),
         "mat_vocab": len(mat_map),
-        "hidden_dim": m["hidden_dim"],
-        "n_blocks": m["n_blocks"],
-        "emb_dim": emb_dim,
-        "dropout": m["dropout"],
-        "k_max": K_MAX,
-        "sec_slot_dim": SEC_SLOT_DIM,
-        "conditioning": conditioning,
-        "router": dict(router_cfg),
-        "expert_hidden_dim": expert_hidden_dim,
-        "expert_n_blocks": expert_n_blocks,
-        # Read by `predict`/`rollout` (which never receive their own --mode
-        # flag) to auto-detect which sampler a checkpoint needs.
-        "mode": t["mode"],
-        "noise_dim": m.get("noise_dim", 64),
+        "conditioning": cfg["conditioning"],
+        "stage1_model": cfg["stage1_model"],
+        "stage2_model": cfg["stage2_model"],
     }
 
-    stage1_model, sec_decoder = build_models(model_config)
-    echo(
-        f"stage1: {sum(p.numel() for p in stage1_model.parameters()):,} parameters  |  "
-        f"sec_decoder: {sum(p.numel() for p in sec_decoder.parameters()):,} parameters"
-    )
-
-    critic = None
-    sec_critic = None
-    if t["mode"] == "wgan":
-        critic, sec_critic = build_critics(model_config)
-        echo(
-            f"critic: {sum(p.numel() for p in critic.parameters()):,} parameters  |  "
-            f"sec_critic: {sum(p.numel() for p in sec_critic.parameters()):,} parameters"
-        )
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    for name, model in models.items():
+        if model is not None:
+            echo(f"{name}: {sum(p.numel() for p in model.parameters()):,} parameters")
 
     out_dir.mkdir(parents=True, exist_ok=True)
     meta = config.build_run_meta(
@@ -415,25 +496,13 @@ def run_train_job(
     config.save_config(cfg, out_dir, meta)
 
     run_training(
-        stage1_model=stage1_model,
-        sec_decoder=sec_decoder,
+        cfg=cfg,
+        models=models,
+        critics=critics,
         train_loader=train_loader,
         val_loader=val_loader,
-        mode=t["mode"],
-        epochs=t["epochs"],
-        lr=t["lr"],
-        weight_decay=t["weight_decay"],
-        ema_decay=t["ema_decay"],
-        warmup_epochs=t["warmup_epochs"],
         device=device,
         out_dir=out_dir,
-        lambda_nsec=t.get("lambda_nsec", 0.1),
-        lambda_s2=t.get("lambda_s2", 1.0),
-        lambda_balance=router_cfg.get("lambda_balance", 0.0),
-        lambda_proc=router_cfg.get("lambda_proc", 0.0),
-        lambda_entropy=router_cfg.get("lambda_entropy", 0.0),
-        gumbel_tau_start=router_cfg.get("gumbel_tau_start", 1.0),
-        gumbel_tau_end=router_cfg.get("gumbel_tau_end", 0.1),
         normalizer_dict={
             "cond": cond_norm.to_dict(),
             "target": tgt_norm.to_dict(),
@@ -442,18 +511,13 @@ def run_train_job(
         pdg_map={str(k): v for k, v in pdg_map.items()},
         mat_map={str(k): v for k, v in mat_map.items()},
         proc_map=proc_map,
+        pdg_topn_map=setup.pdg_topn_map,
+        sec_type_topn_map=setup.sec_type_topn_map,
+        mat_topn_map=setup.mat_topn_map,
         model_config=model_config,
         resume_path=resume,
-        validate_every=t["validate_every"],
-        validate_steps=t["validate_steps"],
-        max_val_batches=t["max_val_batches"],
         total_train_batches=total_train_batches,
-        critic=critic,
-        sec_critic=sec_critic,
-        n_critic=t.get("n_critic", 5),
-        gp_weight=t.get("gp_weight", 10.0),
-        critic_lr=t.get("critic_lr") or None,
-        use_wandb=t.get("wandb", False),
+        use_wandb=t.get("wandb", True),
         wandb_project=t.get("wandb_project", "giant"),
         wandb_run_name=t.get("wandb_run_name", ""),
         wandb_log_every=t.get("wandb_log_every", 50),

giant/rollout.py

@@ -17,7 +17,7 @@ treated as detector leakage and not deposited.
 from __future__ import annotations
 
 from collections import Counter
-from typing import Callable, TypedDict
+from typing import TYPE_CHECKING, Callable, TypedDict
 
 import numpy as np
 import torch
@@ -33,18 +33,156 @@ from giant.data.transforms import (
     Normalizer,
     build_cond_features,
     decode_secondaries,
+    decode_secondary_cont,
     energy_simplex_decode,
     inv_local_frame_rotation,
     inv_log_transform,
     reconstruct_post_pos,
 )
-from giant.particles import nearest_known_pdg, particle_phys_array
-from giant.sample import (
-    sample_flow,
-    sample_secondaries,
-    sample_wgan,
-    sample_secondaries_wgan,
+from giant.particles import (
+    decode_embedding_nearest,
+    decode_topn_class,
+    invert_dense_map,
+    nearest_known_pdg,
+    particle_phys_array,
 )
+from giant.sample import resolve_n_sec, sample_stage1, sample_stage2
+
+if TYPE_CHECKING:
+    from giant.data.loader import TopNMap
+
+
+class L1DistCollector:
+    """Accumulates the L1-distance diagnostic across a whole rollout
+    run: the L1 distance between each emitted secondary's raw predicted
+    embedding vector and the nearest table row it snapped to (only
+    meaningful under `particle_type.target = "embedding"` —
+    `giant.particles.decode_embedding_nearest`). A heavy tail means the
+    decoder is emitting vectors off the embedding manifold — the direct
+    analogue of the species-collapse symptom the v0.3.0 redesign exists to
+    fix.
+
+    Not folded into `rollout()`'s own return value (which is shape-typed as
+    step records, see `_RECORD_KEYS`/`RolloutSummary`) — passed in and read
+    back by the caller instead, mirroring the existing `on_chunk` pattern.
+    O(1) memory via a fixed log-spaced histogram rather than raw samples,
+    since a heavy right tail is exactly what this diagnostic watches for.
+    """
+
+    def __init__(self, n_bins: int = 50, lo: float = 1e-3, hi: float = 1e3) -> None:
+        self.n = 0
+        self.total = 0.0
+        self.total_sq = 0.0
+        self.minimum = float("inf")
+        self.maximum = 0.0
+        self.hist_edges = np.geomspace(lo, hi, n_bins + 1)
+        self.hist_counts = np.zeros(n_bins, dtype=np.int64)
+
+    def add(self, dist: np.ndarray, valid: np.ndarray) -> None:
+        vals = np.asarray(dist)[np.asarray(valid)]
+        if vals.size == 0:
+            return
+        self.n += int(vals.size)
+        self.total += float(vals.sum())
+        self.total_sq += float(np.square(vals).sum())
+        self.minimum = min(self.minimum, float(vals.min()))
+        self.maximum = max(self.maximum, float(vals.max()))
+        self.hist_counts += np.histogram(vals, bins=self.hist_edges)[0]
+
+    def summary(self) -> dict | None:
+        """`None` if nothing was ever added (target != "embedding", or a
+        run with zero secondaries) — the caller should omit the diagnostic
+        entirely rather than write a degenerate summary."""
+        if self.n == 0:
+            return None
+        mean = self.total / self.n
+        variance = max(self.total_sq / self.n - mean**2, 0.0)
+        return {
+            "n": self.n,
+            "mean": mean,
+            "std": variance**0.5,
+            "min": self.minimum,
+            "max": self.maximum,
+            "hist_edges": self.hist_edges.tolist(),
+            "hist_counts": self.hist_counts.tolist(),
+        }
+
+
+def decode_secondary_identity(
+    sec_decoder: torch.nn.Module,
+    sec_cont: torch.Tensor,
+    sec_type: torch.Tensor,
+    n_sec_np: np.ndarray,
+    e_sec: np.ndarray,
+    pre_dir: np.ndarray,
+    sec_phys_norm: Normalizer,
+    pdg_map: dict[int, int],
+    sec_type_topn_map: "TopNMap | None",
+    other_policy: str,
+    rng: np.random.Generator | None,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray | None]:
+    """Decode Stage 2's raw (sec_cont, sec_type) output into physical
+    secondary attributes, branching on `sec_decoder.particle_type_cfg`:
+
+    - `"physical"`: unchanged v0.2 path — `sec_type` already *is* (log_mass,
+      charge), used as the secondary's identity as-is (no snapping).
+    - `"onehot"`: `sec_type` is per-slot class logits — argmax, then
+      `giant.particles.decode_topn_class` (+ `other_policy`) resolves a
+      concrete PDG, whose real physics (log_mass, charge) then come from
+      `giant.particles.particle_phys_array` — unlike "physical", the PDG
+      resolution IS the secondary's identity here, not just a reporting
+      label.
+    - `"embedding"`: `sec_type` is a raw vector in the conditioning's own
+      embedding space — `giant.particles.decode_embedding_nearest` L1-snaps
+      it to the nearest table row for the PDG (+ physics via
+      `particle_phys_array`), and also returns the L1 distance (see this
+      module's `L1DistCollector`).
+
+    Returns (sec_E, sec_dir_world, sec_mass, sec_charge, sec_pdg,
+    sec_type_l1_dist) — the last is `None` except under `"embedding"`.
+    """
+    target = sec_decoder.particle_type_cfg.target
+
+    if target == "physical":
+        sec_full = torch.cat([sec_cont, sec_type], dim=-1).cpu().numpy()
+        sec_E, sec_dir_world, sec_mass, sec_charge, sec_valid = decode_secondaries(
+            sec_full, n_sec_np, e_sec, pre_dir, sec_phys_normalizer=sec_phys_norm
+        )
+        sec_pdg = nearest_known_pdg(sec_mass.reshape(-1), sec_charge.reshape(-1), pdg_map.keys()).reshape(
+            sec_mass.shape
+        )
+        return sec_E, sec_dir_world, sec_mass, sec_charge, sec_pdg, None
+
+    sec_E, sec_dir_world, sec_valid = decode_secondary_cont(sec_cont.cpu().numpy(), n_sec_np, e_sec, pre_dir)
+    sec_type_np = sec_type.cpu().numpy()
+    l1_dist = None
+
+    if target == "onehot":
+        if sec_type_topn_map is None:
+            raise RuntimeError(
+                "particle_type.target='onehot' rollout needs sec_type_topn_map "
+                "(the checkpoint's saved top-N map) — see ckpt['sec_type_topn_map']"
+            )
+        class_idx = sec_type_np.argmax(axis=-1)
+        sec_pdg = decode_topn_class(
+            class_idx,
+            sec_type_topn_map,
+            n_classes=sec_decoder.type_dim,
+            other_policy=other_policy,
+            rng=rng,
+        )
+    else:  # "embedding"
+        idx_to_pdg = invert_dense_map(pdg_map)
+        emb_weight = sec_decoder.cond_enc.pdg_emb.weight.detach().cpu().numpy()
+        sec_pdg, l1_dist = decode_embedding_nearest(sec_type_np, emb_weight, idx_to_pdg)
+        l1_dist = np.where(sec_valid, l1_dist, 0.0).astype(np.float32)
+
+    sec_mass, sec_charge = particle_phys_array(sec_pdg.reshape(-1)).T
+    sec_mass = np.where(sec_valid, sec_mass.reshape(sec_pdg.shape), 0.0).astype(np.float32)
+    sec_charge = np.where(sec_valid, sec_charge.reshape(sec_pdg.shape), 0.0).astype(np.float32)
+    sec_pdg = np.where(sec_valid, sec_pdg, 0).astype(np.int64)
+    return sec_E, sec_dir_world, sec_mass, sec_charge, sec_pdg, l1_dist
+
 
 # Record columns produced per step / per terminal marker.
 _RECORD_KEYS = [
@@ -153,13 +291,9 @@ class _Recorder:
     RAM until the very end.
     """
 
-    def __init__(
-        self, sink: Callable[[dict[str, np.ndarray]], None] | None = None
-    ) -> None:
+    def __init__(self, sink: Callable[[dict[str, np.ndarray]], None] | None = None) -> None:
         self._sink = sink
-        self._cols: dict[str, list] | None = (
-            None if sink is not None else {k: [] for k in _RECORD_KEYS}
-        )
+        self._cols: dict[str, list] | None = None if sink is not None else {k: [] for k in _RECORD_KEYS}
         self.n_rows = 0
         self.termination_reason_counts: Counter[str] = Counter()
 
@@ -167,10 +301,7 @@ class _Recorder:
         n = len(cols["event_id"])
         if n == 0:
             return
-        row = {
-            k: np.asarray(cols[k], dtype=_RECORD_DTYPES[k]).reshape(n)
-            for k in _RECORD_KEYS
-        }
+        row = {k: np.asarray(cols[k], dtype=_RECORD_DTYPES[k]).reshape(n) for k in _RECORD_KEYS}
         self.n_rows += n
         reasons = row["termination_reason"]
         nonempty = reasons[reasons != ""]
@@ -187,8 +318,7 @@ class _Recorder:
 
     def to_dict(self) -> dict[str, np.ndarray]:
         assert self._cols is not None, (
-            "to_dict() is unavailable when streaming to a sink — use "
-            "n_rows/termination_reason_counts instead"
+            "to_dict() is unavailable when streaming to a sink — use n_rows/termination_reason_counts instead"
         )
         out = {}
         for k, chunks in self._cols.items():
@@ -205,7 +335,7 @@ def make_seed_frontier(
     pre_pos: np.ndarray,
     pre_E: np.ndarray,
     pre_dir: np.ndarray,
-    conditioning: str = "embedding",
+    particle_conditioning: str = "embedding",
 ) -> tuple[dict[str, np.ndarray], dict[int, int]]:
     """Build the initial frontier from primary entry states.
 
@@ -225,17 +355,17 @@ def make_seed_frontier(
     dir_ = dir_ / np.clip(np.linalg.norm(dir_, axis=1, keepdims=True), 1e-12, None)
 
     pdg_arr = np.asarray(pdg, dtype=np.int64)
-    if conditioning == "physical":
+    if particle_conditioning == "physical":
         # Real primaries always have a genuine ground-truth PDG code, looked
         # up once here and carried forward unchanged for the track's lifetime
         # (its species never changes mid-track) — same lifecycle as "pdg"
         # itself.
         mass, charge = particle_phys_array(pdg_arr).T
     else:
-        # "embedding" mode never reads mass/charge (see
+        # "embedding"/"onehot" never read mass/charge (see
         # _physical_cond_columns), so resolving them here would only risk
-        # crashing an embedding-mode rollout on a PDG code giant.particles
-        # can't resolve, for a value that's never used.
+        # crashing a rollout on a PDG code giant.particles can't resolve, for
+        # a value that's never used.
         mass = np.zeros(n, dtype=np.float64)
         charge = np.zeros(n, dtype=np.float64)
 
@@ -310,8 +440,16 @@ def rollout(
     max_tracks_per_event: int | None = None,
     escape_threshold: float | None = None,
     on_chunk: Callable[[dict[str, np.ndarray]], None] | None = None,
-    conditioning: str = "embedding",
-    mode: str = "flow",
+    particle_conditioning: str = "embedding",
+    material_conditioning: str = "embedding",
+    pdg_topn_map: "TopNMap | None" = None,
+    mat_topn_map: "TopNMap | None" = None,
+    sec_type_topn_map: "TopNMap | None" = None,
+    other_policy: str = "sample",
+    seed: int | None = None,
+    stage1_ddpm_steps: int = 1000,
+    stage2_ddpm_steps: int = 1000,
+    l1_dist_collector: "L1DistCollector | None" = None,
 ) -> dict[str, np.ndarray] | RolloutSummary:
     """Run showers to completion.
 
@@ -324,12 +462,51 @@ def rollout(
     dict[str, int]}`. Use this for large `--n-events`/`--max-steps` runs,
     where the full record set would otherwise scale with
     `n_events * max_steps * avg_tracks_per_event`.
+
+    There is no `mode` parameter — each stage's generative objective is read
+    directly off the model instance's own `generator_kind` (stage 1 and
+    stage 2 objectives are independent, e.g. `stage1_model.generator="flow"`
+    + `stage2_model.generator="wgan"`), and the decoder (one-shot vs
+    autoregressive) is inferred from `sec_decoder`'s own class — see
+    `sample_stage1`/`sample_stage2` (giant.sample).
+
+    `pdg_topn_map`/`mat_topn_map`/`sec_type_topn_map` serve three independent
+    purposes, no longer required to share one map (see gitea #29):
+    `pdg_topn_map`/`mat_topn_map` are required whenever
+    `particle_conditioning`/`material_conditioning` is `"onehot"` (feeds
+    `build_cond_features`'s extra `cond_cat` top-N columns); `sec_type_topn_map`/
+    `other_policy` are required instead under
+    `stage2_model.particle_type.target = "onehot"` (secondary-species
+    decode) — its class count (`stage2_model.particle_type.n_classes`) may
+    differ from `pdg_topn_map`'s. `seed` seeds the `other_policy = "sample"`
+    draw only (torch/numpy sampling itself is seeded by the caller, same as
+    today).
+
+    `l1_dist_collector`, if given, accumulates the embedding-distance
+    diagnostic across the whole run — see `L1DistCollector`. Only populated
+    under `particle_type.target = "embedding"`; a no-op otherwise.
     """
+    if particle_conditioning == "onehot" and pdg_topn_map is None:
+        raise RuntimeError(
+            "conditioning.particle.type='onehot' rollout needs pdg_topn_map "
+            "(the checkpoint's saved top-N map) — see ckpt['pdg_topn_map']"
+        )
+    if sec_decoder.particle_type_cfg.target == "onehot" and sec_type_topn_map is None:
+        raise RuntimeError(
+            "stage2_model.particle_type.target='onehot' rollout needs sec_type_topn_map "
+            "(the checkpoint's saved top-N map) — see ckpt['sec_type_topn_map']"
+        )
+    if material_conditioning == "onehot" and mat_topn_map is None:
+        raise RuntimeError(
+            "conditioning.material.type='onehot' rollout needs mat_topn_map "
+            "(the checkpoint's saved top-N map) — see ckpt['mat_topn_map']"
+        )
     device = device or torch.device("cpu")
     stage1_model.eval()
     sec_decoder.eval()
     if escape_threshold is not None:
         oracle.escape_threshold = float(escape_threshold)
+    rng = np.random.default_rng(seed)
 
     frontier, counts = make_seed_frontier(
         seeds["event_id"],
@@ -337,7 +514,7 @@ def rollout(
         seeds["pre_pos"],
         seeds["pre_E"],
         seeds["pre_dir"],
-        conditioning=conditioning,
+        particle_conditioning=particle_conditioning,
     )
     rec = _Recorder(sink=on_chunk)
 
@@ -364,8 +541,16 @@ def rollout(
                     steps,
                     device,
                     max_tracks_per_event,
-                    conditioning,
-                    mode,
+                    particle_conditioning,
+                    material_conditioning,
+                    pdg_topn_map,
+                    mat_topn_map,
+                    sec_type_topn_map,
+                    other_policy,
+                    rng,
+                    stage1_ddpm_steps,
+                    stage2_ddpm_steps,
+                    l1_dist_collector,
                 )
             )
         frontier = _concat_frontiers(next_parts)
@@ -395,8 +580,16 @@ def _step_chunk(
     steps,
     device,
     max_tracks_per_event,
-    conditioning,
-    mode="flow",
+    particle_conditioning,
+    material_conditioning,
+    pdg_topn_map,
+    mat_topn_map,
+    sec_type_topn_map,
+    other_policy,
+    rng,
+    stage1_ddpm_steps,
+    stage2_ddpm_steps,
+    l1_dist_collector,
 ) -> dict[str, np.ndarray]:
     """Advance one chunk of tracks by a single step; return the next frontier."""
     n = len(tr["event_id"])
@@ -407,7 +600,7 @@ def _step_chunk(
     tr["_material"] = material
     tr["_layer_id"] = layer_id
 
-    if conditioning == "physical":
+    if particle_conditioning == "physical":
         # Under physical-property conditioning, mass/charge (already resolved
         # on every track — see the cond_dict comment below) drive the model,
         # not a training-vocab PDG embedding — build_cond_features passes
@@ -421,33 +614,19 @@ def _step_chunk(
     # --- Pre-step termination gates (in priority order; each track picks one) ---
     stop = np.zeros(n, dtype=bool)
     escaped_sel = escaped & ~stop
-    rec.add(
-        **_terminal_rows(
-            tr, escaped_sel, TERM_ESCAPED, edep=np.zeros(int(escaped_sel.sum()))
-        )
-    )
+    rec.add(**_terminal_rows(tr, escaped_sel, TERM_ESCAPED, edep=np.zeros(int(escaped_sel.sum()))))
     stop |= escaped_sel
 
     unknown_sel = ~known_pdg & ~stop
-    rec.add(
-        **_terminal_rows(
-            tr, unknown_sel, TERM_UNKNOWN_PDG, edep=tr["pre_E"][unknown_sel]
-        )
-    )
+    rec.add(**_terminal_rows(tr, unknown_sel, TERM_UNKNOWN_PDG, edep=tr["pre_E"][unknown_sel]))
     stop |= unknown_sel
 
     cutoff_sel = (tr["pre_E"] < energy_cutoff) & ~stop
-    rec.add(
-        **_terminal_rows(
-            tr, cutoff_sel, TERM_ENERGY_CUTOFF, edep=tr["pre_E"][cutoff_sel]
-        )
-    )
+    rec.add(**_terminal_rows(tr, cutoff_sel, TERM_ENERGY_CUTOFF, edep=tr["pre_E"][cutoff_sel]))
     stop |= cutoff_sel
 
     maxstep_sel = (tr["step_in_track"] >= max_steps) & ~stop
-    rec.add(
-        **_terminal_rows(tr, maxstep_sel, TERM_MAX_STEPS, edep=tr["pre_E"][maxstep_sel])
-    )
+    rec.add(**_terminal_rows(tr, maxstep_sel, TERM_MAX_STEPS, edep=tr["pre_E"][maxstep_sel]))
     stop |= maxstep_sel
 
     active = ~stop
@@ -475,60 +654,60 @@ def _step_chunk(
         "charge": tr["charge"],
     }
     cond_cont, cond_cat = build_cond_features(
-        cond_dict, pdg_map, mat_map, cond_norm, conditioning=conditioning
+        cond_dict,
+        pdg_map,
+        mat_map,
+        cond_norm,
+        particle_conditioning=particle_conditioning,
+        material_conditioning=material_conditioning,
+        pdg_topn_map=pdg_topn_map.class_map if particle_conditioning == "onehot" else None,
+        mat_topn_map=mat_topn_map.class_map if material_conditioning == "onehot" else None,
     )
     cc = torch.from_numpy(cond_cont).float().to(device)
     ck = torch.from_numpy(cond_cat).long().to(device)
 
-    if mode == "wgan":
-        stage1_norm, n_sec_pred = sample_wgan(stage1_model, cc, ck)
-    else:
-        stage1_norm, n_sec_pred = sample_flow(stage1_model, cc, ck, steps=steps)
+    stage1_norm, n_sec_pred_stage1 = sample_stage1(stage1_model, cc, ck, steps, stage1_ddpm_steps)
     raw = tgt_norm.inverse_transform(stage1_norm.cpu().numpy())
 
     step_length = inv_log_transform(raw[:, 0])
     edep, e_sec, post_E, _delta = energy_simplex_decode(raw[:, 1:3], tr["pre_E"])
 
     post_dir_local = raw[:, 3:6].copy()
-    post_dir_local /= np.clip(
-        np.linalg.norm(post_dir_local, axis=1, keepdims=True), 1e-8, None
-    )
+    post_dir_local /= np.clip(np.linalg.norm(post_dir_local, axis=1, keepdims=True), 1e-8, None)
     post_dir_world = inv_local_frame_rotation(tr["pre_dir"], post_dir_local)
 
     travel_dir_local = raw[:, 6:9].copy()
-    travel_dir_local /= np.clip(
-        np.linalg.norm(travel_dir_local, axis=1, keepdims=True), 1e-8, None
-    )
-    post_pos = reconstruct_post_pos(
-        tr["pre_pos"], tr["pre_dir"], step_length, travel_dir_local
-    )
+    travel_dir_local /= np.clip(np.linalg.norm(travel_dir_local, axis=1, keepdims=True), 1e-8, None)
+    post_pos = reconstruct_post_pos(tr["pre_pos"], tr["pre_dir"], step_length, travel_dir_local)
 
+    n_sec_pred = resolve_n_sec(stage1_model, sec_decoder, cc, ck, stage1_norm, n_sec_pred_stage1)
     n_sec_np = n_sec_pred.cpu().numpy().astype(np.int64)
 
     # --- Secondaries ---
-    # No snapping: sec_mass/sec_charge are the model's raw predicted physical
-    # identity, used as-is for the spawned track's own future conditioning.
-    # sec_pdg_code below is a *separate*, reporting-only nearest-known-PDG
-    # label (never fed back into the model) — see giant/particles.py.
-    if mode == "wgan":
-        sec_cont, sec_phys, _valid = sample_secondaries_wgan(
-            sec_decoder, cc, ck, stage1_norm, n_sec_pred
-        )
-    else:
-        sec_cont, sec_phys, _valid = sample_secondaries(
-            sec_decoder, cc, ck, stage1_norm, n_sec_pred, steps=steps
-        )
-    sec_full = torch.cat([sec_cont, sec_phys], dim=-1).cpu().numpy()
-    sec_E, sec_dir_world, sec_mass, sec_charge, sec_valid = decode_secondaries(
-        sec_full,
+    # No snapping for "physical"/history-facing state elsewhere in the
+    # pipeline: sec_mass/sec_charge (or, for "onehot"/"embedding", the
+    # resolved sec_pdg -> real physics) are the secondary's identity, used
+    # as-is for the spawned track's own future conditioning — see
+    # decode_secondary_identity's docstring for how each
+    # particle_type.target differs on whether PDG resolution is a real
+    # identity decision or just a reporting label.
+    sec_cont, sec_type, _valid = sample_stage2(sec_decoder, cc, ck, stage1_norm, n_sec_pred, steps)
+    sec_E, sec_dir_world, sec_mass, sec_charge, sec_pdg_code, sec_type_l1_dist = decode_secondary_identity(
+        sec_decoder,
+        sec_cont,
+        sec_type,
         n_sec_np,
         e_sec,
         tr["pre_dir"],
-        sec_phys_normalizer=sec_phys_norm,
+        sec_phys_norm,
+        pdg_map,
+        sec_type_topn_map,
+        other_policy,
+        rng,
     )
-    sec_pdg_code = nearest_known_pdg(
-        sec_mass.reshape(-1), sec_charge.reshape(-1), pdg_map.keys()
-    ).reshape(sec_mass.shape)
+    sec_valid = np.arange(sec_E.shape[1])[None, :] < n_sec_np[:, None]
+    if l1_dist_collector is not None and sec_type_l1_dist is not None:
+        l1_dist_collector.add(sec_type_l1_dist, sec_valid)
 
     edep = edep.astype(np.float64)
     post_E = post_E.astype(np.float64)

giant/sample.py

@@ -1,26 +1,22 @@
 import torch
+import torch.nn.functional as F
 
-from giant.constants import K_MAX, SEC_DIM, SEC_SLOT_DIM, X_DIM
+from giant.constants import CONT_SLOT_DIM, X_DIM
+from giant.model.network import DdpmObjective, Stage2Autoregressive, build_objective, stage2_trunk_sec_dim
+from giant.model.schedule import CosineSchedule
 
 
-def _slots_from_flat(
-    x: torch.Tensor, n_sec_pred: torch.Tensor
-) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    """Reshape a flat (B, SEC_DIM) decoder output into per-slot tensors.
-
-    Returns (sec_cont, sec_phys, sec_valid) — see `sample_secondaries`'s
-    docstring for their shapes/meaning. Shared by both the flow-matching and
-    WGAN Stage-2 samplers, which differ only in how `x` was produced.
-    """
-    B = x.size(0)
-    device = x.device
-    x_slots = x.view(B, K_MAX, SEC_SLOT_DIM)
-    sec_cont = x_slots[:, :, :4]
-    sec_phys = x_slots[:, :, 4:]
-    sec_valid = torch.arange(K_MAX, device=device).unsqueeze(0) < n_sec_pred.unsqueeze(
-        1
-    )
-    return sec_cont, sec_phys, sec_valid
+def _predict_n_sec_if_owned(
+    model: torch.nn.Module, cond_cont: torch.Tensor, cond_cat: torch.Tensor
+) -> torch.Tensor | None:
+    """Stage-1 `n_sec_head` is only present on a migrated v0.2 checkpoint
+    (fresh runs move it to stage 2 — see `Stage1Model`'s docstring). `None`
+    here means "ask stage 2 instead", which every caller (`giant/rollout.py`,
+    `giant/cli.py`) must do for a fresh checkpoint."""
+    if getattr(model, "n_sec_head", None) is None:
+        return None
+    logits = model.predict_n_sec(cond_cont, cond_cat)
+    return logits.argmax(dim=-1)
 
 
 @torch.no_grad()
@@ -29,12 +25,14 @@ def sample_flow(
     cond_cont: torch.Tensor,
     cond_cat: torch.Tensor,
     steps: int = 10,
-) -> tuple[torch.Tensor, torch.Tensor]:
+) -> tuple[torch.Tensor, torch.Tensor | None]:
     """Euler integration of the Stage-1 vector field from t=0 to t=1.
 
     Returns (primary_sample, n_sec_pred):
         primary_sample: (B, X_DIM) — normalised 9D primary post-step output
-        n_sec_pred: (B,) int64 — predicted secondary count
+        n_sec_pred: (B,) int64 — predicted secondary count, or `None` if
+            `model` has no `n_sec_head` (a fresh v0.3.0 Stage1Model — see
+            `_predict_n_sec_if_owned`).
     """
     model.eval()
     B = cond_cont.size(0)
@@ -43,11 +41,138 @@ def sample_flow(
     dt = 1.0 / steps
     for i in range(steps):
         t = torch.full((B,), i * dt, device=device)
-        v = model(x, t, cond_cont, cond_cat)
+        v = model(x, cond_cont, cond_cat, t=t)
         x = x + v * dt
-    n_sec_logits = model.predict_n_sec(cond_cont, cond_cat)
-    n_sec_pred = n_sec_logits.argmax(dim=-1)
-    return x, n_sec_pred
+    return x, _predict_n_sec_if_owned(model, cond_cont, cond_cat)
+
+
+@torch.no_grad()
+def sample_ddpm(
+    model: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    schedule,
+) -> tuple[torch.Tensor, torch.Tensor | None]:
+    """Full DDPM ancestral sampling (T reverse steps). Returns (sample, n_sec_pred)
+    — see `sample_flow`'s docstring for the `n_sec_pred` `None` case."""
+    model.eval()
+    B = cond_cont.size(0)
+    device = cond_cont.device
+    x = torch.randn(B, X_DIM, device=device)
+    T = schedule.T
+    for i in reversed(range(T)):
+        t_norm = torch.full((B,), i / T, device=device)
+        eps_pred = model(x, cond_cont, cond_cat, t=t_norm)
+        beta = schedule.betas[i]
+        alpha = schedule.alphas[i]
+        alpha_bar = schedule.alpha_bars[i]
+        z = torch.randn_like(x) if i > 0 else torch.zeros_like(x)
+        x = (1.0 / alpha.sqrt()) * (x - (1.0 - alpha) / (1.0 - alpha_bar).sqrt() * eps_pred) + beta.sqrt() * z
+    return x, _predict_n_sec_if_owned(model, cond_cont, cond_cat)
+
+
+@torch.no_grad()
+def sample_ddim(
+    model: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    schedule,
+    steps: int = 50,
+) -> tuple[torch.Tensor, torch.Tensor | None]:
+    """DDIM deterministic sampling (Song et al. 2020). Returns (sample, n_sec_pred)
+    — see `sample_flow`'s docstring for the `n_sec_pred` `None` case."""
+    model.eval()
+    B = cond_cont.size(0)
+    device = cond_cont.device
+    T = schedule.T
+    timesteps = torch.linspace(T - 1, 0, steps, dtype=torch.long, device=device)
+    x = torch.randn(B, X_DIM, device=device)
+    for step_idx, ts in enumerate(timesteps):
+        t_idx = int(ts.item())
+        t_norm = torch.full((B,), t_idx / T, device=device)
+        eps_pred = model(x, cond_cont, cond_cat, t=t_norm)
+        ab_t = schedule.alpha_bars[t_idx]
+        if step_idx + 1 < len(timesteps):
+            ab_prev = schedule.alpha_bars[int(timesteps[step_idx + 1].item())]
+        else:
+            ab_prev = torch.ones(1, device=device)
+        x0_pred = (x - (1.0 - ab_t).sqrt() * eps_pred) / ab_t.sqrt()
+        x = ab_prev.sqrt() * x0_pred + (1.0 - ab_prev).sqrt() * eps_pred
+    return x, _predict_n_sec_if_owned(model, cond_cont, cond_cat)
+
+
+@torch.no_grad()
+def sample_wgan(
+    generator: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor | None]:
+    """Single-pass Stage-1 WGAN generator sample. Returns (sample, n_sec_pred)
+    — see `sample_flow`'s docstring for the `n_sec_pred` `None` case."""
+    generator.eval()
+    B = cond_cont.size(0)
+    z = torch.randn(B, generator.noise_dim, device=cond_cont.device)
+    x = generator(z, cond_cont, cond_cat)
+    return x, _predict_n_sec_if_owned(generator, cond_cont, cond_cat)
+
+
+def _stage2_flat_width(sec_decoder: torch.nn.Module) -> int:
+    """The width of `sec_decoder`'s own trunk in/out vector — folded
+    (continuous + type) under `particle_type.target = "physical"` or
+    `generator = "wgan"`, continuous-only otherwise (the type slice then
+    comes from `predict_type` instead — see `stage2_trunk_sec_dim`'s
+    docstring)."""
+    return stage2_trunk_sec_dim(
+        sec_decoder.particle_type_cfg,
+        sec_decoder.generator_kind,
+        sec_decoder.k_max,
+        sec_decoder.type_dim,
+    )
+
+
+def _type_folded(sec_decoder: torch.nn.Module) -> bool:
+    target = sec_decoder.particle_type_cfg.target
+    return target == "physical" or build_objective(sec_decoder.generator_kind).folds_type_slice
+
+
+def _decode_stage2_flat(
+    sec_decoder: torch.nn.Module,
+    x: torch.Tensor,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    stage1_out: torch.Tensor,
+    n_sec_pred: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Reshape a flat `(B, flat_width)` `Stage2OneShot` output into per-slot
+    tensors, generator/`particle_type.target`-agnostic: shared by
+    `sample_secondaries`/`sample_secondaries_wgan`, which differ only in how
+    `x` was produced.
+
+    Returns (sec_cont, sec_type, sec_valid):
+        sec_cont:  (B, k_max, CONT_SLOT_DIM) — [stick_logit, local_dir]
+        sec_type:  (B, k_max, type_dim) — under `target="physical"` this is
+            [log_mass, charge] (normalised iff the checkpoint's sec_phys
+            normalizer was applied at training time — denormalize before
+            treating as physical units; see
+            giant.data.transforms.decode_secondaries); under `"onehot"` /
+            `"embedding"` it is raw class logits / an embedding-space vector
+            — decode via giant.particles.decode_topn_class /
+            decode_embedding_nearest (see giant/rollout.py).
+        sec_valid: (B, k_max) bool — True for slots i < n_sec_pred
+    """
+    B = x.size(0)
+    device = x.device
+    k_max = sec_decoder.k_max
+    type_dim = sec_decoder.type_dim
+    if _type_folded(sec_decoder):
+        x_slots = x.view(B, k_max, CONT_SLOT_DIM + type_dim)
+        sec_cont = x_slots[:, :, :CONT_SLOT_DIM]
+        sec_type = x_slots[:, :, CONT_SLOT_DIM:]
+    else:
+        sec_cont = x.view(B, k_max, CONT_SLOT_DIM)
+        sec_type = sec_decoder.predict_type(cond_cont, cond_cat, stage1_out)
+    sec_valid = torch.arange(k_max, device=device).unsqueeze(0) < n_sec_pred.unsqueeze(1)
+    return sec_cont, sec_type, sec_valid
 
 
 @torch.no_grad()
@@ -59,76 +184,25 @@ def sample_secondaries(
     n_sec_pred: torch.Tensor,
     steps: int = 10,
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    """Euler integration of the Stage-2 vector field; return raw slot outputs.
+    """Euler integration of `Stage2OneShot`'s (flow/ddpm) vector field; return
+    raw slot outputs — see `_decode_stage2_flat`'s docstring for the returned
+    (sec_cont, sec_type, sec_valid) shapes/meaning.
 
     n_sec_pred: (B,) int64 — number of valid secondaries per step
-
-    Returns (sec_cont, sec_phys, sec_valid):
-        sec_cont:  (B, K_MAX, 4) — [stick_logit, local_dir_x, local_dir_y, local_dir_z]
-        sec_phys:  (B, K_MAX, PARTICLE_PHYS_DIM) — predicted [log_mass, charge]
-                   per slot (normalised iff the checkpoint's sec_phys
-                   normalizer was applied at training time — denormalize
-                   before treating as physical units; see
-                   giant.data.transforms.decode_secondaries). Used as-is —
-                   no snapping to a discrete PDG code.
-        sec_valid: (B, K_MAX) bool — True for slots i < n_sec_pred
     """
     sec_decoder.eval()
     B = cond_cont.size(0)
     device = cond_cont.device
+    flat_width = _stage2_flat_width(sec_decoder)
 
-    x = torch.randn(B, SEC_DIM, device=device)
+    x = torch.randn(B, flat_width, device=device)
     dt = 1.0 / steps
     for i in range(steps):
         t = torch.full((B,), i * dt, device=device)
-        v = sec_decoder(x, t, cond_cont, cond_cat, stage1_out)
+        v = sec_decoder(x, cond_cont, cond_cat, stage1_out, t=t)
         x = x + v * dt
 
-    return _slots_from_flat(x, n_sec_pred)
-
-
-@torch.no_grad()
-def sample_ddpm(
-    model: torch.nn.Module,
-    cond_cont: torch.Tensor,
-    cond_cat: torch.Tensor,
-    schedule,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Full DDPM ancestral sampling (T reverse steps). Returns (sample, n_sec_pred)."""
-    model.eval()
-    B = cond_cont.size(0)
-    device = cond_cont.device
-    x = torch.randn(B, X_DIM, device=device)
-    T = schedule.T
-    for i in reversed(range(T)):
-        t_norm = torch.full((B,), i / T, device=device)
-        eps_pred = model(x, t_norm, cond_cont, cond_cat)
-        beta = schedule.betas[i]
-        alpha = schedule.alphas[i]
-        alpha_bar = schedule.alpha_bars[i]
-        z = torch.randn_like(x) if i > 0 else torch.zeros_like(x)
-        x = (1.0 / alpha.sqrt()) * (
-            x - (1.0 - alpha) / (1.0 - alpha_bar).sqrt() * eps_pred
-        ) + beta.sqrt() * z
-    n_sec_logits = model.predict_n_sec(cond_cont, cond_cat)
-    n_sec_pred = n_sec_logits.argmax(dim=-1)
-    return x, n_sec_pred
-
-
-@torch.no_grad()
-def sample_wgan(
-    generator: torch.nn.Module,
-    cond_cont: torch.Tensor,
-    cond_cat: torch.Tensor,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Single-pass Stage-1 WGAN generator sample. Returns (sample, n_sec_pred)."""
-    generator.eval()
-    B = cond_cont.size(0)
-    z = torch.randn(B, generator.noise_dim, device=cond_cont.device)
-    x = generator(z, cond_cont, cond_cat)
-    n_sec_logits = generator.predict_n_sec(cond_cont, cond_cat)
-    n_sec_pred = n_sec_logits.argmax(dim=-1)
-    return x, n_sec_pred
+    return _decode_stage2_flat(sec_decoder, x, cond_cont, cond_cat, stage1_out, n_sec_pred)
 
 
 @torch.no_grad()
@@ -139,41 +213,228 @@ def sample_secondaries_wgan(
     stage1_out: torch.Tensor,
     n_sec_pred: torch.Tensor,
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    """Single-pass Stage-2 WGAN generator sample; see `sample_secondaries`'s
-    docstring for the returned (sec_cont, sec_phys, sec_valid) shapes."""
+    """Single-pass `Stage2OneShot` WGAN generator sample; see
+    `_decode_stage2_flat`'s docstring for the returned (sec_cont, sec_type,
+    sec_valid) shapes/meaning."""
     sec_decoder.eval()
     B = cond_cont.size(0)
     z = torch.randn(B, sec_decoder.noise_dim, device=cond_cont.device)
     x = sec_decoder(z, cond_cont, cond_cat, stage1_out)
-    return _slots_from_flat(x, n_sec_pred)
+    return _decode_stage2_flat(sec_decoder, x, cond_cont, cond_cat, stage1_out, n_sec_pred)
 
 
 @torch.no_grad()
-def sample_ddim(
-    model: torch.nn.Module,
+def sample_secondaries_ar(
+    sec_decoder: torch.nn.Module,
     cond_cont: torch.Tensor,
     cond_cat: torch.Tensor,
-    schedule,
-    steps: int = 50,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """DDIM deterministic sampling (Song et al. 2020). Returns (sample, n_sec_pred)."""
-    model.eval()
+    stage1_out: torch.Tensor,
+    n_sec_pred: torch.Tensor,
+    steps: int = 10,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """`Stage2Autoregressive` inference loop: one token at a time, in
+    descending-energy slot order, `k_max` sequential calls. Unlike training
+    (teacher forcing — a single parallel pass over ground-truth tokens, see
+    `giant.training.stage2_inputs._assemble_stage2_ar_inputs`), there is no
+    ground truth at inference: each token's conditioning is built
+    free-running, from the PREVIOUS TOKEN'S OWN just-generated output — the
+    train/inference gap that is the cost of markov history's
+    expressiveness.
+
+    A `{flow,ddpm}` token costs `steps` ODE substeps; `wgan` costs one pass —
+    the "K sequential forwards" cost applies per-token here, not
+    once, so a flow/ddpm AR run costs ~`k_max * steps` model calls per
+    physics step.
+
+    Under `history="attention"` the history encoding is computed once per
+    slot via `Stage2Autoregressive.history_step` (a KV-cache append)
+    rather than re-derived by every model call inside that slot — so an ODE
+    loop's `steps` substeps, and the separate `predict_type` call when the
+    type slice isn't folded into the trunk output, all reuse the SAME `hist`
+    tensor for a given `k`. Recomputing per call instead would be merely
+    wasteful under markov (its per-call cost is already O(1)) but wrong under
+    attention: `AttentionHistory.step` mutates the cache by appending, so
+    calling it more than once per slot would double-count that slot's own
+    (not-yet-existing) predecessor.
+
+    The free-running history feature stays UNSNAPPED (mirrors the
+    established "no snapping" precedent for `particle_type.target =
+    "physical"` secondaries feeding their own future conditioning):
+    `"physical"` carries the raw (log_mass, charge) forward as-is;
+    `"embedding"` carries the raw predicted vector as-is; `"onehot"` is the
+    one exception — its history slot must be a probability-simplex-shaped
+    vector (that's what `MarkovHistory`/`AttentionHistory` were trained on,
+    `_type_repr`'s `F.one_hot` ground truth), so it's the hard one-hot of
+    `argmax(logits)`, not the raw logits themselves. Discretizing further,
+    into a concrete PDG code, only ever happens once — at secondary-spawn
+    time in `giant/rollout.py` — never inside this loop.
+
+    Returns (sec_cont, sec_type, sec_valid) — same shapes/meaning as
+    `sample_secondaries`/`sample_secondaries_wgan`'s (see
+    `_decode_stage2_flat`'s docstring); `sec_type` is raw per-slot output in
+    all three `particle_type.target` cases (never one-hot-collapsed), so the
+    caller decodes it exactly the same way regardless of which decoder
+    produced it.
+    """
+    sec_decoder.eval()
     B = cond_cont.size(0)
     device = cond_cont.device
-    T = schedule.T
-    timesteps = torch.linspace(T - 1, 0, steps, dtype=torch.long, device=device)
-    x = torch.randn(B, X_DIM, device=device)
-    for step_idx, ts in enumerate(timesteps):
-        t_idx = int(ts.item())
-        t_norm = torch.full((B,), t_idx / T, device=device)
-        eps_pred = model(x, t_norm, cond_cont, cond_cat)
-        ab_t = schedule.alpha_bars[t_idx]
-        if step_idx + 1 < len(timesteps):
-            ab_prev = schedule.alpha_bars[int(timesteps[step_idx + 1].item())]
+    k_max = sec_decoder.k_max
+    type_dim = sec_decoder.type_dim
+    objective = build_objective(sec_decoder.generator_kind)
+    target = sec_decoder.particle_type_cfg.target
+    type_folded = _type_folded(sec_decoder)
+    token_dim = CONT_SLOT_DIM + type_dim if type_folded else CONT_SLOT_DIM
+
+    sec_cont = torch.zeros(B, k_max, CONT_SLOT_DIM, device=device)
+    sec_type = torch.zeros(B, k_max, type_dim, device=device)
+
+    # Running per-token state, threaded from one slot to the next.
+    prev_repr = torch.zeros(B, CONT_SLOT_DIM + type_dim, device=device)
+    remaining = torch.ones(B, device=device)
+    history_cache = sec_decoder.init_history_cache()
+
+    for k in range(k_max):
+        has_prev = torch.full((B, 1), k >= 1, dtype=torch.bool, device=device)
+        history_feat = prev_repr.unsqueeze(1)  # (B, 1, CONT_SLOT_DIM + type_dim)
+        remaining_frac = remaining.unsqueeze(1)  # (B, 1)
+        slot_idx = torch.full((B, 1), k / max(k_max - 1, 1), device=device, dtype=torch.float32)
+        hist, history_cache = sec_decoder.history_step(history_feat, has_prev, history_cache)
+
+        if objective.is_adversarial:
+            z = torch.randn(B, 1, sec_decoder.noise_dim, device=device)
+            token = sec_decoder(
+                z,
+                cond_cont,
+                cond_cat,
+                stage1_out,
+                history_feat,
+                has_prev,
+                remaining_frac,
+                slot_idx,
+                hist=hist,
+            )
         else:
-            ab_prev = torch.ones(1, device=device)
-        x0_pred = (x - (1.0 - ab_t).sqrt() * eps_pred) / ab_t.sqrt()
-        x = ab_prev.sqrt() * x0_pred + (1.0 - ab_prev).sqrt() * eps_pred
-    n_sec_logits = model.predict_n_sec(cond_cont, cond_cat)
-    n_sec_pred = n_sec_logits.argmax(dim=-1)
-    return x, n_sec_pred
+            x = torch.randn(B, 1, token_dim, device=device)
+            dt = 1.0 / steps
+            for i in range(steps):
+                t = torch.full((B, 1), i * dt, device=device)
+                v = sec_decoder(
+                    x,
+                    cond_cont,
+                    cond_cat,
+                    stage1_out,
+                    history_feat,
+                    has_prev,
+                    remaining_frac,
+                    slot_idx,
+                    t=t,
+                    hist=hist,
+                )
+                x = x + v * dt
+            token = x
+
+        token = token.squeeze(1)  # (B, token_dim)
+        cont_k = token[:, :CONT_SLOT_DIM]
+        if type_folded:
+            type_k = token[:, CONT_SLOT_DIM:]
+        else:
+            type_k = sec_decoder.predict_type(
+                cond_cont,
+                cond_cat,
+                stage1_out,
+                history_feat,
+                has_prev,
+                remaining_frac,
+                slot_idx,
+                hist=hist,
+            ).squeeze(1)
+
+        sec_cont[:, k] = cont_k
+        sec_type[:, k] = type_k
+
+        if target == "onehot":
+            type_for_history = F.one_hot(type_k.argmax(dim=-1), num_classes=type_dim).float()
+        else:
+            type_for_history = type_k
+
+        stick_fraction = torch.sigmoid(cont_k[:, 0])
+        prev_repr = torch.cat([stick_fraction.unsqueeze(-1), cont_k[:, 1:4], type_for_history], dim=-1)
+        remaining = torch.clamp(remaining * (1.0 - stick_fraction), min=0.0)
+
+    sec_valid = torch.arange(k_max, device=device).unsqueeze(0) < n_sec_pred.unsqueeze(1)
+    return sec_cont, sec_type, sec_valid
+
+
+# ---------------------------------------------------------------------------
+# Per-stage dispatch — shared by giant/rollout.py and giant/cli.py's
+# `predict` command, since both need "given a stage model, produce a
+# sample" without hand-picking the sampler themselves (each stage's
+# generative objective is independent, read off the model's own
+# `generator_kind`, not a caller-supplied `mode` string).
+# ---------------------------------------------------------------------------
+
+
+def sample_stage1(
+    stage1_model: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    steps: int,
+    ddpm_steps: int = 1000,
+) -> tuple[torch.Tensor, torch.Tensor | None]:
+    """Dispatches on `stage1_model.generator_kind`."""
+    objective = build_objective(stage1_model.generator_kind)
+    if objective.is_adversarial:
+        return sample_wgan(stage1_model, cond_cont, cond_cat)
+    if isinstance(objective, DdpmObjective):
+        schedule = CosineSchedule(T=ddpm_steps).to(cond_cont.device)
+        return sample_ddpm(stage1_model, cond_cont, cond_cat, schedule)
+    return sample_flow(stage1_model, cond_cont, cond_cat, steps=steps)
+
+
+def sample_stage2(
+    sec_decoder: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    stage1_out: torch.Tensor,
+    n_sec_pred: torch.Tensor,
+    steps: int,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Dispatches on `decoder` (one-shot vs autoregressive — the class
+    itself, via `isinstance`) and `sec_decoder.generator_kind` (flow/ddpm/
+    wgan). DDPM secondaries aren't supported — no `Stage2*` class was ever
+    built with `generator="ddpm"` in practice and `flow_matching_loss_secondary*`
+    is the only stage-2 training path that exists for the non-adversarial
+    case, so there's nothing to dispatch to here.
+    """
+    if isinstance(sec_decoder, Stage2Autoregressive):
+        return sample_secondaries_ar(sec_decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=steps)
+    if build_objective(sec_decoder.generator_kind).is_adversarial:
+        return sample_secondaries_wgan(sec_decoder, cond_cont, cond_cat, stage1_out, n_sec_pred)
+    return sample_secondaries(sec_decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=steps)
+
+
+def resolve_n_sec(
+    stage1_model: torch.nn.Module,
+    sec_decoder: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    stage1_out: torch.Tensor,
+    n_sec_pred: torch.Tensor | None,
+) -> torch.Tensor:
+    """`n_sec_pred` is already populated when `stage1_model` owns a legacy
+    `n_sec_head` (a migrated v0.2 checkpoint — see `Stage1Model`'s
+    docstring); otherwise ask stage 2, which owns it by default. Raises if
+    neither stage owns a head at all — the only way that happens is
+    `stage2_model.n_sec.mode` other than `"head"` (`"truth"`/`"stop_token"`),
+    neither of which is a valid rollout-/predict-capable checkpoint."""
+    if n_sec_pred is not None:
+        return n_sec_pred
+    if getattr(sec_decoder, "n_sec_head", None) is None:
+        raise RuntimeError(
+            "checkpoint has no n_sec_head on either stage — needs "
+            "stage2_model.n_sec.mode = 'head' (the default); 'truth' is "
+            "standalone-evaluation-only and 'stop_token' isn't implemented"
+        )
+    logits = sec_decoder.predict_n_sec(cond_cont, cond_cat, stage1_out)
+    return logits.argmax(dim=-1)

giant/tools/bump_dataset_version.py

@@ -1,5 +1,5 @@
 """Cut a new raw generation or processed schema version for the geant_steps
-dataset tree (see scripts/migrate_geant_steps.py for the layout):
+dataset tree (see giant/tools/migrate_geant_steps.py for the layout):
 
     raw/<kind>/<gen>/<detector>/shard-NNN.root
     processed/<kind>/<gen>/<schema>/<detector>/shard-NNN.parquet
@@ -82,9 +82,7 @@ def _max_index(parent: Path, pattern: re.Pattern) -> int:
 
 def _git_user_name() -> str | None:
     try:
-        out = subprocess.run(
-            ["git", "config", "user.name"], capture_output=True, text=True, timeout=2
-        )
+        out = subprocess.run(["git", "config", "user.name"], capture_output=True, text=True, timeout=2)
     except (OSError, subprocess.SubprocessError):
         # OSError (e.g. git not on PATH) and subprocess.SubprocessError
         # (e.g. TimeoutExpired) are unrelated hierarchies — TimeoutExpired
@@ -142,9 +140,7 @@ def plan_bump_schema(
     raw_gen_dir = root / "raw" / kind / gen_tag
     processed_gen_dir = root / "processed" / kind / gen_tag
     if not raw_gen_dir.is_dir() and not processed_gen_dir.is_dir():
-        raise SystemExit(
-            f"error: {gen_tag} doesn't exist yet for kind={kind} — run bump-gen first"
-        )
+        raise SystemExit(f"error: {gen_tag} doesn't exist yet for kind={kind} — run bump-gen first")
     if target is not None:
         if not SCHEMA_RE.match(target):
             raise SystemExit(f"error: --to must look like 'schemaN', got {target!r}")
@@ -154,9 +150,7 @@ def plan_bump_schema(
         schema_tag = f"schema{next_schema}"
     new_dirs = [processed_gen_dir / schema_tag]
     by_suffix = f" ({by})" if by else ""
-    log_line = (
-        f"- `{gen_tag}`/`{schema_tag}` (kind={kind}) — {date} — {reason}{by_suffix}"
-    )
+    log_line = f"- `{gen_tag}`/`{schema_tag}` (kind={kind}) — {date} — {reason}{by_suffix}"
     return new_dirs, log_line
 
 
@@ -212,9 +206,7 @@ def _manifest_referenced_files(pools_root: Path) -> set[Path]:
     return referenced
 
 
-def _referenced_root_count(
-    raw_gen_dir: Path, processed_gen_dir: Path
-) -> tuple[int, int]:
+def _referenced_root_count(raw_gen_dir: Path, processed_gen_dir: Path) -> tuple[int, int]:
     """(total .root files, count with a same-named .parquet under any schema) for one gen."""
     if not raw_gen_dir.is_dir():
         return 0, 0
@@ -243,9 +235,7 @@ def _referenced_root_count(
     return total, referenced
 
 
-def _referenced_parquet_count(
-    schema_dir: Path, manifest_referenced: set[Path]
-) -> tuple[int, int]:
+def _referenced_parquet_count(schema_dir: Path, manifest_referenced: set[Path]) -> tuple[int, int]:
     """(total .parquet files, count listed in at least one manifest) for one schema dir."""
     if not schema_dir.is_dir():
         return 0, 0
@@ -342,11 +332,7 @@ def print_status(root: Path) -> None:
     grand_files = 0
     for kind_dir in sorted(p for p in raw_root.iterdir() if p.is_dir()):
         kind = kind_dir.name
-        gens = sorted(
-            int(m.group(1))
-            for m in (GEN_RE.match(p.name) for p in kind_dir.iterdir() if p.is_dir())
-            if m
-        )
+        gens = sorted(int(m.group(1)) for m in (GEN_RE.match(p.name) for p in kind_dir.iterdir() if p.is_dir()) if m)
         print(_colorize(f"{kind}/", "kind"))
         kind_total = 0
         kind_files = 0
@@ -359,25 +345,18 @@ def print_status(root: Path) -> None:
             schemas = sorted(
                 int(m.group(1))
                 for m in (
-                    SCHEMA_RE.match(p.name)
-                    for p in (schema_dir.iterdir() if schema_dir.is_dir() else [])
-                    if p.is_dir()
+                    SCHEMA_RE.match(p.name) for p in (schema_dir.iterdir() if schema_dir.is_dir() else []) if p.is_dir()
                 )
                 if m
             )
             schema_sizes = {s: _du(schema_dir / f"schema{s}") for s in schemas}
             schema_counts = {
-                s: _referenced_parquet_count(
-                    schema_dir / f"schema{s}", manifest_referenced
-                )
-                for s in schemas
+                s: _referenced_parquet_count(schema_dir / f"schema{s}", manifest_referenced) for s in schemas
             }
             processed_size = sum(schema_sizes.values())
             processed_files = sum(c[0] for c in schema_counts.values())
             processed_referenced = sum(c[1] for c in schema_counts.values())
-            raw_files, raw_referenced = _referenced_root_count(
-                raw_gen_dir, processed_gen_dir
-            )
+            raw_files, raw_referenced = _referenced_root_count(raw_gen_dir, processed_gen_dir)
             gen_total = raw_size + processed_size
             gen_files = raw_files + processed_files
             kind_total += gen_total
@@ -425,9 +404,7 @@ def print_status(root: Path) -> None:
                         print(_reason_line(schema_reason, indent=4))
             else:
                 print(_colorize("      (none)", "schema"))
-        print(
-            _row(f"{kind} total", kind_total, indent=1, level="gen", count=kind_files)
-        )
+        print(_row(f"{kind} total", kind_total, indent=1, level="gen", count=kind_files))
         print()
         grand_total += kind_total
         grand_files += kind_files
@@ -526,13 +503,8 @@ def plan_update_manifest(
     return result, missing
 
 
-def apply_update_manifest(
-    manifest_path: Path, lines: list[tuple[str, str | None]]
-) -> None:
-    out = [
-        replacement if replacement is not None else original
-        for original, replacement in lines
-    ]
+def apply_update_manifest(manifest_path: Path, lines: list[tuple[str, str | None]]) -> None:
+    out = [replacement if replacement is not None else original for original, replacement in lines]
     manifest_path.write_text("\n".join(out) + "\n")
 
 
@@ -552,9 +524,7 @@ def _resolve_manifest_files(manifest_path: Path) -> list[Path]:
     return files
 
 
-def plan_create_manifest(
-    output_path: Path, parquet_files: list[Path]
-) -> tuple[list[str], list[Path], list[Path]]:
+def plan_create_manifest(output_path: Path, parquet_files: list[Path]) -> tuple[list[str], list[Path], list[Path]]:
     """Return (relative_lines, missing_files, resolved_abs_paths)."""
     manifest_dir = output_path.resolve().parent
     lines: list[str] = []
@@ -569,9 +539,7 @@ def plan_create_manifest(
     return lines, missing, resolved
 
 
-def check_holdout_overlap(
-    output_path: Path, resolved_new_files: list[Path]
-) -> list[tuple[str, Path]]:
+def check_holdout_overlap(output_path: Path, resolved_new_files: list[Path]) -> list[tuple[str, Path]]:
     """Return (other_manifest_name, file) pairs where new files clash with existing manifests.
 
     The check is triggered when output_path is (or will be) holdout.manifest, or when a
@@ -611,7 +579,7 @@ def apply_create_manifest(output_path: Path, lines: list[str]) -> None:
 
 
 # ---------------------------------------------------------------------------
-# CLI entry points (called from scripts/dwarf.py)
+# CLI entry points (called from giant/tools/dwarf.py)
 # ---------------------------------------------------------------------------
 
 
@@ -641,9 +609,7 @@ def _run_bump(
     if gen is None:
         new_dirs, log_line = plan_bump_gen(root_path, kind, reason, by, date, to)
     else:
-        new_dirs, log_line = plan_bump_schema(
-            root_path, kind, gen, reason, by, date, to
-        )
+        new_dirs, log_line = plan_bump_schema(root_path, kind, gen, reason, by, date, to)
 
     print(f"=== {'EXECUTING' if execute else 'DRY RUN'} ===")
     print("new directories:")

giant/tools/check_migration_v02_v03.py

@@ -0,0 +1,161 @@
+"""Portal-machine follow-up for v0.3.0 step 2: diff a real v0.2 checkpoint's
+outputs against the new `build_models` on the same input batch.
+
+`tests/test_migration_v02_v03.py` already proves this bit-identical with
+synthetic random weights, but that test can't run where it matters (no
+`/ceph` on local dev machines — see CLAUDE.md's Compute environment
+section). This script is the real-checkpoint counterpart: run it on a portal
+machine against an actual trained checkpoint before merging
+`v0.3.0-stage2-autoregressive` to `master`.
+
+Usage (from the repo root, on a portal machine):
+
+    uv run python giant/tools/check_migration_v02_v03.py /ceph/lbogner/.../best.pt
+    uv run python giant/tools/check_migration_v02_v03.py /ceph/lbogner/.../best.pt --ema
+    uv run python giant/tools/check_migration_v02_v03.py /ceph/lbogner/.../best.pt --batch 32 --seed 1
+
+Run it once against a flow (or ddpm) checkpoint and once against a wgan
+checkpoint ("one flow checkpoint and one WGAN checkpoint").
+A routed checkpoint (`model_config["router"]["enabled"]`) is only checked for
+successful construction — `giant.model.network.migrate_legacy_state_dict`
+doesn't yet remap routed (Expert-per-router) state dicts, so the
+bit-identical assertion is skipped with a clear warning in that case (see the
+function's own docstring for why).
+"""
+
+import argparse
+import sys
+from pathlib import Path
+
+import torch
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from giant.constants import COND_DIM, SEC_SLOT_DIM, X_DIM  # noqa: E402
+from giant.model import network as net  # noqa: E402
+from tests.legacy import network_v02_snapshot as legacy  # noqa: E402
+
+
+def _random_batch(model_config: dict, batch: int, seed: int):
+    g = torch.Generator().manual_seed(seed)
+    pdg_vocab = model_config["pdg_vocab"]
+    mat_vocab = model_config["mat_vocab"]
+    k_max = model_config.get("k_max", 15)
+    noise_dim = model_config.get("noise_dim", 64)
+
+    cond_cont = torch.randn(batch, COND_DIM, generator=g)
+    cond_cat = torch.stack(
+        [
+            torch.randint(0, pdg_vocab, (batch,), generator=g),
+            torch.randint(0, mat_vocab, (batch,), generator=g),
+        ],
+        dim=1,
+    )
+    x1 = torch.randn(batch, X_DIM, generator=g)
+    x2 = torch.randn(batch, k_max * SEC_SLOT_DIM, generator=g)
+    t = torch.rand(batch, generator=g)
+    z1 = torch.randn(batch, noise_dim, generator=g)
+    z2 = torch.randn(batch, noise_dim, generator=g)
+    return cond_cont, cond_cat, x1, x2, t, z1, z2
+
+
+def _max_abs_diff(a: torch.Tensor, b: torch.Tensor) -> float:
+    return (a - b).abs().max().item()
+
+
+def main() -> int:
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("checkpoint", type=Path, help="Path to a v0.2 best.pt/last.pt")
+    p.add_argument(
+        "--ema",
+        action="store_true",
+        help="Use the checkpoint's EMA weights (model_ema/sec_decoder_ema) — "
+        "what predict/rollout actually sample from — instead of raw weights.",
+    )
+    p.add_argument("--batch", type=int, default=16)
+    p.add_argument("--seed", type=int, default=0)
+    args = p.parse_args()
+
+    ckpt = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
+    if "model_config" not in ckpt:
+        print(f"FAIL: {args.checkpoint} has no 'model_config' key — can't migrate it")
+        return 1
+    model_config = ckpt["model_config"]
+    mode = model_config.get("mode", "flow")
+    routed = bool((model_config.get("router") or {}).get("enabled"))
+    print(f"checkpoint: {args.checkpoint}")
+    print(f"  mode={mode!r}  conditioning={model_config.get('conditioning')!r}  routed={routed}  ema={args.ema}")
+
+    stage1_key = "model_ema" if args.ema and "model_ema" in ckpt else "model"
+    stage2_key = "sec_decoder_ema" if args.ema and "sec_decoder_ema" in ckpt else "sec_decoder"
+    if args.ema and stage1_key == "model":
+        print("  warning: --ema requested but no model_ema in checkpoint, using raw weights")
+
+    # --- old side: the frozen v0.2 snapshot, loaded with the checkpoint's own weights ---
+    old_stage1, old_stage2 = legacy.build_models(model_config)
+    old_stage1.load_state_dict(ckpt[stage1_key])
+    old_stage2.load_state_dict(ckpt[stage2_key])
+    old_stage1.eval()
+    old_stage2.eval()
+
+    # --- new side: migrated config + remapped state dict, through the new build_models ---
+    new_models = net.build_models(model_config)
+    new_stage1, new_stage2 = new_models["stage1"], new_models["stage2"]
+    assert new_stage1 is not None and new_stage2 is not None
+
+    if routed:
+        print(
+            "  routed checkpoint: migrate_legacy_state_dict only handles the "
+            "monolithic trunk shape — verifying construction only, skipping "
+            "the bit-identical weight/output comparison."
+        )
+        print("PASS (construction only, routed checkpoint)")
+        return 0
+
+    remapped1, remapped2 = net.migrate_legacy_state_dict(ckpt[stage1_key], ckpt[stage2_key])
+    missing1, unexpected1 = new_stage1.load_state_dict(remapped1, strict=True)
+    missing2, unexpected2 = new_stage2.load_state_dict(remapped2, strict=True)
+    if missing1 or unexpected1 or missing2 or unexpected2:
+        print("FAIL: state dict mismatch after remap")
+        print(f"  stage1 missing={missing1} unexpected={unexpected1}")
+        print(f"  stage2 missing={missing2} unexpected={unexpected2}")
+        return 1
+    new_stage1.eval()
+    new_stage2.eval()
+
+    cond_cont, cond_cat, x1, x2, t, z1, z2 = _random_batch(model_config, args.batch, args.seed)
+
+    ok = True
+    with torch.no_grad():
+        if mode == "wgan":
+            old_out1 = old_stage1(z1, cond_cont, cond_cat)
+            new_out1 = new_stage1(z1, cond_cont, cond_cat)
+        else:
+            old_out1 = old_stage1(x1, t, cond_cont, cond_cat)
+            new_out1 = new_stage1(x1, cond_cont, cond_cat, t=t)
+        old_n_sec = old_stage1.predict_n_sec(cond_cont, cond_cat)
+        new_n_sec = new_stage1.predict_n_sec(cond_cont, cond_cat)
+        if mode == "wgan":
+            old_out2 = old_stage2(z2, cond_cont, cond_cat, old_out1)
+            new_out2 = new_stage2(z2, cond_cont, cond_cat, new_out1)
+        else:
+            old_out2 = old_stage2(x2, t, cond_cont, cond_cat, old_out1)
+            new_out2 = new_stage2(x2, cond_cont, cond_cat, new_out1, t=t)
+
+    for label, old_out, new_out in [
+        ("stage1 output", old_out1, new_out1),
+        ("n_sec logits", old_n_sec, new_n_sec),
+        ("stage2 output", old_out2, new_out2),
+    ]:
+        identical = torch.equal(old_out, new_out)
+        diff = _max_abs_diff(old_out, new_out)
+        status = "OK" if identical else "MISMATCH"
+        print(f"  {label}: {status} (max abs diff = {diff:.3e})")
+        ok = ok and identical
+
+    print("PASS" if ok else "FAIL")
+    return 0 if ok else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

giant/tools/create_root_files.py

@@ -32,7 +32,7 @@ from dataclasses import dataclass
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
-# Must match scripts/bump_dataset_version.py's GEN_RE.
+# Must match giant/tools/bump_dataset_version.py's GEN_RE.
 GEN_RE = re.compile(r"^gen\d+$")
 SHARD_RE = re.compile(r"^shard-(\d+)\.root$")
 
@@ -64,9 +64,7 @@ def parse_detector_spec(spec: str) -> tuple[str, str | None]:
     if ":" in spec:
         label, config = spec.split(":", 1)
         if not label or not config:
-            raise PlanError(
-                f"invalid --detector spec {spec!r}: expected NAME or NAME:CONFIG"
-            )
+            raise PlanError(f"invalid --detector spec {spec!r}: expected NAME or NAME:CONFIG")
         return label, config
     return spec, None
 
@@ -94,9 +92,7 @@ def plan_jobs(
         raise PlanError(f"--gen must look like 'genN', got {gen!r}")
     gen_dir = dataset_root / "raw" / kind / gen
     if not gen_dir.is_dir():
-        raise PlanError(
-            f"{gen_dir} doesn't exist — run bump_dataset_version.py bump-gen first"
-        )
+        raise PlanError(f"{gen_dir} doesn't exist — run bump_dataset_version.py bump-gen first")
 
     jobs = []
     for spec in detector_specs:
@@ -124,9 +120,7 @@ def job_seed(kind: str, gen: str, job: SimJob, energy_gev: float | None) -> int:
     return zlib.crc32(key.encode()) & 0x7FFFFFFF
 
 
-def build_cmd(
-    executable: Path, job: SimJob, events_per_file: int, energy_gev: float | None
-) -> list[str]:
+def build_cmd(executable: Path, job: SimJob, events_per_file: int, energy_gev: float | None) -> list[str]:
     """minicalosim executables take positional `[configName] nEvents [energy_GeV]`."""
     cmd = [str(executable)]
     if job.config:
@@ -147,10 +141,7 @@ def run_job(
     gen: str,
     tmp_root: Path,
 ) -> JobResult:
-    workdir = (
-        tmp_root
-        / f"{kind}-{gen}-{job.detector}-{job.shard_index:03d}-{uuid.uuid4().hex[:8]}"
-    )
+    workdir = tmp_root / f"{kind}-{gen}-{job.detector}-{job.shard_index:03d}-{uuid.uuid4().hex[:8]}"
     workdir.mkdir(parents=True)
 
     cmd = build_cmd(executable, job, events_per_file, energy_gev)
@@ -174,20 +165,12 @@ def run_job(
             job,
             False,
             None,
-            f"expected exactly one .root output in {workdir}, found {len(produced)}: "
-            f"{[p.name for p in produced]}",
+            f"expected exactly one .root output in {workdir}, found {len(produced)}: {[p.name for p in produced]}",
             result.stdout,
             result.stderr,
         )
 
-    dest = (
-        dataset_root
-        / "raw"
-        / kind
-        / gen
-        / job.detector
-        / f"shard-{job.shard_index:03d}.root"
-    )
+    dest = dataset_root / "raw" / kind / gen / job.detector / f"shard-{job.shard_index:03d}.root"
     if dest.exists():
         return JobResult(
             job,
@@ -279,14 +262,7 @@ def run_make_root(
     print(f"executable: {executable}")
     for job in planned_jobs:
         cmd = build_cmd(executable, job, events_per_file, energy_gev)
-        dest = (
-            dataset_root_path
-            / "raw"
-            / kind
-            / gen
-            / job.detector
-            / f"shard-{job.shard_index:03d}.root"
-        )
+        dest = dataset_root_path / "raw" / kind / gen / job.detector / f"shard-{job.shard_index:03d}.root"
         seed = job_seed(kind, gen, job, energy_gev)
         print(f"  MINICALOSIM_SEED={seed} {' '.join(cmd)}  ->  {dest}")
 

giant/tools/dwarf.py

@@ -1,6 +1,6 @@
 """dwarf — little helper to `giant`: dataset/tooling CLI for the geant_steps pipeline.
 
-Unifies the standalone scripts/*.py conversion, migration, versioning, and
+Unifies the standalone giant/tools/*.py conversion, migration, versioning, and
 simulation-fanout tools into one Typer app so there's a single command name
 (and `--help`) to remember instead of five differently-hyphenated ones.
 """
@@ -14,20 +14,20 @@ import typer
 from typing_extensions import Annotated
 
 from giant.config import Conditioning
-from scripts.bump_dataset_version import (
+from giant.tools.bump_dataset_version import (
     run_bump_gen,
     run_bump_schema,
     run_create_manifest,
     run_status,
     run_update_manifest,
 )
-from scripts.create_root_files import run_make_root
-from scripts.geometry_oracle import run_build_geometry_oracle
-from scripts.hparam_scan import DATA_DEFAULT, SCAN_DIR_DEFAULT, run_hparam_scan
-from scripts.migrate_geant_steps import run_migration
-from scripts.steps_to_parquet import convert_steps_to_parquet
-from scripts.steps_to_parquet_parallel import run_parallel_job
-from scripts.warm_setup_cache import run_warm_setup_cache
+from giant.tools.create_root_files import run_make_root
+from giant.tools.geometry_oracle import run_build_geometry_oracle
+from giant.tools.hparam_scan import DATA_DEFAULT, SCAN_DIR_DEFAULT, run_hparam_scan
+from giant.tools.migrate_geant_steps import run_migration
+from giant.tools.steps_to_parquet import convert_steps_to_parquet
+from giant.tools.steps_to_parquet_parallel import run_parallel_job
+from giant.tools.warm_setup_cache import run_warm_setup_cache
 
 app = typer.Typer(no_args_is_help=True)
 
@@ -80,8 +80,7 @@ def convert(
         typer.Option(
             "--output",
             "-o",
-            help="Output Parquet file (default: <input>.parquet). Only valid "
-            "with a single input file and --jobs 1.",
+            help="Output Parquet file (default: <input>.parquet). Only valid with a single input file and --jobs 1.",
         ),
     ] = None,
     batch_size: Annotated[
@@ -91,9 +90,7 @@ def convert(
             help="Uproot read batch size, e.g. '100 MB' or '500000' (rows)",
         ),
     ] = "100 MB",
-    tree: Annotated[
-        str, typer.Option("--tree", help="Tree name inside the ROOT file")
-    ] = "Steps",
+    tree: Annotated[str, typer.Option("--tree", help="Tree name inside the ROOT file")] = "Steps",
     compression: Annotated[
         Compression, typer.Option("--compression", help="Parquet compression codec")
     ] = Compression.snappy,
@@ -129,15 +126,11 @@ def convert(
         raise typer.Exit(1)
     _warn_if_exceeds_shared_quota(jobs, "--jobs")
 
-    compression_value = (
-        "uncompressed" if compression is Compression.none else compression.value
-    )
+    compression_value = "uncompressed" if compression is Compression.none else compression.value
 
     if jobs == 1:
         if output is not None and len(root_files) > 1:
-            typer.echo(
-                "error: --output can only be used with a single input file", err=True
-            )
+            typer.echo("error: --output can only be used with a single input file", err=True)
             raise typer.Exit(1)
         total_orphaned = 0
         for root_file in root_files:
@@ -150,10 +143,7 @@ def convert(
             )
             total_orphaned += n_orphaned
         if total_orphaned:
-            typer.echo(
-                f"\n{total_orphaned} orphaned child track(s) dropped across "
-                f"{len(root_files)} file(s)."
-            )
+            typer.echo(f"\n{total_orphaned} orphaned child track(s) dropped across {len(root_files)} file(s).")
         return
 
     if output is not None:
@@ -176,9 +166,7 @@ def convert(
 
 @app.command()
 def migrate(
-    root: Annotated[
-        Path, typer.Argument(help="Dataset root to migrate in place")
-    ] = _DATASET_ROOT_DEFAULT,
+    root: Annotated[Path, typer.Argument(help="Dataset root to migrate in place")] = _DATASET_ROOT_DEFAULT,
     execute: Annotated[
         bool,
         typer.Option(
@@ -190,8 +178,7 @@ def migrate(
         bool,
         typer.Option(
             "--copy",
-            help="Copy instead of move, leaving the originals in place "
-            "(e.g. if another process is still reading them)",
+            help="Copy instead of move, leaving the originals in place (e.g. if another process is still reading them)",
         ),
     ] = False,
 ) -> None:
@@ -202,12 +189,8 @@ def migrate(
 @app.command("bump-gen")
 def bump_gen(
     reason: Annotated[str, typer.Option("--reason", help="Why this gen exists")],
-    kind: Annotated[
-        str, typer.Option("--kind", help="steps | hits | ... (default: steps)")
-    ] = "steps",
-    by: Annotated[
-        Optional[str], typer.Option("--by", help="Attribution (default: git user.name)")
-    ] = None,
+    kind: Annotated[str, typer.Option("--kind", help="steps | hits | ... (default: steps)")] = "steps",
+    by: Annotated[Optional[str], typer.Option("--by", help="Attribution (default: git user.name)")] = None,
     date: Annotated[
         Optional[str],
         typer.Option("--date", help="Override date (default: today, ISO)"),
@@ -220,12 +203,8 @@ def bump_gen(
             help="Target gen tag (default: one past the current highest)",
         ),
     ] = None,
-    execute: Annotated[
-        bool, typer.Option("--execute", help="Apply (default: dry run)")
-    ] = False,
-    root: Annotated[
-        Path, typer.Option("--root", help="Dataset root")
-    ] = _DATASET_ROOT_DEFAULT,
+    execute: Annotated[bool, typer.Option("--execute", help="Apply (default: dry run)")] = False,
+    root: Annotated[Path, typer.Option("--root", help="Dataset root")] = _DATASET_ROOT_DEFAULT,
 ) -> None:
     """Cut a new raw generation."""
     run_bump_gen(
@@ -243,12 +222,8 @@ def bump_gen(
 def bump_schema(
     gen: Annotated[str, typer.Option("--gen", help="Existing gen tag, e.g. gen1")],
     reason: Annotated[str, typer.Option("--reason", help="Why this schema exists")],
-    kind: Annotated[
-        str, typer.Option("--kind", help="steps | hits | ... (default: steps)")
-    ] = "steps",
-    by: Annotated[
-        Optional[str], typer.Option("--by", help="Attribution (default: git user.name)")
-    ] = None,
+    kind: Annotated[str, typer.Option("--kind", help="steps | hits | ... (default: steps)")] = "steps",
+    by: Annotated[Optional[str], typer.Option("--by", help="Attribution (default: git user.name)")] = None,
     date: Annotated[
         Optional[str],
         typer.Option("--date", help="Override date (default: today, ISO)"),
@@ -261,12 +236,8 @@ def bump_schema(
             help="Target schema tag (default: one past the current highest)",
         ),
     ] = None,
-    execute: Annotated[
-        bool, typer.Option("--execute", help="Apply (default: dry run)")
-    ] = False,
-    root: Annotated[
-        Path, typer.Option("--root", help="Dataset root")
-    ] = _DATASET_ROOT_DEFAULT,
+    execute: Annotated[bool, typer.Option("--execute", help="Apply (default: dry run)")] = False,
+    root: Annotated[Path, typer.Option("--root", help="Dataset root")] = _DATASET_ROOT_DEFAULT,
 ) -> None:
     """Cut a new schema within a gen."""
     run_bump_schema(
@@ -283,9 +254,7 @@ def bump_schema(
 
 @app.command()
 def status(
-    root: Annotated[
-        Path, typer.Option("--root", help="Dataset root")
-    ] = _DATASET_ROOT_DEFAULT,
+    root: Annotated[Path, typer.Option("--root", help="Dataset root")] = _DATASET_ROOT_DEFAULT,
 ) -> None:
     """List existing gens/schemas per kind."""
     run_status(str(root))
@@ -293,9 +262,7 @@ def status(
 
 @app.command("update-manifest")
 def update_manifest(
-    manifests: Annotated[
-        list[Path], typer.Argument(help="One or more .manifest files to update")
-    ],
+    manifests: Annotated[list[Path], typer.Argument(help="One or more .manifest files to update")],
     schema: Annotated[
         Optional[str],
         typer.Option(
@@ -306,9 +273,7 @@ def update_manifest(
     ] = None,
     gen: Annotated[
         Optional[str],
-        typer.Option(
-            "--gen", metavar="genN", help="Target gen tag (default: keep existing gen)"
-        ),
+        typer.Option("--gen", metavar="genN", help="Target gen tag (default: keep existing gen)"),
     ] = None,
     execute: Annotated[
         bool,
@@ -316,9 +281,7 @@ def update_manifest(
     ] = False,
 ) -> None:
     """Repoint manifest(s) to a new gen and/or schema, verifying all target files exist."""
-    run_update_manifest(
-        [str(m) for m in manifests], schema=schema, execute=execute, gen=gen
-    )
+    run_update_manifest([str(m) for m in manifests], schema=schema, execute=execute, gen=gen)
 
 
 @app.command("create-manifest")
@@ -333,22 +296,15 @@ def create_manifest(
         typer.Option(
             "--pool",
             metavar="DETECTOR",
-            help="Detector name; combined with --type and --root to form "
-            "<root>/pools/<detector>/<type>.manifest",
+            help="Detector name; combined with --type and --root to form <root>/pools/<detector>/<type>.manifest",
         ),
     ] = None,
     type_: Annotated[
         Optional[PoolType],
-        typer.Option(
-            "--type", help="Pool type — full, holdout, or dev (required with --pool)"
-        ),
+        typer.Option("--type", help="Pool type — full, holdout, or dev (required with --pool)"),
     ] = None,
-    root: Annotated[
-        Path, typer.Option("--root", help="Dataset root (used with --pool)")
-    ] = _DATASET_ROOT_DEFAULT,
-    execute: Annotated[
-        bool, typer.Option("--execute", help="Write the manifest (default: dry run)")
-    ] = False,
+    root: Annotated[Path, typer.Option("--root", help="Dataset root (used with --pool)")] = _DATASET_ROOT_DEFAULT,
+    execute: Annotated[bool, typer.Option("--execute", help="Write the manifest (default: dry run)")] = False,
     force: Annotated[
         bool,
         typer.Option("--force", help="Overwrite the manifest if it already exists"),
@@ -368,9 +324,7 @@ def create_manifest(
 
 @app.command("make-root")
 def make_root(
-    executable: Annotated[
-        Path, typer.Option("--executable", help="Built minicalosim run_* executable")
-    ],
+    executable: Annotated[Path, typer.Option("--executable", help="Built minicalosim run_* executable")],
     detector: Annotated[
         list[str],
         typer.Option(
@@ -382,15 +336,9 @@ def make_root(
             "Repeatable.",
         ),
     ],
-    num_files: Annotated[
-        int, typer.Option("--num-files", help="New shards to create per detector")
-    ],
-    events_per_file: Annotated[
-        int, typer.Option("--events-per-file", help="nEvents passed to the executable")
-    ],
-    gen: Annotated[
-        str, typer.Option("--gen", help="Existing gen tag under raw/<kind>/, e.g. gen1")
-    ],
+    num_files: Annotated[int, typer.Option("--num-files", help="New shards to create per detector")],
+    events_per_file: Annotated[int, typer.Option("--events-per-file", help="nEvents passed to the executable")],
+    gen: Annotated[str, typer.Option("--gen", help="Existing gen tag under raw/<kind>/, e.g. gen1")],
     energy_gev: Annotated[
         float | None,
         typer.Option(
@@ -401,20 +349,12 @@ def make_root(
             "to name the dataset accordingly.",
         ),
     ] = None,
-    kind: Annotated[
-        str, typer.Option("--kind", help="steps | hits | ... (default: steps)")
-    ] = "steps",
-    dataset_root: Annotated[
-        Path, typer.Option("--dataset-root", help="Dataset root")
-    ] = _DATASET_ROOT_DEFAULT,
-    jobs: Annotated[
-        int, typer.Option("--jobs", "-j", help="Parallel simulation runs (default: 4)")
-    ] = 4,
+    kind: Annotated[str, typer.Option("--kind", help="steps | hits | ... (default: steps)")] = "steps",
+    dataset_root: Annotated[Path, typer.Option("--dataset-root", help="Dataset root")] = _DATASET_ROOT_DEFAULT,
+    jobs: Annotated[int, typer.Option("--jobs", "-j", help="Parallel simulation runs (default: 4)")] = 4,
     execute: Annotated[
         bool,
-        typer.Option(
-            "--execute", help="Actually run jobs (default: dry run / print plan)"
-        ),
+        typer.Option("--execute", help="Actually run jobs (default: dry run / print plan)"),
     ] = False,
 ) -> None:
     """Generate new ROOT shards via a minicalosim executable."""
@@ -441,9 +381,7 @@ class OracleMethod(str, Enum):
 
 @app.command("build-geometry-oracle")
 def build_geometry_oracle(
-    data: Annotated[
-        Path, typer.Argument(help="Steps parquet file or directory of steps files")
-    ],
+    data: Annotated[Path, typer.Argument(help="Steps parquet file or directory of steps files")],
     out: Annotated[Path, typer.Option("--out", "-o", help="Output oracle .pkl path")],
     method: Annotated[
         OracleMethod,
@@ -456,9 +394,7 @@ def build_geometry_oracle(
             ),
         ),
     ] = OracleMethod.slab,
-    k: Annotated[
-        int, typer.Option("--k", help="Neighbours for the knn classifier")
-    ] = 1,
+    k: Annotated[int, typer.Option("--k", help="Neighbours for the knn classifier")] = 1,
     subsample: Annotated[
         int,
         typer.Option("--subsample", help="Max reference points sampled from the data"),
@@ -504,9 +440,7 @@ def build_geometry_oracle(
 def warm_cache(
     data: Annotated[
         Path,
-        typer.Argument(
-            help="Parquet file, directory, or .manifest — same as `giant train`'s"
-        ),
+        typer.Argument(help="Parquet file, directory, or .manifest — same as `giant train`'s"),
     ],
     val_fraction: Annotated[
         float,
@@ -518,49 +452,52 @@ def warm_cache(
     ] = 0.1,
     seed: Annotated[
         int,
-        typer.Option(
-            "--seed", "-s", help="Must match the `giant train` run(s) to warm for"
-        ),
+        typer.Option("--seed", "-s", help="Must match the `giant train` run(s) to warm for"),
     ] = 0,
-    conditioning: Annotated[
+    particle_conditioning: Annotated[
         Conditioning,
         typer.Option(
-            "--conditioning", help="Must match the `giant train` run(s) to warm for"
+            "--particle-conditioning",
+            help="Must match the `giant train` run(s)' conditioning.particle.type to warm for",
+        ),
+    ] = Conditioning.physical,
+    material_conditioning: Annotated[
+        Conditioning,
+        typer.Option(
+            "--material-conditioning",
+            help="Must match the `giant train` run(s)' conditioning.material.type "
+            "to warm for — independent of --particle-conditioning "
+            "(the two axes may differ)",
         ),
     ] = Conditioning.physical,
     router: Annotated[
         bool,
         typer.Option(
             "--router/--no-router",
-            help="Warm the process vocabulary too (only takes effect with "
-            "--router-type process)",
+            help="Warm the process vocabulary too (only takes effect with --router-type process)",
         ),
     ] = False,
-    router_type: Annotated[
-        str, typer.Option("--router-type", help="Router implementation name")
-    ] = "energy",
-    n_experts: Annotated[
-        int, typer.Option("--n-experts", help="Number of routed experts")
-    ] = 4,
+    router_type: Annotated[str, typer.Option("--router-type", help="Router implementation name")] = "energy",
+    n_experts: Annotated[int, typer.Option("--n-experts", help="Number of routed experts")] = 4,
     rebuild: Annotated[
         bool,
-        typer.Option(
-            "--rebuild", help="Ignore any existing sidecar and recompute every section"
-        ),
+        typer.Option("--rebuild", help="Ignore any existing sidecar and recompute every section"),
     ] = False,
 ) -> None:
     """Precompute `giant train`'s setup-stage sidecar for `data` ahead of time.
 
     Warms the vocab maps, event-id split index, and the normalizer entry for
-    the given --val-fraction/--seed/--conditioning, so a later `giant train`
-    run (or a `dwarf hparam-scan` sweep, which shares one such entry across
-    every run) skips straight to training. See giant/data/setup_cache.py.
+    the given --val-fraction/--seed/--particle-conditioning/
+    --material-conditioning, so a later `giant train` run (or a `dwarf
+    hparam-scan` sweep, which shares one such entry across every run) skips
+    straight to training. See giant/data/setup_cache.py.
     """
     run_warm_setup_cache(
         data=str(data),
         val_fraction=val_fraction,
         seed=seed,
-        conditioning=conditioning.value,
+        particle_conditioning=particle_conditioning.value,
+        material_conditioning=material_conditioning.value,
         router_enabled=router,
         router_type=router_type,
         n_experts=n_experts,

giant/tools/geometry_oracle.py

@@ -38,9 +38,7 @@ def run_build_geometry_oracle(
         n_bins=n_bins,
     )
 
-    print(
-        f"method: {method}  reference points: {oracle.metadata['n_reference_points']:,}"
-    )
+    print(f"method: {method}  reference points: {oracle.metadata['n_reference_points']:,}")
     print("classes (material, layer_id):")
     for material, layer_id in oracle.classes:
         print(f"  {material:<12} layer_id={layer_id}")

giant/tools/hparam_scan.py

@@ -68,8 +68,8 @@ def final_metrics(metrics_path: Path) -> tuple[int, float, float]:
     with open(metrics_path, newline="") as f:
         rows = list(csv.DictReader(f))
     epochs_completed = int(rows[-1]["epoch"])
-    final_val_loss = float(rows[-1]["val_loss"])
-    best_val_loss = min(float(r["val_loss"]) for r in rows)
+    final_val_loss = float(rows[-1]["val/loss"])
+    best_val_loss = min(float(r["val/loss"]) for r in rows)
     return epochs_completed, final_val_loss, best_val_loss
 
 
@@ -154,9 +154,7 @@ def run_hparam_scan(
         wall_time_s = time.monotonic() - start
 
         if metrics_path.exists():
-            epochs_completed, final_val_loss, best_val_loss = final_metrics(
-                metrics_path
-            )
+            epochs_completed, final_val_loss, best_val_loss = final_metrics(metrics_path)
             append_summary(
                 summary_path,
                 {
@@ -171,11 +169,6 @@ def run_hparam_scan(
                     "wall_time_s": round(wall_time_s, 1),
                 },
             )
-            print(
-                f"[{i}/{len(runs)}] {name} — val_loss {final_val_loss:.4f} "
-                f"({wall_time_s:.1f}s)"
-            )
+            print(f"[{i}/{len(runs)}] {name} — val_loss {final_val_loss:.4f} ({wall_time_s:.1f}s)")
         else:
-            print(
-                f"[{i}/{len(runs)}] {name} — no metrics.csv produced, check train.log"
-            )
+            print(f"[{i}/{len(runs)}] {name} — no metrics.csv produced, check train.log")

giant/tools/migrate_geant_steps.py

@@ -50,12 +50,8 @@ PREDICTED_RE = re.compile(
     r"^(?P<detector>[a-z0-9]+(?:_[a-z0-9]+)*)_10k_(?P<shard>\d+)"
     r"_predicted(?P<local>_local)?\.parquet$"
 )
-SHARD_RE = re.compile(
-    r"^(?P<detector>[a-z0-9]+(?:_[a-z0-9]+)*)_10k_(?P<shard>\d+)\.(?P<ext>root|parquet)$"
-)
-LEGACY_PREDICTED_RE = re.compile(
-    r"^pbwo4_10000events_hits_predicted(?P<local>_local)?\.parquet$"
-)
+SHARD_RE = re.compile(r"^(?P<detector>[a-z0-9]+(?:_[a-z0-9]+)*)_10k_(?P<shard>\d+)\.(?P<ext>root|parquet)$")
+LEGACY_PREDICTED_RE = re.compile(r"^pbwo4_10000events_hits_predicted(?P<local>_local)?\.parquet$")
 LEGACY_RE = re.compile(r"^pbwo4_10000events_hits\.(?P<ext>root|parquet)$")
 
 
@@ -110,24 +106,9 @@ def plan_moves(src_root: Path) -> tuple[list[tuple[Path, Path]], list[Path]]:
         if m:
             detector, shard, ext = m["detector"], int(m["shard"]), m["ext"]
             if ext == "root":
-                dst = (
-                    src_root
-                    / "raw"
-                    / "steps"
-                    / GEN
-                    / detector
-                    / f"shard-{shard:03d}.root"
-                )
+                dst = src_root / "raw" / "steps" / GEN / detector / f"shard-{shard:03d}.root"
             else:
-                dst = (
-                    src_root
-                    / "processed"
-                    / "steps"
-                    / GEN
-                    / SCHEMA
-                    / detector
-                    / f"shard-{shard:03d}.parquet"
-                )
+                dst = src_root / "processed" / "steps" / GEN / SCHEMA / detector / f"shard-{shard:03d}.parquet"
             moves.append((path, dst))
             continue
 
@@ -135,19 +116,9 @@ def plan_moves(src_root: Path) -> tuple[list[tuple[Path, Path]], list[Path]]:
         if m:
             ext = m["ext"]
             if ext == "root":
-                dst = (
-                    src_root / "raw" / "hits" / LEGACY_GEN / "pbwo4" / "shard-000.root"
-                )
+                dst = src_root / "raw" / "hits" / LEGACY_GEN / "pbwo4" / "shard-000.root"
             else:
-                dst = (
-                    src_root
-                    / "processed"
-                    / "hits"
-                    / LEGACY_GEN
-                    / LEGACY_SCHEMA
-                    / "pbwo4"
-                    / "shard-000.parquet"
-                )
+                dst = src_root / "processed" / "hits" / LEGACY_GEN / LEGACY_SCHEMA / "pbwo4" / "shard-000.parquet"
             moves.append((path, dst))
             continue
 
@@ -165,20 +136,9 @@ def plan_manifests(src_root: Path) -> dict[Path, list[str]]:
         for pool, shards in rules.items():
             manifest_path = manifest_dir / f"{pool}{MANIFEST_SUFFIX}"
             for shard in shards:
-                dst = (
-                    src_root
-                    / "processed"
-                    / "steps"
-                    / GEN
-                    / SCHEMA
-                    / detector
-                    / f"shard-{shard:03d}.parquet"
-                )
+                dst = src_root / "processed" / "steps" / GEN / SCHEMA / detector / f"shard-{shard:03d}.parquet"
                 manifests[manifest_path].append((shard, dst))
-    return {
-        k: [os.path.relpath(dst, start=k.parent) for _, dst in sorted(v)]
-        for k, v in manifests.items()
-    }
+    return {k: [os.path.relpath(dst, start=k.parent) for _, dst in sorted(v)] for k, v in manifests.items()}
 
 
 def run_migration(root: str, execute: bool, copy: bool) -> None:

giant/tools/profile_analysis_costs.py

@@ -13,7 +13,7 @@ real checkpoint) they short-circuit almost instantly and are excluded here —
 see `runtime_estimate.py`'s `_ROUTER_FIXED_S` for how those are handled
 instead.
 
-Usage: ``uv run python scripts/profile_analysis_costs.py``
+Usage: ``uv run python giant/tools/profile_analysis_costs.py``
 """
 
 from __future__ import annotations
@@ -94,9 +94,7 @@ def _make_rollout(n: int, n_events: int, seed: int) -> pl.DataFrame:
             "post_dx": post_dir[:, 0],
             "post_dy": post_dir[:, 1],
             "post_dz": post_dir[:, 2],
-            "edep": np.where(
-                is_synthetic, np.where(reasons == "escaped", 0.0, pre_E), edep
-            ),
+            "edep": np.where(is_synthetic, np.where(reasons == "escaped", 0.0, pre_E), edep),
             "step_length": np.where(is_synthetic, 0.0, step_length),
             "material": rng.choice(_MATERIALS, size=n),
             "layer_id": rng.integers(0, 30, size=n),
@@ -163,9 +161,7 @@ def _make_reference(n: int, n_events: int, seed: int) -> pl.DataFrame:
     )
 
 
-def _time(
-    spec_id: str, rollout: Path, reference: Path, shared: Path, out: Path
-) -> float:
+def _time(spec_id: str, rollout: Path, reference: Path, shared: Path, out: Path) -> float:
     t0 = time.perf_counter()
     compute_reduced(
         spec_id,

giant/tools/steps_to_parquet_parallel.py

@@ -2,11 +2,11 @@
 
 A single `dwarf convert` call converts a list of files one at a time; this
 module runs up to --jobs conversions concurrently, each as its own `dwarf
-convert` subprocess (invoked via `python -m scripts.dwarf`, so it picks up
+convert` subprocess (invoked via `python -m giant.tools.dwarf`, so it picks up
 the active venv/uv environment automatically).
 
 Inputs must live under <dataset-root>/raw/<kind>/<gen>/<detector>/<file>.root
-(see scripts/migrate_geant_steps.py) — each is written to the matching
+(see giant/tools/migrate_geant_steps.py) — each is written to the matching
 processed/<kind>/<gen>/<schema>/<detector>/<file>.parquet, where <schema>
 defaults to the highest schemaN already under processed/<kind>/<gen>/ (pass
 --schema to pick a specific one, e.g. one just created by `dwarf bump-schema`).
@@ -22,7 +22,7 @@ import sys
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from pathlib import Path
 
-# Must match scripts/bump_dataset_version.py's GEN_RE / SCHEMA_RE.
+# Must match giant/tools/bump_dataset_version.py's GEN_RE / SCHEMA_RE.
 GEN_RE = re.compile(r"^gen\d+$")
 SCHEMA_RE = re.compile(r"^schema(\d+)$")
 
@@ -49,9 +49,7 @@ def latest_schema_tag(processed_gen_dir: Path) -> str | None:
     return best_tag
 
 
-def resolve_destination(
-    root_file: Path, dataset_root: Path, schema_override: str | None
-) -> Path:
+def resolve_destination(root_file: Path, dataset_root: Path, schema_override: str | None) -> Path:
     """Map raw/<kind>/<gen>/<detector>/<file>.root (relative to *dataset_root*)
     to processed/<kind>/<gen>/<schema>/<detector>/<file>.parquet.
 
@@ -66,12 +64,7 @@ def resolve_destination(
         raise DestinationError(f"{root_file} is not under dataset root {dataset_root}")
 
     parts = rel.parts
-    if (
-        len(parts) != 5
-        or parts[0] != "raw"
-        or not GEN_RE.match(parts[2])
-        or not parts[4].endswith(".root")
-    ):
+    if len(parts) != 5 or parts[0] != "raw" or not GEN_RE.match(parts[2]) or not parts[4].endswith(".root"):
         raise DestinationError(
             f"{root_file} does not match raw/<kind>/<gen>/<detector>/<file>.root "
             f"under {dataset_root} (got relative path: {rel})"
@@ -89,7 +82,7 @@ def resolve_destination(
     return processed_gen_dir / schema_tag / detector / f"{shard_stem}.parquet"
 
 
-_DWARF_CONVERT_CMD = [sys.executable, "-m", "scripts.dwarf", "convert"]
+_DWARF_CONVERT_CMD = [sys.executable, "-m", "giant.tools.dwarf", "convert"]
 
 
 def _convert_one(
@@ -134,7 +127,7 @@ def run_parallel(
     written next to the input .root).
 
     *cmd_prefix* overrides the subprocess command run per file (defaults to
-    `python -m scripts.dwarf convert`) — used by tests to substitute a fake
+    `python -m giant.tools.dwarf convert`) — used by tests to substitute a fake
     conversion script.
 
     Returns one (root_file, returncode, stdout, stderr) tuple per file, in
@@ -216,13 +209,7 @@ def run_parallel_job(
             print(f"  {root_file}", file=sys.stderr)
         raise SystemExit(1)
 
-    total_orphaned = sum(
-        int(m.group(1))
-        for _, _, stdout, _ in results
-        for m in _ORPHAN_RE.finditer(stdout)
-    )
+    total_orphaned = sum(int(m.group(1)) for _, _, stdout, _ in results for m in _ORPHAN_RE.finditer(stdout))
     if total_orphaned:
-        print(
-            f"\n{total_orphaned} orphaned child track(s) dropped across {len(results)} file(s)."
-        )
+        print(f"\n{total_orphaned} orphaned child track(s) dropped across {len(results)} file(s).")
     print(f"\nAll {len(results)} conversion(s) completed.")

giant/tools/warm_setup_cache.py

@@ -9,6 +9,8 @@ for the sidecar itself.
 
 from pathlib import Path
 
+from giant import config as gconfig
+from giant.constants import K_MAX
 from giant.pipeline import run_setup_stage
 
 
@@ -16,7 +18,8 @@ def run_warm_setup_cache(
     data: str,
     val_fraction: float = 0.1,
     seed: int = 0,
-    conditioning: str = "physical",
+    particle_conditioning: str = "physical",
+    material_conditioning: str = "physical",
     router_enabled: bool = False,
     router_type: str = "energy",
     n_experts: int = 4,
@@ -25,9 +28,11 @@ def run_warm_setup_cache(
 ) -> None:
     """Populate (or refresh) the setup cache sidecar for `data`.
 
-    `val_fraction`/`seed`/`conditioning` select the normalizer cache entry
+    `val_fraction`/`seed`/`particle_conditioning`/`material_conditioning`
+    select the normalizer cache entry
     (`giant.data.setup_cache.normalizer_key`) — pass the same values a later
-    `giant train` invocation will use so it hits this warmed entry.
+    `giant train` invocation will use so it hits this warmed entry. The two
+    conditioning axes are independent and may differ.
     `router_enabled`/`router_type`/`n_experts` only matter for
     `router_type == "process"` (warms that `n_experts`'s process map); the
     energy-router quantile summary is always collected regardless, so a
@@ -39,12 +44,30 @@ def run_warm_setup_cache(
         "type": router_type,
         "n_experts": n_experts,
     }
+    # Merged against DEFAULT_CONFIG (not a hand-rolled partial dict) so
+    # run_setup_stage always sees every key it might read (e.g.
+    # conditioning.particle.emb_dim, stage2_model.particle_type.target) at
+    # its real default, not silently missing/None — see issues.md Issue 1.
+    # This CLI only ever configures one router (matching today's single
+    # --router-type flag), so it's placed on stage1_model; stage2_model's
+    # stays disabled.
+    cfg = gconfig.merge_cli_overrides(
+        gconfig.DEFAULT_CONFIG,
+        None,
+        {
+            "conditioning": {
+                "particle": {"type": particle_conditioning},
+                "material": {"type": material_conditioning},
+            },
+            "stage1_model": {"router": router_cfg},
+            "stage2_model": {"router": {"enabled": False}, "k_max": K_MAX},
+        },
+    )
     run_setup_stage(
         Path(data),
         val_fraction=val_fraction,
         seed=seed,
-        conditioning=conditioning,
-        router_cfg=router_cfg,
+        cfg=cfg,
         cache_setup=True,
         rebuild_setup_cache=rebuild,
         echo=echo,

giant/training/__init__.py

@@ -0,0 +1,30 @@
+"""Training: per-stage trainers, metric collection, checkpointing, the loop.
+
+Split out of the former single-module `giant/train.py`. The public surface is
+`train` (the entry point `giant.pipeline` calls) plus the trainer/spec types
+that tests and tooling construct directly.
+"""
+
+from giant.training.checkpoint import build_checkpoint, load_checkpoint
+from giant.training.metrics import MetricsCollector, MetricSpec
+from giant.training.loop import train
+from giant.training.trainers import (
+    FlowDDPMStageTrainer,
+    StageSpec,
+    StageTrainer,
+    WGANStageTrainer,
+    build_stage_trainers,
+)
+
+__all__ = [
+    "FlowDDPMStageTrainer",
+    "MetricSpec",
+    "MetricsCollector",
+    "StageSpec",
+    "StageTrainer",
+    "WGANStageTrainer",
+    "build_checkpoint",
+    "build_stage_trainers",
+    "load_checkpoint",
+    "train",
+]

giant/training/checkpoint.py

@@ -0,0 +1,68 @@
+"""Checkpoint assembly and restore.
+
+The on-disk layout is unchanged from v0.2/v0.3.0 and is read by
+`giant/cli.py`, `giant/rollout.py`, `giant/sample.py` and
+`giant/analysis/router_gating.py` — stage 1's weights live under `model`,
+stage 2's under `sec_decoder`, with `_ema`/`critic`/`sec_critic` companions
+and per-stage `optimizer_<stage>` / `optimizer_d_<stage>` / `lr_sched_<stage>`
+entries.
+"""
+
+from giant.training.trainers import StageTrainer
+
+#: Stage name -> the checkpoint key its weights live under. Historical: stage
+#: 1 predates the two-stage split, so it kept the bare "model" key.
+_STAGE_KEY = {"stage1": "model", "stage2": "sec_decoder"}
+_CRITIC_KEY = {"stage1": "critic", "stage2": "sec_critic"}
+
+
+def build_checkpoint(
+    trainers: dict[str, StageTrainer],
+    epoch: int,
+    global_step: int,
+    best_val_loss: float,
+    extras: dict,
+) -> dict:
+    """`extras` carries the dataset-level sidecars (normalizer, vocab maps,
+    model_config) that `train()` receives as arguments; `None` values are
+    omitted so an absent sidecar leaves no key behind."""
+    ckpt: dict = {
+        "epoch": epoch,
+        "best_val_loss": best_val_loss,
+        "global_step": global_step,
+    }
+    for name, trainer in trainers.items():
+        sd = trainer.state_dict()
+        key = _STAGE_KEY[name]
+        ckpt[key] = sd["model"]
+        if "model_ema" in sd:
+            ckpt[f"{key}_ema"] = sd["model_ema"]
+        if "critic" in sd:
+            ckpt[_CRITIC_KEY[name]] = sd["critic"]
+            ckpt[f"optimizer_d_{name}"] = sd["optimizer_d"]
+        ckpt[f"optimizer_{name}"] = sd["optimizer"]
+        ckpt[f"lr_sched_{name}"] = sd["lr_sched"]
+    ckpt.update({k: v for k, v in extras.items() if v is not None})
+    return ckpt
+
+
+def load_checkpoint(trainers: dict[str, StageTrainer], ckpt: dict, lr: float) -> None:
+    """Restore every active stage, then hand `lr`'s authority back to the
+    config — `load_state_dict` would otherwise leave the checkpoint's own
+    base LR in place, silently ignoring `--lr` on resume."""
+    for name, trainer in trainers.items():
+        key = _STAGE_KEY[name]
+        sd = {
+            "model": ckpt[key],
+            "optimizer": ckpt[f"optimizer_{name}"],
+            "lr_sched": ckpt[f"lr_sched_{name}"],
+        }
+        ema_key = f"{key}_ema"
+        if ema_key in ckpt:
+            sd["model_ema"] = ckpt[ema_key]
+        crit_key = _CRITIC_KEY[name]
+        if crit_key in ckpt:
+            sd["critic"] = ckpt[crit_key]
+            sd["optimizer_d"] = ckpt[f"optimizer_d_{name}"]
+        trainer.load_state_dict(sd)
+        trainer.resume_lr(lr)

giant/training/loop.py

@@ -0,0 +1,297 @@
+"""The training loop.
+
+`train()` owns the epoch structure and nothing else: the per-stage step is
+`giant.training.trainers`' job, every number reported is
+`giant.training.metrics`' job, and the on-disk checkpoint is
+`giant.training.checkpoint`'s.
+"""
+
+import os
+import signal
+import time
+from pathlib import Path
+from types import FrameType
+from typing import Callable
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from giant.data.loader import TopNMap
+from giant.data.setup_cache import topnmap_to_json
+from giant.training.checkpoint import build_checkpoint, load_checkpoint
+from giant.training.metrics import MetricsCollector
+from giant.training.trainers import (
+    FlowDDPMStageTrainer,
+    StageTrainer,
+    build_stage_trainers,
+)
+from giant.validate import validate_marginals
+
+_CATCHABLE_SIGNALS = (signal.SIGINT, signal.SIGTERM)
+
+
+class _GracefulShutdown:
+    """Turns SIGINT/SIGTERM into a flag check instead of an immediate crash.
+
+    A second signal while already shutting down restores the default
+    handler and re-sends the signal, so an unresponsive run can still be
+    force-killed.
+    """
+
+    def __init__(self) -> None:
+        self.requested = False
+        self._previous: dict[
+            int,
+            Callable[[int, FrameType | None], object] | signal.Handlers | int | None,
+        ] = {}
+
+    def __enter__(self) -> "_GracefulShutdown":
+        for sig in _CATCHABLE_SIGNALS:
+            self._previous[sig] = signal.getsignal(sig)
+            signal.signal(sig, self._handle)
+        return self
+
+    def __exit__(self, *exc_info) -> None:
+        for sig, handler in self._previous.items():
+            signal.signal(sig, handler)
+
+    def _handle(self, signum: int, frame) -> None:
+        if self.requested:
+            signal.signal(signum, self._previous[signum])
+            os.kill(os.getpid(), signum)
+            return
+        self.requested = True
+        print(
+            f"\nreceived {signal.Signals(signum).name} — finishing the current "
+            "batch, then saving a checkpoint and exiting (send again to force-quit)"
+        )
+
+
+def _try_validate_marginals(trainer: StageTrainer, val_loader, device, **kwargs):
+    """Runs `validate_marginals` on `trainer`'s sampling model (EMA model if
+    present, else the raw model). `validate_marginals` itself dispatches
+    through `giant.sample.sample_stage1`/`sample_stage2`/`resolve_n_sec`, so
+    this is generator- and one-shot-vs-autoregressive-agnostic."""
+    model = trainer.sampling_model()
+    return validate_marginals(model, val_loader, device=device, **kwargs)
+
+
+def _marginal_kl(trainers: dict[str, StageTrainer], val_loader, device, **kwargs) -> float:
+    """Mean marginal KL over the stage-1 sampling chain, or NaN when stage 1
+    is inactive or `validate_marginals` declined to produce a result."""
+    stage1 = trainers.get("stage1")
+    if stage1 is None:
+        return float("nan")
+    result = _try_validate_marginals(
+        stage1,
+        val_loader,
+        device,
+        sec_decoder=trainers["stage2"].sampling_model() if "stage2" in trainers else None,
+        **kwargs,
+    )
+    if result is None:
+        return float("nan")
+    return float(np.mean(result["kl_divergence"]))
+
+
+def train(
+    cfg: dict,
+    models: dict[str, torch.nn.Module | None],
+    critics: dict[str, torch.nn.Module | None],
+    train_loader: DataLoader,
+    val_loader: DataLoader,
+    device: torch.device,
+    out_dir: str | Path,
+    normalizer_dict: dict | None = None,
+    pdg_map: dict | None = None,
+    mat_map: dict | None = None,
+    proc_map: dict | None = None,
+    pdg_topn_map: TopNMap | None = None,
+    sec_type_topn_map: TopNMap | None = None,
+    mat_topn_map: TopNMap | None = None,
+    model_config: dict | None = None,
+    resume_path: str | Path | None = None,
+    total_train_batches: int = 0,
+    use_wandb: bool = False,
+    wandb_project: str = "giant",
+    wandb_run_name: str = "",
+    wandb_log_every: int = 50,
+) -> None:
+    """Train whichever of stage1/stage2 are active, each through its own
+    `StageTrainer`. `models`/`critics` are the dicts
+    `giant.model.network.build_models`/`build_critics` return — a `None`
+    entry means that stage is `active = false`.
+    """
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    t = cfg["train"]
+    epochs = t["epochs"]
+    validate_every = t.get("validate_every", 0)
+    validate_steps = t.get("validate_steps", 10)
+    max_val_batches = t.get("max_val_batches", 0)
+
+    trainers = build_stage_trainers(cfg, models, critics, device, total_train_batches)
+    if not trainers:
+        raise ValueError("no active stage — stage1_model.active and stage2_model.active are both false")
+    has_adversarial = any(not tr.supports_val_loss for tr in trainers.values())
+
+    checkpoint_extras = {
+        "normalizer": normalizer_dict,
+        "pdg_map": pdg_map,
+        "mat_map": mat_map,
+        "proc_map": proc_map,
+        "pdg_topn_map": topnmap_to_json(pdg_topn_map) if pdg_topn_map is not None else None,
+        "sec_type_topn_map": topnmap_to_json(sec_type_topn_map) if sec_type_topn_map is not None else None,
+        "mat_topn_map": topnmap_to_json(mat_topn_map) if mat_topn_map is not None else None,
+        "model_config": model_config,
+    }
+
+    start_epoch = 1
+    best_val_loss = float("inf")
+    global_step = 0
+    if resume_path is not None:
+        ckpt = torch.load(resume_path, map_location=device, weights_only=False)
+        load_checkpoint(trainers, ckpt, t["lr"])
+        start_epoch = ckpt.get("epoch", 0) + 1
+        best_val_loss = ckpt.get("best_val_loss", float("inf"))
+        global_step = ckpt.get("global_step", 0)
+        if start_epoch > epochs:
+            print(f"checkpoint already completed epoch {start_epoch - 1} (>= --epochs {epochs}) — nothing to train")
+            return
+
+    collector = MetricsCollector.create(
+        trainers,
+        out_dir,
+        cfg,
+        model_config,
+        resume=resume_path is not None,
+        use_wandb=use_wandb,
+        wandb_project=wandb_project,
+        wandb_run_name=wandb_run_name,
+        wandb_log_every=wandb_log_every,
+    )
+
+    epoch_w = len(str(epochs))
+    last_completed_epoch = start_epoch - 1
+    with _GracefulShutdown() as shutdown:
+        for epoch in range(start_epoch, epochs + 1):
+            epoch_start = time.monotonic()
+            if device.type == "cuda":
+                torch.cuda.reset_peak_memory_stats(device)
+            collector.start_epoch(epoch)
+            for trainer in trainers.values():
+                trainer.train_mode()
+
+            bar = tqdm(
+                train_loader,
+                desc=f"  epoch {epoch:{epoch_w}d}/{epochs}",
+                total=total_train_batches or None,
+                leave=False,
+                unit="batch",
+                dynamic_ncols=True,
+            )
+            for batch in bar:
+                B = batch[0].size(0)
+                collector.add_train_batch(
+                    {name: trainer.step(batch, device, global_step) for name, trainer in trainers.items()},
+                    B,
+                )
+                bar.set_postfix_str(collector.postfix(), refresh=False)
+                global_step += 1
+                collector.log_batch(global_step, batch, device)
+                if shutdown.requested:
+                    break
+            bar.close()
+
+            if shutdown.requested:
+                ckpt = build_checkpoint(trainers, epoch - 1, global_step, best_val_loss, checkpoint_extras)
+                torch.save(ckpt, out_dir / "last.pt")
+                last_completed_epoch = epoch - 1
+                print(
+                    f"saved in-progress weights from partway through epoch "
+                    f"{epoch} to {out_dir / 'last.pt'} "
+                    f"(resume will restart epoch {epoch})"
+                )
+                break
+
+            for trainer in trainers.values():
+                trainer.eval_mode()
+
+            # --- per-stage validation ---
+            scored = {name: tr for name, tr in trainers.items() if tr.supports_val_loss}
+            if scored:
+                with torch.no_grad():
+                    for val_batch_idx, batch in enumerate(val_loader):
+                        if max_val_batches > 0 and val_batch_idx >= max_val_batches:
+                            break
+                        B = batch[0].size(0)
+                        collector.add_val_batch(
+                            {name: tr.val_loss(batch, device) for name, tr in scored.items()},
+                            B,
+                        )
+                        collector.observe_routers(batch[0].to(device), batch[1].to(device), B)
+
+            # An adversarial stage has no averageable validation loss, so it
+            # needs the marginal-KL signal every epoch to pick a best
+            # checkpoint at all; a purely non-adversarial run only pays for
+            # it every `validate_every` epochs.
+            marginal_kl = float("nan")
+            if has_adversarial:
+                marginal_kl = _marginal_kl(trainers, val_loader, device)
+            elif validate_every > 0 and epoch % validate_every == 0:
+                stage1 = trainers.get("stage1")
+                ddpm_steps = 1000
+                if isinstance(stage1, FlowDDPMStageTrainer) and stage1.ddpm_schedule is not None:
+                    ddpm_steps = stage1.ddpm_schedule.T
+                marginal_kl = _marginal_kl(
+                    trainers,
+                    val_loader,
+                    device,
+                    steps=validate_steps,
+                    ddpm_steps=ddpm_steps,
+                )
+
+            val_loss = sum(
+                trainer.val_objective(
+                    collector.train_means(name),
+                    collector.val_means(name),
+                    marginal_kl,
+                )
+                for name, trainer in trainers.items()
+            )
+
+            epoch_time = time.monotonic() - epoch_start
+            is_best = val_loss < best_val_loss
+            collector.set("val/loss", val_loss)
+            collector.set("val/marginal_kl", marginal_kl)
+            collector.set(
+                "gpu_mem_mb",
+                torch.cuda.max_memory_allocated(device) / (1024 * 1024) if device.type == "cuda" else 0.0,
+            )
+            collector.set("samples_per_sec", collector.train_samples / max(epoch_time, 1e-8))
+            collector.set("is_best", int(is_best))
+            collector.set("epoch_time_s", epoch_time)
+
+            print(collector.summary_line(val_loss, epoch_time, is_best))
+            collector.write_epoch(global_step)
+
+            ckpt = build_checkpoint(trainers, epoch, global_step, best_val_loss, checkpoint_extras)
+            if is_best:
+                best_val_loss = val_loss
+                ckpt["best_val_loss"] = best_val_loss
+                torch.save(ckpt, out_dir / "best.pt")
+            torch.save(ckpt, out_dir / "last.pt")
+            last_completed_epoch = epoch
+
+            if shutdown.requested:
+                break
+
+    collector.close()
+
+    if shutdown.requested:
+        print(
+            f"stopped after epoch {last_completed_epoch} due to shutdown signal — "
+            f"resume with --resume {out_dir / 'last.pt'}"
+        )

giant/training/metrics.py

@@ -0,0 +1,384 @@
+"""Per-epoch metric accumulation, `metrics.csv`, and W&B logging.
+
+Every scalar a training run reports is declared exactly once, as a
+`MetricSpec` on the `StageTrainer` that computes it (see
+`giant.training.trainers`). `MetricsCollector` derives the CSV/W&B column set
+from those declarations, so adding a metric means adding one line next to the
+code that produces it — there is no second list to keep in sync.
+
+Column naming is uniform: `<stage>/train/<key>`, `<stage>/val/<key>`,
+`<stage>/<key>` for point-in-time values (`lr`, `critic_lr`),
+`<stage>/router/<key>` for routing diagnostics, and an unprefixed run-level
+tail (`val/loss`, `grad_norm`, `epoch_time_s`, ...). W&B groups panels on
+`/`, so the same names read well there.
+"""
+
+import csv
+from dataclasses import dataclass
+from pathlib import Path
+
+import torch
+
+_ROUTER_KEYS = ("entropy", "util_min", "util_max", "util_std")
+
+# Written after every stage's columns, by `MetricsCollector` itself rather
+# than by any one trainer — these describe the run, not a stage.
+_RUN_COLUMNS = (
+    "val/loss",
+    "val/marginal_kl",
+    "grad_norm",
+    "gpu_mem_mb",
+    "samples_per_sec",
+    "is_best",
+    "epoch_time_s",
+)
+
+# tqdm/W&B batch-granularity smoothing, matching v0.2/v0.3.0's inline EMA.
+_EMA_ALPHA = 0.05
+
+
+@dataclass(frozen=True)
+class MetricSpec:
+    """One scalar a trainer emits per batch, and how it is reported.
+
+    `key` indexes the dict `StageTrainer.step()` / `.val_loss()` returns;
+    `column` is the CSV/W&B column suffix, joined to the stage name with
+    "/". `reduce` is either "mean" (batch-size-weighted average over the
+    epoch) or "last" (the most recent value — for point-in-time quantities
+    like the learning rate, which is a schedule readout, not a statistic).
+    """
+
+    key: str
+    column: str
+    reduce: str = "mean"
+
+
+def train_metric(key: str, column: str | None = None) -> MetricSpec:
+    return MetricSpec(key, column or f"train/{key}")
+
+
+def val_metric(key: str, column: str | None = None) -> MetricSpec:
+    return MetricSpec(key, column or f"val/{key}")
+
+
+def stage_metric(key: str, column: str | None = None) -> MetricSpec:
+    """A point-in-time stage-level readout (`lr`, `critic_lr`) — reported
+    unprefixed by split, as `<stage>/<key>`."""
+    return MetricSpec(key, column or key, reduce="last")
+
+
+def _wandb_run_config(cfg: dict, model_config: dict | None, param_counts: dict) -> dict:
+    return {
+        "train": cfg["train"],
+        "conditioning": cfg["conditioning"],
+        "stage1_model": cfg["stage1_model"],
+        "stage2_model": cfg["stage2_model"],
+        "model_config": model_config or {},
+        "param_counts": param_counts,
+    }
+
+
+class _Accumulator:
+    """Batch-size-weighted sums for one stage and one split."""
+
+    def __init__(self) -> None:
+        self.sums: dict[str, float] = {}
+        self.n = 0
+        self.last: dict[str, float] = {}
+
+    def add(self, stats: dict, keys: set[str], batch_size: int) -> None:
+        for key in keys:
+            if key in stats:
+                self.sums[key] = self.sums.get(key, 0.0) + stats[key] * batch_size
+        self.last.update(stats)
+        self.n += batch_size
+
+    def mean(self, key: str) -> float:
+        return self.sums.get(key, 0.0) / max(self.n, 1)
+
+    def means(self) -> dict[str, float]:
+        return {key: self.mean(key) for key in self.sums}
+
+    def reset(self) -> None:
+        self.sums.clear()
+        self.last.clear()
+        self.n = 0
+
+
+class _RouterAccumulator:
+    """Gate-diagnostic sums for one routed stage."""
+
+    def __init__(self, n_experts: int) -> None:
+        self.n_experts = n_experts
+        self.entropy = 0.0
+        self.importance: torch.Tensor | None = None
+        self.n = 0
+
+    def add(self, entropy: torch.Tensor, importance: torch.Tensor, n: int) -> None:
+        self.entropy += entropy.item() * n
+        self.importance = importance.clone() if self.importance is None else self.importance + importance
+        self.n += n
+
+    def stats(self) -> dict[str, float]:
+        if self.importance is None or self.n == 0:
+            return dict.fromkeys(_ROUTER_KEYS, 0.0)
+        util = self.importance / self.importance.sum().clamp_min(1e-8)
+        return {
+            "entropy": self.entropy / self.n,
+            "util_min": util.min().item(),
+            "util_max": util.max().item(),
+            "util_std": util.std().item() if self.n_experts > 1 else 0.0,
+        }
+
+    def reset(self) -> None:
+        self.entropy = 0.0
+        self.importance = None
+        self.n = 0
+
+
+class MetricsCollector:
+    """Owns every number a training run reports.
+
+    Accumulates per-batch stats from each stage, writes one `metrics.csv` row
+    per epoch, mirrors it to W&B, and formats the tqdm postfix and the epoch
+    summary line — so `giant.training.loop.train` never carries a running
+    sum, a column name, or a W&B call of its own.
+    """
+
+    def __init__(
+        self,
+        trainers: dict,
+        out_dir: Path,
+        *,
+        epochs: int,
+        resume: bool = False,
+        wandb_run=None,
+        wandb_log_every: int = 50,
+    ) -> None:
+        self.trainers = trainers
+        self.epochs = epochs
+        self.wandb_run = wandb_run
+        self.wandb_log_every = wandb_log_every
+        self.epoch_width = len(str(epochs))
+
+        self._train = {name: _Accumulator() for name in trainers}
+        self._val = {name: _Accumulator() for name in trainers}
+        self._routers = {
+            name: _RouterAccumulator(tr.router.n_experts) for name, tr in trainers.items() if tr.router is not None
+        }
+        # Only "mean" specs need summing; "last" specs are read straight off
+        # the accumulator's most recent stats dict. "grad_norm" is always
+        # summed — it feeds the run-level `grad_norm` column whether or not
+        # a trainer reports it per stage.
+        self._train_keys = {
+            name: {spec.key for spec in tr.train_metrics if spec.reduce == "mean"} | {"grad_norm"}
+            for name, tr in trainers.items()
+        }
+        self._val_keys = {
+            name: {spec.key for spec in tr.val_metrics if spec.reduce == "mean"} for name, tr in trainers.items()
+        }
+        self._run_values: dict[str, float] = {}
+        self._epoch = 0
+        self._ema_loss = 0.0
+        self._ema_grad_norm = 0.0
+        self._ema_seeded = False
+        self._batch_loss = 0.0
+        self._batch_grad_norm = 0.0
+
+        self.fieldnames = self._build_fieldnames()
+        metrics_path = out_dir / "metrics.csv"
+        append = resume and metrics_path.exists()
+        self._file = open(metrics_path, "a" if append else "w", newline="")
+        self._writer = csv.DictWriter(self._file, fieldnames=self.fieldnames)
+        if not append:
+            self._writer.writeheader()
+
+    # --- construction ---------------------------------------------------
+
+    @classmethod
+    def create(
+        cls,
+        trainers: dict,
+        out_dir: Path,
+        cfg: dict,
+        model_config: dict | None,
+        *,
+        resume: bool = False,
+        use_wandb: bool = False,
+        wandb_project: str = "giant",
+        wandb_run_name: str = "",
+        wandb_log_every: int = 50,
+    ) -> "MetricsCollector":
+        """Build the collector, starting a W&B run first when enabled."""
+        wandb_run = None
+        if use_wandb:
+            try:
+                import wandb
+            except ImportError as exc:
+                raise RuntimeError(
+                    "train.wandb = true (--wandb) requires the 'wandb' package — install it via `uv sync --extra wandb`"
+                ) from exc
+            param_counts = {name: sum(p.numel() for p in tr.model.parameters()) for name, tr in trainers.items()}
+            param_counts["total"] = sum(param_counts.values())
+            wandb_run = wandb.init(
+                project=wandb_project,
+                name=wandb_run_name or out_dir.name,
+                id=out_dir.name,
+                resume="allow",
+                config=_wandb_run_config(cfg, model_config, param_counts),
+            )
+        return cls(
+            trainers,
+            out_dir,
+            epochs=cfg["train"]["epochs"],
+            resume=resume,
+            wandb_run=wandb_run,
+            wandb_log_every=wandb_log_every,
+        )
+
+    def _build_fieldnames(self) -> list[str]:
+        fields = ["epoch"]
+        for name, trainer in self.trainers.items():
+            for spec in trainer.train_metrics:
+                fields.append(f"{name}/{spec.column}")
+            for spec in trainer.val_metrics:
+                fields.append(f"{name}/{spec.column}")
+            if trainer.router is not None:
+                fields += [f"{name}/router/{key}" for key in _ROUTER_KEYS]
+            for spec in trainer.stage_metrics:
+                fields.append(f"{name}/{spec.column}")
+        fields += list(_RUN_COLUMNS)
+        return fields
+
+    def close(self) -> None:
+        self._file.close()
+        if self.wandb_run is not None:
+            self.wandb_run.finish()
+
+    # --- per-batch ------------------------------------------------------
+
+    def start_epoch(self, epoch: int) -> None:
+        self._epoch = epoch
+        for acc in self._train.values():
+            acc.reset()
+        for acc in self._val.values():
+            acc.reset()
+        for acc in self._routers.values():
+            acc.reset()
+        self._run_values.clear()
+        self._ema_seeded = False
+
+    def add_train_batch(self, stats: dict[str, dict], batch_size: int) -> None:
+        """`stats` maps stage name -> the dict that stage's `step()` returned."""
+        self._batch_loss = 0.0
+        self._batch_grad_norm = 0.0
+        for name, stage_stats in stats.items():
+            self._train[name].add(stage_stats, self._train_keys[name], batch_size)
+            self._batch_loss += self.trainers[name].batch_loss(stage_stats)
+            self._batch_grad_norm += stage_stats.get("grad_norm", 0.0)
+        if self._ema_seeded:
+            self._ema_loss += _EMA_ALPHA * (self._batch_loss - self._ema_loss)
+            self._ema_grad_norm += _EMA_ALPHA * (self._batch_grad_norm - self._ema_grad_norm)
+        else:
+            self._ema_loss = self._batch_loss
+            self._ema_grad_norm = self._batch_grad_norm
+            self._ema_seeded = True
+
+    def add_val_batch(self, stats: dict[str, dict], batch_size: int) -> None:
+        for name, stage_stats in stats.items():
+            self._val[name].add(stage_stats, self._val_keys[name], batch_size)
+
+    @torch.no_grad()
+    def observe_routers(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, batch_size: int) -> None:
+        """Record gate diagnostics for every routed stage on this batch.
+
+        Called from the validation pass only (as in v0.2/v0.3.0), so a stage
+        whose trainer has no validation pass — i.e. WGAN — reports zeros.
+        """
+        for name, acc in self._routers.items():
+            router = self.trainers[name].router
+            entropy, importance = router.gate_stats(cond_cont, cond_cat)
+            acc.add(entropy, importance, batch_size)
+
+    def postfix(self) -> str:
+        """tqdm postfix for the training bar."""
+        return f"loss={self._ema_loss:.4f} gnorm={self._ema_grad_norm:.3f}"
+
+    def log_batch(self, global_step: int, batch: tuple, device: torch.device) -> None:
+        """Batch-granularity W&B log, throttled to every `wandb_log_every`
+        optimizer steps (a single epoch can be tens of thousands). `batch` is
+        the raw training batch, needed only to re-derive routing entropy for
+        routed stages — it is never moved to `device` otherwise."""
+        if self.wandb_run is None or self.wandb_log_every <= 0:
+            return
+        if global_step % self.wandb_log_every != 0:
+            return
+        payload = {
+            "batch/epoch": self._epoch,
+            "batch/loss": self._batch_loss,
+            "batch/loss_ema": self._ema_loss,
+            "batch/grad_norm": self._batch_grad_norm,
+        }
+        for name, trainer in self.trainers.items():
+            payload[f"batch/{name}/lr"] = trainer.optimizer.param_groups[0]["lr"]
+            if trainer.router is not None:
+                with torch.no_grad():
+                    entropy, _ = trainer.router.gate_stats(batch[0].to(device), batch[1].to(device))
+                payload[f"batch/{name}/router/entropy"] = entropy.item()
+        self.wandb_run.log(payload, step=global_step)
+
+    # --- per-epoch ------------------------------------------------------
+
+    def train_means(self, stage: str) -> dict[str, float]:
+        return self._train[stage].means()
+
+    def val_means(self, stage: str) -> dict[str, float]:
+        return self._val[stage].means()
+
+    @property
+    def train_samples(self) -> int:
+        """Samples seen this epoch — identical across stages (every stage
+        steps on every batch), so any one accumulator's count will do."""
+        return max((acc.n for acc in self._train.values()), default=0)
+
+    def set(self, column: str, value: float) -> None:
+        """Record a run-level value for this epoch's row (`val/loss`,
+        `gpu_mem_mb`, ...). Must name a column in `_RUN_COLUMNS`."""
+        if column not in _RUN_COLUMNS:
+            raise KeyError(f"{column!r} is not a run-level metrics column")
+        self._run_values[column] = value
+
+    def summary_line(self, val_loss: float, epoch_time: float, is_best: bool) -> str:
+        bits = [trainer.summary(self.train_means(name)) for name, trainer in self.trainers.items()]
+        marker = "  [best]" if is_best else ""
+        return (
+            f"epoch {self._epoch:{self.epoch_width}d}/{self.epochs}  "
+            + "  ".join(bits)
+            + f"  val {val_loss:.4f}  {epoch_time:.1f}s{marker}"
+        )
+
+    def write_epoch(self, global_step: int) -> None:
+        """Assemble, write, and flush this epoch's row; mirror it to W&B."""
+        row: dict = {"epoch": self._epoch}
+        grad_norm_total = 0.0
+        for name, trainer in self.trainers.items():
+            train_acc, val_acc = self._train[name], self._val[name]
+            for spec in trainer.train_metrics:
+                row[f"{name}/{spec.column}"] = train_acc.mean(spec.key)
+            for spec in trainer.val_metrics:
+                row[f"{name}/{spec.column}"] = val_acc.mean(spec.key)
+            if trainer.router is not None:
+                for key, value in self._routers[name].stats().items():
+                    row[f"{name}/router/{key}"] = value
+            for spec in trainer.stage_metrics:
+                row[f"{name}/{spec.column}"] = train_acc.last.get(spec.key, 0.0)
+            grad_norm_total += train_acc.mean("grad_norm")
+
+        for column in _RUN_COLUMNS:
+            row[column] = self._run_values.get(column, float("nan"))
+        row["grad_norm"] = grad_norm_total
+
+        self._writer.writerow(row)
+        self._file.flush()
+        if self.wandb_run is not None:
+            self.wandb_run.log(row, step=global_step)

giant/training/stage2_inputs.py

@@ -0,0 +1,321 @@
+"""Ground-truth tensor assembly for stage-2 training.
+
+Pure functions, no optimizer/model state: they turn a batch's ground-truth
+secondary tensors into the per-token targets and autoregressive conditioning
+inputs `giant.training.trainers` feeds to `Stage2OneShot` /
+`Stage2Autoregressive`. Split out of the trainers so the (target, generator,
+decoder) width rules — the fiddliest part of this codebase —
+live in one place and stay unit-testable on their own.
+"""
+
+import torch
+import torch.nn.functional as F
+
+from giant.config import ParticleTypeConfig
+from giant.constants import CONT_SLOT_DIM, PARTICLE_PHYS_DIM
+from giant.model.objectives import build_objective
+from giant.sample import sample_secondaries_ar
+
+
+def _gumbel_tau(step: int, total_steps: int, tau_start: float, tau_end: float) -> float:
+    """Linear anneal of the straight-through Gumbel-softmax temperature.
+
+    Deterministic in `step`/`total_steps` alone (no extra state), so it
+    recomputes correctly on `--resume` from a checkpoint's saved `global_step`
+    without needing to persist anything new (see
+    giant.model.network.Router.combine_weights).
+    """
+    progress = min(step / max(total_steps, 1), 1.0)
+    return tau_start + (tau_end - tau_start) * progress
+
+
+def _type_repr(
+    sec_type_idx: torch.Tensor,
+    sec_cont: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+    cond_enc: torch.nn.Module,
+    emb_dim: int,
+) -> torch.Tensor:
+    """(B, K_MAX, type_dim) ground-truth type representation, generator-
+    independent (unlike `_assemble_stage2_ar_target`'s training *target*,
+    which varies by generator/objective — see its docstring): `"physical"` ->
+    `(log_mass, charge)`; `"onehot"` -> one-hot of the true class;
+    `"embedding"` -> the conditioning's own detached embedding-table row.
+
+    Used both to build `_assemble_stage2_ar_target`'s wgan+onehot/embedding
+    branch and as the AR history features' previous-secondary identity — the
+    latter must always reflect the true physical secondary that came before,
+    regardless of what the *current* token's own training objective is.
+    """
+    target = particle_type_cfg.target
+    if target == "physical":
+        return sec_cont[..., CONT_SLOT_DIM : CONT_SLOT_DIM + PARTICLE_PHYS_DIM]
+    if target == "onehot":
+        return F.one_hot(sec_type_idx, num_classes=emb_dim).float()
+    return cond_enc.pdg_emb(sec_type_idx).detach()
+
+
+def _assemble_stage2_ar_target(
+    sec_cont: torch.Tensor,
+    sec_type_idx: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+    generator: str,
+    cond_enc: torch.nn.Module,
+    emb_dim: int,
+) -> torch.Tensor:
+    """(B, K_MAX, token_dim) ground-truth per-token target — the unflattened
+    analogue of `_assemble_stage2_real` (defined below in terms of this),
+    matching whatever width `Stage2Autoregressive`'s (or `Stage2OneShot`'s)
+    own trunk produces for this (target, generator) combination
+    (`giant.model.network.stage2_trunk_sec_dim`):
+
+    - `target = "physical"`: unchanged from v0.2 — `sec_cont` (stick_logit,
+      dir, log_mass, charge) as-is.
+    - `target` in `("onehot", "embedding")` + an objective that doesn't fold
+      the type slice (flow/ddpm): just the continuous stick/dir slots — the
+      type slice isn't part of this tensor at all (`type_head` handles it
+      separately).
+    - `target` in `("onehot", "embedding")` + a folding objective (wgan):
+      stick/dir slots concatenated with the per-slot type representation (a
+      one-hot of the true class, relaxed on the *generated* side only, by the
+      caller; or the conditioning's own detached embedding-table row).
+    """
+    target = particle_type_cfg.target
+    if target == "physical":
+        return sec_cont
+    cont = sec_cont[..., :CONT_SLOT_DIM]
+    if not build_objective(generator).folds_type_slice:
+        return cont
+    type_repr = _type_repr(sec_type_idx, sec_cont, particle_type_cfg, cond_enc, emb_dim)
+    return torch.cat([cont, type_repr], dim=-1)
+
+
+def _assemble_stage2_real(
+    sec_cont: torch.Tensor,
+    sec_type_idx: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+    generator: str,
+    cond_enc: torch.nn.Module,
+    emb_dim: int,
+) -> torch.Tensor:
+    """Ground-truth flattened stage-2 vector for `Stage2OneShot` — the
+    flattened form of `_assemble_stage2_ar_target`, which
+    `Stage2Autoregressive`'s per-token target also uses; the two must stay in
+    lockstep. See `_assemble_stage2_ar_target`'s docstring for the
+    (target, generator) width rules."""
+    return _assemble_stage2_ar_target(sec_cont, sec_type_idx, particle_type_cfg, generator, cond_enc, emb_dim).flatten(
+        1
+    )
+
+
+def _stick_fraction(sec_cont: torch.Tensor) -> torch.Tensor:
+    """(B, K_MAX) — sigmoid of each slot's own stick-breaking logit
+    (`sec_cont[...,0]`); scale-free (see `giant.data.transforms.
+    encode_secondaries`), so this needs no absolute `e_sec`."""
+    return torch.sigmoid(sec_cont[..., 0])
+
+
+def _remaining_energy_fraction(fraction: torch.Tensor) -> torch.Tensor:
+    """(B, K_MAX) — fraction of the original e_sec budget unclaimed entering
+    slot i: `1.0` at `i=0`, `prod_{j<i}(1-fraction_j)` for `i>=1`
+    ("no re-derivation needed": the existing
+    stick-breaking encoding is already scale-free, so this is derivable from
+    the batch's ground-truth stick logits alone, no `e_sec` required)."""
+    cumprod = torch.cumprod(1.0 - fraction, dim=1)
+    return torch.cat([torch.ones_like(cumprod[:, :1]), cumprod[:, :-1]], dim=1)
+
+
+def _shift_prev(x: torch.Tensor) -> torch.Tensor:
+    """`(B, K, ...)` -> same shape, slot i holds slot i-1's value; slot 0 gets
+    an arbitrary zero placeholder (never read as-is — see `_ar_has_prev`;
+    `MarkovHistory` substitutes its own learned start vector there instead)."""
+    return torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1)
+
+
+def _ar_has_prev(k_max: int, device: torch.device) -> torch.Tensor:
+    """`(1, K_MAX)` bool: True for slot index `>= 1`. Correct without
+    `n_sec`: `sec_mask` is a prefix mask, so any *valid* token at `k>=1`
+    always has a valid predecessor at `k-1`; the only wrong cases are tokens
+    that are themselves padding, already masked out of every loss."""
+    return (torch.arange(k_max, device=device) >= 1).unsqueeze(0)
+
+
+def _ar_meta(k_max: int, batch: int, device: torch.device, fraction: torch.Tensor) -> dict[str, torch.Tensor]:
+    """`has_prev`/`remaining_frac`/`slot_idx` — the three per-token AR
+    conditioning tensors that don't depend on *which* history representation
+    (ground truth vs. the scheduled-sampling mix) produced `fraction`.
+    Shared by `_assemble_stage2_ar_inputs` and
+    `_assemble_stage2_ar_inputs_scheduled`, which differ only in
+    `history_feat`."""
+    slot_idx = (torch.arange(k_max, device=device).float() / max(k_max - 1, 1)).unsqueeze(0)
+    return {
+        "has_prev": _ar_has_prev(k_max, device).expand(batch, -1),
+        "remaining_frac": _remaining_energy_fraction(fraction),
+        "slot_idx": slot_idx.expand(batch, -1),
+    }
+
+
+def _assemble_stage2_ar_inputs(
+    sec_cont: torch.Tensor,
+    sec_type_idx: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+    cond_enc: torch.nn.Module,
+    emb_dim: int,
+) -> dict[str, torch.Tensor]:
+    """Ground-truth per-token AR conditioning tensors — all `(B, K_MAX, ...)`
+    or `(B, K_MAX)`, built in one vectorized pass (teacher forcing means
+    every token's input is ground truth).
+    Keys match `Stage2Autoregressive.forward`'s trailing kwargs."""
+    device = sec_cont.device
+    B, K = sec_cont.shape[0], sec_cont.shape[1]
+    fraction = _stick_fraction(sec_cont)
+    type_repr = _type_repr(sec_type_idx, sec_cont, particle_type_cfg, cond_enc, emb_dim)
+    history_feat = torch.cat(
+        [
+            _shift_prev(fraction).unsqueeze(-1),
+            _shift_prev(sec_cont[..., 1:CONT_SLOT_DIM]),
+            _shift_prev(type_repr),
+        ],
+        dim=-1,
+    )
+    return {"history_feat": history_feat, **_ar_meta(K, B, device, fraction)}
+
+
+def _stage2_tf_prob(mode: str, p_start: float, p_end: float, epoch: int, total_epochs: int) -> float:
+    """P(condition slot k+1 on the TRUE token k rather than the model's own
+    prediction), for the current epoch
+    (`stage2_model.autoregressive.teacher_forcing`).
+    `"always"`/`"never"` are the two degenerate constants; `"scheduled"`
+    linearly interpolates
+    `p_start` (epoch 0) to `p_end` (the final epoch) — standard scheduled
+    sampling (Bengio et al. 2015)."""
+    if mode == "always":
+        return 1.0
+    if mode == "never":
+        return 0.0
+    frac = epoch / max(total_epochs - 1, 1)
+    frac = min(max(frac, 0.0), 1.0)
+    return p_start + (p_end - p_start) * frac
+
+
+def _history_repr_from_ar_sample(
+    sec_cont_pred: torch.Tensor,
+    sec_type_pred: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """`(fraction, direction, type_repr)` — the same triple `_type_repr` /
+    `_stick_fraction` derive from ground truth, but from a free-running
+    `sample_secondaries_ar` self-sample instead, so the two can be mixed
+    slot-by-slot under scheduled sampling (`_assemble_stage2_ar_inputs_scheduled`).
+    `target="onehot"` collapses the raw per-slot type logits to a hard
+    one-hot of `argmax` — `sample_secondaries_ar`'s own history convention
+    (see its docstring), matching what `MarkovHistory`/`AttentionHistory`
+    were trained on; the other two targets are already the right
+    representation."""
+    fraction = torch.sigmoid(sec_cont_pred[..., 0])
+    direction = sec_cont_pred[..., 1:4]
+    if particle_type_cfg.target == "onehot":
+        type_dim = sec_type_pred.size(-1)
+        type_repr = F.one_hot(sec_type_pred.argmax(-1), num_classes=type_dim).float()
+    else:
+        type_repr = sec_type_pred
+    return fraction, direction, type_repr
+
+
+def _assemble_stage2_ar_inputs_scheduled(
+    model: torch.nn.Module,
+    cond_cont: torch.Tensor,
+    cond_cat: torch.Tensor,
+    stage1_ctx: torch.Tensor,
+    sec_cont: torch.Tensor,
+    sec_type_idx: torch.Tensor,
+    n_sec: torch.Tensor,
+    particle_type_cfg: ParticleTypeConfig,
+    cond_enc: torch.nn.Module,
+    emb_dim: int,
+    p_tf: float,
+    sample_steps: int,
+) -> dict[str, torch.Tensor]:
+    """Scheduled-sampling counterpart of `_assemble_stage2_ar_inputs`
+    (`teacher_forcing` = "scheduled"/"never"):
+    each slot's history is the TRUE previous token with probability `p_tf`
+    (an independent per-example, per-slot Bernoulli draw) and the model's own
+    free-running prediction otherwise — closing the train/inference gap that
+    `teacher_forcing="always"` (ground truth throughout training) never sees.
+    `p_tf >= 1.0` degenerates exactly to `_assemble_stage2_ar_inputs` (and
+    skips self-sampling entirely), so callers can call this unconditionally.
+
+    The free-running estimate is a REAL autoregressive self-sample —
+    `giant.sample.sample_secondaries_ar` under `torch.no_grad()` — not a
+    cheap one-step proxy, so building it costs the same `k_max` (`* steps`
+    for flow) sequential forwards `sample.py` pays at inference, EVERY batch
+    this is called on (paid at train time too whenever teacher_forcing !=
+    "always"). Fully detached: gradient only ever flows
+    through the "real" target path each stage trainer already uses
+    (`_assemble_stage2_ar_target`), never through this self-sample.
+    """
+    device = sec_cont.device
+    B, K = sec_cont.shape[0], sec_cont.shape[1]
+    if p_tf >= 1.0:
+        return _assemble_stage2_ar_inputs(sec_cont, sec_type_idx, particle_type_cfg, cond_enc, emb_dim)
+
+    was_training = model.training
+    sec_cont_pred, sec_type_pred, _ = sample_secondaries_ar(
+        model, cond_cont, cond_cat, stage1_ctx, n_sec, steps=sample_steps
+    )
+    if was_training:
+        model.train()
+
+    fraction_gt = _stick_fraction(sec_cont)
+    dir_gt = sec_cont[..., 1:CONT_SLOT_DIM]
+    type_repr_gt = _type_repr(sec_type_idx, sec_cont, particle_type_cfg, cond_enc, emb_dim)
+    fraction_pred, dir_pred, type_repr_pred = _history_repr_from_ar_sample(
+        sec_cont_pred, sec_type_pred, particle_type_cfg
+    )
+
+    use_gt = torch.rand(B, K, device=device) < p_tf
+    fraction = torch.where(use_gt, fraction_gt, fraction_pred)
+    direction = torch.where(use_gt.unsqueeze(-1), dir_gt, dir_pred)
+    type_repr = torch.where(use_gt.unsqueeze(-1), type_repr_gt, type_repr_pred)
+
+    own_feat = torch.cat([fraction.unsqueeze(-1), direction, type_repr], dim=-1)
+    return {
+        "history_feat": _shift_prev(own_feat),
+        **_ar_meta(K, B, device, fraction),
+    }
+
+
+def _relax_onehot_type_slice(
+    x_flat: torch.Tensor,
+    k_max: int,
+    cont_dim: int,
+    type_dim: int,
+    tau: float,
+    grad_probe: dict[str, float] | None = None,
+) -> torch.Tensor:
+    """Straight-through Gumbel-softmax relaxation of the per-slot type slice
+    inside a flattened `(B, k_max * (cont_dim + type_dim))` WGAN generator
+    output: the forward pass is a
+    hard one-hot (matching what the critic sees from real data), the
+    backward pass flows smooth gradient. Continuous slots (stick/dir, and
+    the type slice itself under `target = "embedding"`, which never calls
+    this) pass through unchanged.
+
+    `grad_probe`, if given, gets `["cont"]`/`["type"]` populated with the L2
+    norm of the gradient reaching this split point during the next
+    `.backward()` call that touches it — a backward hook, not a second
+    backward pass. This is the differentiability validation-obligation
+    instrumentation: the trunk-gradient contribution
+    from the type slice vs. the continuous slices, for
+    `particle_type.target="onehot"` + `generator="wgan"`. Only ever populated
+    on a `did_g_step` batch — the critic step backprops through
+    `fake.detach()`, which never reaches these hooks — so it stays empty
+    (callers default to `0.0`) otherwise."""
+    B = x_flat.size(0)
+    x = x_flat.view(B, k_max, cont_dim + type_dim)
+    cont, type_logits = x[..., :cont_dim], x[..., cont_dim:]
+    if grad_probe is not None:
+        cont.register_hook(lambda g: grad_probe.__setitem__("cont", g.norm().item()))
+        type_logits.register_hook(lambda g: grad_probe.__setitem__("type", g.norm().item()))
+    type_soft = F.gumbel_softmax(type_logits, tau=tau, hard=True, dim=-1)
+    return torch.cat([cont, type_soft], dim=-1).reshape(B, -1)

giant/training/trainers.py

@@ -0,0 +1,949 @@
+"""Per-stage trainers: optimizer(s), EMA, LR schedule, and the per-batch step.
+
+`StageSpec` resolves one stage's slice of the config once, so the two
+concrete trainers share a single constructor shape instead of ~24 keyword
+arguments each, and `StageTrainer` carries every piece that used to be
+copy-pasted between them (cosine warmup, EMA, checkpoint state, LR resume,
+train/eval toggling).
+
+Each trainer also *declares* the metrics it emits, as `MetricSpec` lists —
+that declaration is the single source of truth for `metrics.csv` and W&B
+columns (see `giant.training.metrics`) — and exposes the three small hooks
+(`batch_loss`, `summary`, `val_objective`) that let the epoch loop treat
+adversarial and non-adversarial stages identically.
+"""
+
+import copy
+import math
+from dataclasses import dataclass, field
+from typing import NamedTuple
+
+import torch
+import torch.nn.functional as F
+import torch.optim as optim
+
+from giant.config import ParticleTypeConfig, Stage1ModelConfig, Stage2ModelConfig, TrainConfig
+from giant.constants import CONT_SLOT_DIM
+from giant.data.dataset import StepBatch
+from giant.model.network import Router, build_objective, resolve_type_n_classes, stage2_type_dim
+from giant.model.wgan import generator_loss, gradient_penalty
+from giant.training.metrics import MetricSpec, stage_metric, train_metric, val_metric
+from giant.training.stage2_inputs import (
+    _assemble_stage2_ar_inputs_scheduled,
+    _assemble_stage2_ar_target,
+    _gumbel_tau,
+    _relax_onehot_type_slice,
+    _stage2_tf_prob,
+)
+
+
+@torch.no_grad()
+def _update_ema(ema_model: torch.nn.Module, model: torch.nn.Module, decay: float) -> None:
+    for ema_p, p in zip(ema_model.parameters(), model.parameters()):
+        ema_p.mul_(decay).add_(p, alpha=1 - decay)
+
+
+def _stage_router(model: torch.nn.Module) -> Router | None:
+    """A stage model's Router, if its trunk is routed — else None.
+
+    Post-step-2 refactor the router lives at `model.trunk.router`
+    (`giant.model.network.RoutedTrunk`), not `model.router` directly.
+    """
+    trunk = getattr(model, "trunk", None)
+    return getattr(trunk, "router", None)
+
+
+def _cosine_warmup_lambda(warmup_steps: int, total_steps: int):
+    """Linear warmup for `warmup_steps`, then cosine decay to zero over the
+    remainder — the LR schedule both trainers use, in their own step units
+    (optimizer steps for flow/ddpm, generator steps for WGAN)."""
+
+    def _lr_lambda(step: int) -> float:
+        if warmup_steps > 0 and step < warmup_steps:
+            return (step + 1) / warmup_steps
+        t = step - warmup_steps
+        T = max(total_steps - warmup_steps, 1)
+        return 0.5 * (1.0 + math.cos(math.pi * min(t, T) / T))
+
+    return _lr_lambda
+
+
+def _batch_to_device(batch: StepBatch, device: torch.device) -> StepBatch:
+    return type(batch)(*(t.to(device) for t in batch))
+
+
+@dataclass(frozen=True)
+class StageSpec:
+    """One stage's resolved training configuration.
+
+    Built once by `StageSpec.from_config`, which is the only place that reads
+    the `cfg` dict — so a new config key means one new field and one new read,
+    not another argument threaded through two constructors.
+    """
+
+    name: str
+    is_stage2: bool
+    generator: str
+    decoder: str = "one_shot"
+
+    # loss weights
+    lambda_weight: float = 1.0
+    n_sec_lambda: float = 0.1
+
+    # particle-type target (stage 2 only)
+    particle_type: ParticleTypeConfig = field(default_factory=ParticleTypeConfig)
+    particle_type_n_classes: int = 16
+
+    # optimization
+    lr: float = 3e-4
+    weight_decay: float = 0.01
+    ema_decay: float = 0.9999
+    warmup_epochs: int = 0
+    epochs: int = 1
+    steps_per_epoch: int = 1
+
+    # routing auxiliaries
+    lambda_balance: float = 0.0
+    lambda_proc: float = 0.0
+    lambda_entropy: float = 0.0
+    gumbel_tau_start: float = 1.0
+    gumbel_tau_end: float = 0.1
+
+    # autoregressive stage 2
+    teacher_forcing: str = "always"
+    tf_p_start: float = 1.0
+    tf_p_end: float = 1.0
+    ar_sample_steps: int = 10
+
+    # generator-specific
+    ddpm_n_steps: int = 1000
+    n_critic: int = 5
+    gp_weight: float = 10.0
+    critic_lr: float = 0.0
+    type_gumbel_tau_start: float = 1.0
+    type_gumbel_tau_end: float = 0.1
+
+    @classmethod
+    def from_config(cls, cfg: dict, name: str, is_stage2: bool, steps_per_epoch: int) -> "StageSpec":
+        t = TrainConfig.from_dict(cfg["train"])
+        # n_sec/particle_type/decoder/autoregressive/wgan's gumbel_tau_* are
+        # stage-2-only concepts, always read off s2_spec (guarded by
+        # is_stage2 where the stage-1 StageSpec needs a different value) —
+        # historically n_sec/particle_type were read from stage2_model
+        # unconditionally even for the stage-1 StageSpec, preserved here for
+        # behavioral parity. stage_spec covers the fields both stage configs
+        # share structurally (generator, lambda, router, ddpm, and wgan's
+        # base fields — Stage2ModelConfig's sub-configs all subclass
+        # stage 1's).
+        s2_spec = Stage2ModelConfig.from_dict(cfg["stage2_model"])
+        stage_spec = s2_spec if is_stage2 else Stage1ModelConfig.from_dict(cfg["stage1_model"])
+        return cls(
+            name=name,
+            is_stage2=is_stage2,
+            generator=stage_spec.generator,
+            decoder=s2_spec.decoder if is_stage2 else "one_shot",
+            lambda_weight=stage_spec.lambda_weight,
+            n_sec_lambda=s2_spec.n_sec.lambda_weight,
+            particle_type=s2_spec.particle_type,
+            particle_type_n_classes=resolve_type_n_classes(
+                s2_spec.particle_type, cfg["conditioning"]["particle"]["emb_dim"]
+            ),
+            # train.* keys are all guaranteed by DEFAULT_CONFIG's deep-merge
+            # (giant/config.py), so TrainConfig.from_dict never has to fall
+            # back to a literal here; the field defaults below exist only
+            # for tests that construct StageSpec by hand.
+            lr=t.lr,
+            weight_decay=t.weight_decay,
+            ema_decay=t.ema_decay,
+            warmup_epochs=t.warmup_epochs,
+            epochs=t.epochs,
+            steps_per_epoch=max(steps_per_epoch, 1),
+            lambda_balance=stage_spec.router.lambda_balance,
+            lambda_proc=stage_spec.router.lambda_proc,
+            lambda_entropy=stage_spec.router.lambda_entropy,
+            gumbel_tau_start=stage_spec.router.gumbel_tau_start,
+            gumbel_tau_end=stage_spec.router.gumbel_tau_end,
+            teacher_forcing=s2_spec.autoregressive.teacher_forcing if is_stage2 else cls.teacher_forcing,
+            tf_p_start=s2_spec.autoregressive.tf_p_start if is_stage2 else cls.tf_p_start,
+            tf_p_end=s2_spec.autoregressive.tf_p_end if is_stage2 else cls.tf_p_end,
+            # AR self-sampling under scheduled/never teacher forcing reuses
+            # train.validate_steps as its flow-matching ODE step count — no
+            # dedicated config key for this (the autoregressive config lists
+            # tf_p_start/tf_p_end/attn_n_heads/attn_n_layers only).
+            ar_sample_steps=t.validate_steps,
+            ddpm_n_steps=stage_spec.ddpm.n_steps,
+            n_critic=stage_spec.wgan.n_critic,
+            gp_weight=stage_spec.wgan.gp_weight,
+            critic_lr=stage_spec.wgan.critic_lr,
+            type_gumbel_tau_start=s2_spec.wgan.gumbel_tau_start if is_stage2 else cls.type_gumbel_tau_start,
+            type_gumbel_tau_end=s2_spec.wgan.gumbel_tau_end if is_stage2 else cls.type_gumbel_tau_end,
+        )
+
+
+class StageTrainer:
+    """One active stage's optimizer(s), EMA, and per-batch step.
+
+    Reads only the shared `StepBatch` (`giant.data.dataset`) — stage 2 always
+    conditions on the ground-truth `x1_s1` (`stage2_model.stage1_context =
+    "truth"`, stage-level teacher forcing; `"sampled"` is not implemented),
+    so stage trainers never need each other's output at train time. This means
+    "stage-2-only training is a cheap ablation, not new plumbing" falls out
+    for free: a trainer only exists for active stages, and inactive stages
+    are simply never constructed.
+
+    Grad-norm clipping is per-stage here — v0.2's single shared optimizer
+    clipped both stages' gradients jointly; splitting per stage is a small,
+    disclosed behavior change. It doesn't affect Adam's per-parameter update
+    math itself (no cross-parameter coupling), only the clip threshold's
+    scope.
+    """
+
+    #: Metrics this trainer emits, declared once — `giant.training.metrics`
+    #: derives every CSV/W&B column from these. Instance attributes rather
+    #: than class constants because some are conditional on the stage's own
+    #: configuration (see `WGANStageTrainer.__init__`).
+    train_metrics: list[MetricSpec]
+    val_metrics: list[MetricSpec]
+    stage_metrics: list[MetricSpec]
+
+    #: False for adversarial stages, which have no monotone per-batch
+    #: validation loss worth averaging (see `val_objective`).
+    supports_val_loss: bool = True
+
+    #: Built by the subclass (the optimizer flavour differs) and wired to the
+    #: schedule via `_init_lr_schedule`.
+    optimizer: optim.Optimizer
+    lr_sched: optim.lr_scheduler.LambdaLR
+    total_steps: int
+
+    def __init__(
+        self,
+        spec: StageSpec,
+        model: torch.nn.Module,
+        device: torch.device,
+        extra_modules: tuple[torch.nn.Module, ...] = (),
+    ) -> None:
+        self.spec = spec
+        self.name = spec.name
+        self.is_stage2 = spec.is_stage2
+        self.generator = spec.generator
+        self.decoder = spec.decoder
+        self.device = device
+        self.model = model.to(device)
+        self.router = _stage_router(self.model)
+        self._modules = (self.model, *extra_modules)
+
+        self.particle_type_cfg = spec.particle_type
+        self.particle_type_n_classes = spec.particle_type_n_classes
+        self.ema_decay = spec.ema_decay
+
+        self.ema_model: torch.nn.Module | None = None
+        if spec.ema_decay > 0:
+            self.ema_model = copy.deepcopy(self.model).eval()
+            for p in self.ema_model.parameters():
+                p.requires_grad_(False)
+
+    # --- schedule -------------------------------------------------------
+
+    def _init_lr_schedule(self, optimizer: optim.Optimizer, warmup_steps: int, total_steps: int) -> None:
+        self._lr_lambda = _cosine_warmup_lambda(warmup_steps, total_steps)
+        self.total_steps = total_steps
+        self.lr_sched = optim.lr_scheduler.LambdaLR(optimizer, self._lr_lambda)
+
+    # --- per-batch (subclass responsibility) ----------------------------
+
+    def step(self, batch: StepBatch, device: torch.device, global_step: int) -> dict:
+        raise NotImplementedError
+
+    def val_loss(self, batch: StepBatch, device: torch.device) -> dict:
+        raise NotImplementedError
+
+    # --- reporting hooks ------------------------------------------------
+
+    def batch_loss(self, stats: dict) -> float:
+        """The single number this stage contributes to the progress bar's
+        smoothed loss."""
+        raise NotImplementedError
+
+    def summary(self, means: dict) -> str:
+        """This stage's fragment of the end-of-epoch console line."""
+        raise NotImplementedError
+
+    def val_objective(self, train_means: dict, val_means: dict, marginal_kl: float) -> float:
+        """This stage's contribution to the best-checkpoint selection score."""
+        raise NotImplementedError
+
+    # --- mode / state ---------------------------------------------------
+
+    def sampling_model(self) -> torch.nn.Module:
+        return self.ema_model if self.ema_model is not None else self.model
+
+    def train_mode(self) -> None:
+        for module in self._modules:
+            module.train()
+
+    def eval_mode(self) -> None:
+        for module in self._modules:
+            module.eval()
+
+    # --- stage-2 secondary assembly (shared by both trainer subclasses) ---
+
+    def _ar_inputs(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_ctx: torch.Tensor,
+        sec_cont: torch.Tensor,
+        sec_type_idx: torch.Tensor,
+        n_sec: torch.Tensor,
+        epoch: int | None,
+    ) -> dict[str, torch.Tensor]:
+        """Per-token AR conditioning for this stage's secondary decoder.
+
+        `epoch=None` means full teacher forcing (`p_tf=1.0`) regardless of
+        `spec.teacher_forcing` — the val-loss convention, kept in this one
+        place so both trainer subclasses honor it identically.
+        """
+        p_tf = (
+            1.0
+            if epoch is None
+            else _stage2_tf_prob(
+                self.spec.teacher_forcing,
+                self.spec.tf_p_start,
+                self.spec.tf_p_end,
+                epoch,
+                self.spec.epochs,
+            )
+        )
+        return _assemble_stage2_ar_inputs_scheduled(
+            self.model,
+            cond_cont,
+            cond_cat,
+            stage1_ctx,
+            sec_cont,
+            sec_type_idx,
+            n_sec,
+            self.particle_type_cfg,
+            self.model.cond_enc,
+            self.particle_type_n_classes,
+            p_tf,
+            self.spec.ar_sample_steps,
+        )
+
+    def _sec_target(
+        self,
+        sec_cont: torch.Tensor,
+        sec_type_idx: torch.Tensor,
+        generator: str,
+        *,
+        flatten: bool,
+    ) -> torch.Tensor:
+        """Ground-truth stage-2 target for this stage's secondary decoder,
+        per the (particle-type target, generator) width rules in
+        `_assemble_stage2_ar_target`. `flatten=True` gives `Stage2OneShot`'s
+        flattened `(B, K*token_dim)` form (the old `_real`); `flatten=False`
+        gives `Stage2Autoregressive`'s per-token `(B, K, token_dim)` form (the
+        old `_ar_target`) — the two are the same tensor modulo `.flatten(1)`,
+        so the width rules live in one place (`stage2_inputs.py`)."""
+        target = _assemble_stage2_ar_target(
+            sec_cont,
+            sec_type_idx,
+            self.particle_type_cfg,
+            generator,
+            self.model.cond_enc,
+            self.particle_type_n_classes,
+        )
+        return target.flatten(1) if flatten else target
+
+    @staticmethod
+    def _sec_mask(n_sec: torch.Tensor, k_max: int, device: torch.device) -> torch.Tensor:
+        """`(B, K_MAX)` bool prefix mask: slot k is valid iff `k < n_sec`."""
+        return torch.arange(k_max, device=device).unsqueeze(0) < n_sec.unsqueeze(1)
+
+    def _n_sec_loss(
+        self,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        stage1_ctx: torch.Tensor,
+        n_sec: torch.Tensor,
+        device: torch.device,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """`(l_nsec, nsec_acc)` for this stage's multiplicity classifier —
+        zeros when the stage owns no `n_sec_head` (stage 1 now that n_sec
+        defaults to stage 2, or any stage without the head). Owns the only
+        stage1-vs-stage2 `predict_n_sec` signature split, shared by the
+        non-adversarial and WGAN trainers.
+
+        Gated on `n_sec_head is None`, not on `n_sec.mode`: a future
+        `mode="stop_token"` model (currently rejected in
+        `validate_config`) carries no head and would train its EOS signal in
+        the generator/AR loss path instead, so this correctly stays zero.
+        """
+        if self.model.n_sec_head is None:
+            zero = torch.zeros((), device=device)
+            return zero, zero
+        logits = (
+            self.model.predict_n_sec(cond_cont, cond_cat, stage1_ctx)
+            if self.is_stage2
+            else self.model.predict_n_sec(cond_cont, cond_cat)
+        )
+        l_nsec = F.cross_entropy(logits, n_sec)
+        nsec_acc = (logits.argmax(dim=-1) == n_sec).float().mean()
+        return l_nsec, nsec_acc
+
+    @staticmethod
+    def _step_optimizer(optimizer: optim.Optimizer, loss: torch.Tensor, params: list) -> float:
+        """`zero_grad -> backward -> clip_grad_norm_(1.0) -> step`, returning
+        the pre-clip grad norm. The one place the grad-clip constant lives."""
+        optimizer.zero_grad()
+        loss.backward()
+        grad_norm = torch.nn.utils.clip_grad_norm_(params, 1.0)
+        optimizer.step()
+        return grad_norm.item()
+
+    def _extra_state(self) -> dict:
+        """Subclass state beyond model/optimizer/lr_sched/EMA."""
+        return {}
+
+    def _load_extra_state(self, sd: dict) -> None:
+        return None
+
+    def state_dict(self) -> dict:
+        sd = {
+            "model": self.model.state_dict(),
+            "optimizer": self.optimizer.state_dict(),
+            "lr_sched": self.lr_sched.state_dict(),
+        }
+        if self.ema_model is not None:
+            sd["model_ema"] = self.ema_model.state_dict()
+        sd.update(self._extra_state())
+        return sd
+
+    def load_state_dict(self, sd: dict) -> None:
+        self.model.load_state_dict(sd["model"])
+        self.optimizer.load_state_dict(sd["optimizer"])
+        self.lr_sched.load_state_dict(sd["lr_sched"])
+        if self.ema_model is not None:
+            self.ema_model.load_state_dict(sd.get("model_ema", sd["model"]))
+        self._load_extra_state(sd)
+
+    def _resume_extra_lr(self, lr: float) -> None:
+        return None
+
+    def resume_lr(self, lr: float) -> None:
+        """Restore the configured `lr`'s authority after `load_state_dict`
+        restored the checkpoint's own base LR."""
+        self.lr_sched.base_lrs = [lr for _ in self.lr_sched.base_lrs]
+        resumed_lr = lr * self._lr_lambda(self.lr_sched.last_epoch)
+        for group in self.optimizer.param_groups:
+            group["lr"] = resumed_lr
+        self._resume_extra_lr(lr)
+
+
+class FlowDDPMStageTrainer(StageTrainer):
+    """flow or ddpm generator for a single stage."""
+
+    def __init__(self, spec: StageSpec, model: torch.nn.Module, device: torch.device) -> None:
+        objective = build_objective(spec.generator, n_steps=spec.ddpm_n_steps)
+        if spec.is_stage2 and not objective.supports_stage2_decoder:
+            raise NotImplementedError(
+                f"stage2_model.generator={spec.generator!r} is accepted by the "
+                "schema but not implemented in v0.3.0 for stage 2 (only "
+                "'flow' and 'wgan' have a stage-2 secondary-decoder loss)"
+            )
+        super().__init__(spec, model, device)
+        self.objective = objective
+        self.particle_type_lambda = self.particle_type_cfg.lambda_weight
+        # Width of the type slice actually folded into x1_s2 by _sec_target,
+        # under this trainer's objective (flow/ddpm only — see the
+        # NotImplementedError above, neither folds the type slice): "physical"
+        # keeps it folded in (PARTICLE_PHYS_DIM wide, unchanged from v0.2);
+        # "onehot"/"embedding" pull it out into model.type_head instead (0
+        # here).
+        self._flow_type_dim = None if self.particle_type_cfg.target == "physical" else 0
+
+        self.params = list(self.model.parameters())
+        self.optimizer = optim.AdamW(self.params, lr=spec.lr, weight_decay=spec.weight_decay)
+        self._init_lr_schedule(
+            self.optimizer,
+            warmup_steps=spec.warmup_epochs * spec.steps_per_epoch,
+            total_steps=max(spec.epochs * spec.steps_per_epoch, 1),
+        )
+        self.ddpm_schedule = self.objective.build_schedule(spec.ddpm_n_steps, device)
+
+        self.train_metrics = [
+            train_metric(key)
+            for key in (
+                "loss",
+                "loss_gen",
+                "loss_nsec",
+                "loss_balance",
+                "loss_proc",
+                "loss_entropy",
+                "nsec_acc",
+                "loss_type",
+                "type_acc",
+                "grad_norm",
+            )
+        ]
+        self.val_metrics = [
+            val_metric(key)
+            for key in (
+                "loss",
+                "loss_gen",
+                "loss_nsec",
+                "nsec_acc",
+                "loss_type",
+                "type_acc",
+            )
+        ]
+        self.stage_metrics = [stage_metric("lr")]
+
+    def _generator_loss(self, cond_cont, cond_cat, x1_s1, x1_s2, sec_mask, stage1_ctx, ar_inputs=None):
+        if not self.is_stage2:
+            return self.objective.stage1_loss(self.model, x1_s1, cond_cont, cond_cat, schedule=self.ddpm_schedule)
+        assert self.decoder != "autoregressive" or ar_inputs is not None
+        return self.objective.stage2_loss(
+            self.model,
+            x1_s2,
+            cond_cont,
+            cond_cat,
+            stage1_ctx,
+            sec_mask,
+            type_dim=self._flow_type_dim,
+            ar_inputs=ar_inputs,
+        )
+
+    def _type_loss(
+        self,
+        cond_cont,
+        cond_cat,
+        stage1_ctx,
+        sec_type_idx,
+        sec_mask,
+        device,
+        ar_inputs=None,
+    ):
+        """CE (`target="onehot"`) or MSE (`target="embedding"`) loss for the
+        stage-2 model's `type_head` — the non-adversarial counterpart to
+        WGANStageTrainer's ST-Gumbel-into-the-critic path.
+        Zero when this stage has no `type_head` (stage 1, or
+        `particle_type.target = "physical"`)."""
+        l_type = torch.zeros((), device=device)
+        type_acc = torch.zeros((), device=device)
+        type_head = getattr(self.model, "type_head", None)
+        if not self.is_stage2 or type_head is None:
+            return l_type, type_acc
+        if self.decoder == "autoregressive":
+            assert ar_inputs is not None
+            type_out = self.model.predict_type(
+                cond_cont,
+                cond_cat,
+                stage1_ctx,
+                ar_inputs["history_feat"],
+                ar_inputs["has_prev"],
+                ar_inputs["remaining_frac"],
+                ar_inputs["slot_idx"],
+            )
+        else:
+            type_out = self.model.predict_type(cond_cont, cond_cat, stage1_ctx)
+        mask = sec_mask.float()
+        denom = mask.sum().clamp(min=1)
+        if self.particle_type_cfg.target == "onehot":
+            ce = F.cross_entropy(type_out.transpose(1, 2), sec_type_idx, reduction="none")
+            l_type = (ce * mask).sum() / denom
+            type_acc = ((type_out.argmax(-1) == sec_type_idx).float() * mask).sum() / denom
+        else:  # "embedding"
+            target_vec = self.model.cond_enc.pdg_emb(sec_type_idx).detach()
+            se = ((type_out - target_vec) ** 2).mean(-1)
+            l_type = (se * mask).sum() / denom
+        return l_type, type_acc
+
+    def _compute(self, batch: StepBatch, device: torch.device, epoch: int | None = None) -> dict:
+        """`epoch=None` (the `val_loss` path) always uses full teacher
+        forcing (`p_tf=1.0`) regardless of `spec.teacher_forcing` — validation
+        should stay a stable, non-stochastic ground-truth comparison; only
+        the training `step` path schedules `p_tf` by epoch."""
+        (
+            cond_cont,
+            cond_cat,
+            x1_s1,
+            n_sec,
+            sec_cont,
+            proc_idx,
+            sec_type_idx,
+        ) = _batch_to_device(batch, device)
+        sec_mask = self._sec_mask(n_sec, sec_cont.size(1), device)
+        stage1_ctx = x1_s1.detach()
+
+        x1_s2 = None
+        ar_inputs = None
+        if self.is_stage2 and self.decoder == "autoregressive":
+            ar_inputs = self._ar_inputs(cond_cont, cond_cat, stage1_ctx, sec_cont, sec_type_idx, n_sec, epoch)
+            x1_s2 = self._sec_target(sec_cont, sec_type_idx, self.generator, flatten=False)
+        elif self.is_stage2:
+            x1_s2 = self._sec_target(sec_cont, sec_type_idx, self.generator, flatten=True)
+
+        l_gen = self._generator_loss(cond_cont, cond_cat, x1_s1, x1_s2, sec_mask, stage1_ctx, ar_inputs=ar_inputs)
+        l_nsec, nsec_acc = self._n_sec_loss(cond_cont, cond_cat, stage1_ctx, n_sec, device)
+
+        l_type, type_acc = self._type_loss(
+            cond_cont,
+            cond_cat,
+            stage1_ctx,
+            sec_type_idx,
+            sec_mask,
+            device,
+            ar_inputs=ar_inputs,
+        )
+
+        l_balance = l_proc = l_entropy = torch.zeros((), device=device)
+        if self.router is not None:
+            if self.spec.lambda_balance > 0:
+                l_balance = self.router.balance_loss(cond_cont, cond_cat)
+            if self.spec.lambda_proc > 0:
+                l_proc = self.router.classify_loss(cond_cont, cond_cat, proc_idx)
+            if self.spec.lambda_entropy > 0:
+                l_entropy = self.router.entropy_loss(cond_cont, cond_cat)
+
+        total = self.spec.lambda_weight * l_gen + self.spec.n_sec_lambda * l_nsec + self.particle_type_lambda * l_type
+        if self.spec.lambda_balance > 0:
+            total = total + self.spec.lambda_balance * l_balance
+        if self.spec.lambda_proc > 0:
+            total = total + self.spec.lambda_proc * l_proc
+        if self.spec.lambda_entropy > 0:
+            total = total + self.spec.lambda_entropy * l_entropy
+
+        return {
+            "loss": total,
+            "loss_gen": l_gen,
+            "loss_nsec": l_nsec,
+            "loss_type": l_type,
+            "type_acc": type_acc,
+            "loss_balance": l_balance,
+            "loss_proc": l_proc,
+            "loss_entropy": l_entropy,
+            "nsec_acc": nsec_acc,
+        }
+
+    def step(self, batch: StepBatch, device: torch.device, global_step: int) -> dict:
+        if self.router is not None:
+            self.router.gumbel_tau = _gumbel_tau(
+                global_step,
+                self.total_steps,
+                self.spec.gumbel_tau_start,
+                self.spec.gumbel_tau_end,
+            )
+        epoch = global_step // self.spec.steps_per_epoch
+        out = self._compute(batch, device, epoch=epoch)
+        grad_norm = self._step_optimizer(self.optimizer, out["loss"], self.params)
+        self.lr_sched.step()
+        if self.ema_model is not None:
+            _update_ema(self.ema_model, self.model, self.ema_decay)
+        stats = {key: value.item() for key, value in out.items()}
+        stats["grad_norm"] = grad_norm
+        stats["lr"] = self.optimizer.param_groups[0]["lr"]
+        return stats
+
+    @torch.no_grad()
+    def val_loss(self, batch: StepBatch, device: torch.device) -> dict:
+        return {key: value.item() for key, value in self._compute(batch, device).items()}
+
+    # --- reporting ------------------------------------------------------
+
+    def batch_loss(self, stats: dict) -> float:
+        return stats["loss"]
+
+    def summary(self, means: dict) -> str:
+        return f"{self.name}[loss={means.get('loss', 0.0):.3f}]"
+
+    def val_objective(self, train_means: dict, val_means: dict, marginal_kl: float) -> float:
+        return val_means.get("loss", 0.0)
+
+
+class _Stage2RealFakeBatch(NamedTuple):
+    """Subset of `StepBatch` that `_stage2_real_and_fake` needs."""
+
+    cond_cont: torch.Tensor
+    cond_cat: torch.Tensor
+    n_sec: torch.Tensor
+    sec_cont: torch.Tensor
+    sec_type_idx: torch.Tensor
+
+
+class WGANStageTrainer(StageTrainer):
+    """WGAN-GP generator+critic for a single stage (see giant/model/wgan.py).
+
+    Ports `_wgan_train_step` to operate on one stage instead of two fused
+    together — the critic updates every batch; every `n_critic`-th batch
+    additionally updates the generator (`did_g_step`). The (non-adversarial)
+    n_sec classifier, when this stage's model owns it, updates every batch
+    regardless — folded into whichever generator optimizer step happens this
+    batch, same precedent as v0.2.
+    """
+
+    supports_val_loss = False
+
+    def __init__(
+        self,
+        spec: StageSpec,
+        model: torch.nn.Module,
+        critic: torch.nn.Module,
+        device: torch.device,
+    ) -> None:
+        self.critic = critic.to(device)
+        super().__init__(spec, model, device, extra_modules=(self.critic,))
+        self.n_critic = max(spec.n_critic, 1)
+        self.gp_weight = spec.gp_weight
+        self.critic_lr = spec.critic_lr
+
+        self.g_params = list(self.model.parameters())
+        self.d_params = list(self.critic.parameters())
+        # WGAN-GP recipe (Gulrajani et al. 2017): Adam, beta1=0, no weight decay.
+        self.optimizer = optim.Adam(self.g_params, lr=spec.lr, betas=(0.0, 0.9))
+        self.optimizer_d = optim.Adam(
+            self.d_params,
+            lr=spec.critic_lr if spec.critic_lr > 0 else spec.lr,
+            betas=(0.0, 0.9),
+        )
+
+        # Generator steps fire every n_critic-th batch, so warmup/decay must
+        # be counted in those units, matching v0.2.
+        gen_steps_per_epoch = max(spec.steps_per_epoch // self.n_critic, 1)
+        self._init_lr_schedule(
+            self.optimizer,
+            warmup_steps=spec.warmup_epochs * gen_steps_per_epoch,
+            total_steps=max(spec.epochs * gen_steps_per_epoch, 1),
+        )
+
+        train_keys = [
+            "d_loss",
+            "g_loss",
+            "wasserstein",
+            "gp_loss",
+            "loss_nsec",
+            "nsec_acc",
+            "grad_norm_d",
+            "grad_norm_g",
+        ]
+        if self.is_stage2 and self.particle_type_cfg.target == "onehot":
+            # Differentiability instrumentation — only meaningful when the
+            # type slice is a straight-through Gumbel relaxation.
+            train_keys += ["grad_norm_type_slice", "grad_norm_cont_slice"]
+        self.train_metrics = [train_metric(key) for key in train_keys]
+        self.val_metrics = []
+        self.stage_metrics = [stage_metric("lr"), stage_metric("critic_lr")]
+
+    def _stage2_real_and_fake(self, batch_tensors: _Stage2RealFakeBatch, stage1_ctx, global_step, device):
+        """Build `(real, fake_raw, mask, critic_fn)` for stage 2, covering
+        both decoders and all three particle-type targets. `fake_raw` still
+        needs the caller's straight-through relaxation under
+        `particle_type.target = "onehot"`, and neither tensor is masked-and-
+        multiplied on the fake side yet."""
+        cond_cont, cond_cat, n_sec, sec_cont, sec_type_idx = batch_tensors
+        B = cond_cont.size(0)
+        type_dim = stage2_type_dim(self.particle_type_cfg, self.particle_type_n_classes)
+        slot_width = CONT_SLOT_DIM + type_dim
+        k_max = sec_cont.size(1)
+
+        sec_mask = self._sec_mask(n_sec, k_max, device)
+        mask = sec_mask.unsqueeze(-1).expand(-1, -1, slot_width).reshape(B, -1).float()
+
+        def critic_fn(x):
+            return self.critic(x, cond_cont, cond_cat, stage1_ctx)
+
+        if self.decoder == "autoregressive":
+            epoch = global_step // self.spec.steps_per_epoch
+            ar = self._ar_inputs(cond_cont, cond_cat, stage1_ctx, sec_cont, sec_type_idx, n_sec, epoch)
+            real = self._sec_target(sec_cont, sec_type_idx, self.generator, flatten=False).reshape(B, -1) * mask
+            z = torch.randn(B, k_max, self.model.noise_dim, device=device)
+            fake_raw = self.model(
+                z,
+                cond_cont,
+                cond_cat,
+                stage1_ctx,
+                ar["history_feat"],
+                ar["has_prev"],
+                ar["remaining_frac"],
+                ar["slot_idx"],
+            ).reshape(B, -1)
+        else:
+            real = self._sec_target(sec_cont, sec_type_idx, self.generator, flatten=True) * mask
+            z = torch.randn(B, self.model.noise_dim, device=device)
+            fake_raw = self.model(z, cond_cont, cond_cat, stage1_ctx)
+
+        return real, fake_raw, mask, critic_fn
+
+    def step(self, batch: StepBatch, device: torch.device, global_step: int) -> dict:
+        (
+            cond_cont,
+            cond_cat,
+            x1_s1,
+            n_sec,
+            sec_cont,
+            _proc_idx,
+            sec_type_idx,
+        ) = _batch_to_device(batch, device)
+        B = cond_cont.size(0)
+        stage1_ctx = x1_s1.detach()
+        grad_probe: dict[str, float] = {}
+
+        if not self.is_stage2:
+            real = x1_s1
+
+            def critic_fn(x):
+                return self.critic(x, cond_cont, cond_cat)
+
+            z = torch.randn(B, self.model.noise_dim, device=device)
+            fake = self.model(z, cond_cont, cond_cat)
+            mask = None
+        else:
+            real, fake_raw, mask, critic_fn = self._stage2_real_and_fake(
+                _Stage2RealFakeBatch(cond_cont, cond_cat, n_sec, sec_cont, sec_type_idx),
+                stage1_ctx,
+                global_step,
+                device,
+            )
+            if self.particle_type_cfg.target == "onehot":
+                # Straight-through Gumbel-softmax relaxation of the type
+                # slice only — the critic must see a hard one-hot forward
+                # (matching what "real" data looks like) while gradient
+                # still flows smoothly to the generator. grad_probe captures
+                # the gradient-magnitude instrumentation — see
+                # _relax_onehot_type_slice's docstring.
+                tau = _gumbel_tau(
+                    global_step,
+                    self.total_steps,
+                    self.spec.type_gumbel_tau_start,
+                    self.spec.type_gumbel_tau_end,
+                )
+                fake_raw = _relax_onehot_type_slice(
+                    fake_raw,
+                    sec_cont.size(1),
+                    CONT_SLOT_DIM,
+                    stage2_type_dim(self.particle_type_cfg, self.particle_type_n_classes),
+                    tau,
+                    grad_probe=grad_probe,
+                )
+            fake = fake_raw * mask
+
+        # --- critic step (every batch) ---
+        fake_detached = fake.detach()
+        real_score = critic_fn(real)
+        fake_score = critic_fn(fake_detached)
+        gp = gradient_penalty(critic_fn, real, fake_detached, mask=mask)
+        d_loss = fake_score.mean() - real_score.mean() + self.gp_weight * gp
+        wasserstein = (real_score.mean() - fake_score.mean()).detach()
+
+        grad_norm_d = self._step_optimizer(self.optimizer_d, d_loss, self.d_params)
+
+        # --- generator (+ n_sec) step ---
+        did_g_step = global_step % self.n_critic == 0
+        l_nsec, nsec_acc = self._n_sec_loss(cond_cont, cond_cat, stage1_ctx, n_sec, device)
+
+        # On a non-generator-step batch with no n_sec_head on this stage
+        # (n_sec now defaults to stage 2), there's nothing for
+        # the generator optimizer to do this batch — g_loss would otherwise
+        # be a graph-less zero tensor, which .backward() rejects outright.
+        skip_g_step = not did_g_step and self.model.n_sec_head is None
+        if did_g_step:
+            g_loss_adv = generator_loss(critic_fn, fake)
+            g_loss = self.spec.lambda_weight * g_loss_adv + self.spec.n_sec_lambda * l_nsec
+        else:
+            g_loss_adv = torch.zeros((), device=device)
+            g_loss = self.spec.n_sec_lambda * l_nsec
+        if skip_g_step:
+            grad_norm_g = 0.0
+        else:
+            grad_norm_g = self._step_optimizer(self.optimizer, g_loss, self.g_params)
+
+        if did_g_step:
+            self.lr_sched.step()
+            if self.ema_model is not None:
+                _update_ema(self.ema_model, self.model, self.ema_decay)
+
+        return {
+            "d_loss": d_loss.item(),
+            "g_loss": g_loss_adv.item(),
+            "wasserstein": wasserstein.item(),
+            "gp_loss": gp.item(),
+            "loss_nsec": l_nsec.item(),
+            "nsec_acc": nsec_acc.item(),
+            "did_g_step": did_g_step,
+            "grad_norm": grad_norm_d + grad_norm_g,
+            "grad_norm_d": grad_norm_d,
+            "grad_norm_g": grad_norm_g,
+            "grad_norm_type_slice": grad_probe.get("type", 0.0),
+            "grad_norm_cont_slice": grad_probe.get("cont", 0.0),
+            "lr": self.optimizer.param_groups[0]["lr"],
+            "critic_lr": self.optimizer_d.param_groups[0]["lr"],
+        }
+
+    # --- reporting ------------------------------------------------------
+
+    def batch_loss(self, stats: dict) -> float:
+        return stats["d_loss"] + stats["g_loss"]
+
+    def summary(self, means: dict) -> str:
+        return f"{self.name}[d={means.get('d_loss', 0.0):.3f} g={means.get('g_loss', 0.0):.3f}]"
+
+    def val_objective(self, train_means: dict, val_means: dict, marginal_kl: float) -> float:
+        """No monotone per-batch WGAN loss fit for averaging, so
+        best-checkpoint selection uses the real marginal-KL signal when
+        `validate_marginals` produced one, and falls back to this epoch's own
+        Wasserstein-distance magnitude otherwise.
+
+        Behavior change vs. every run up to v0.3.0: the pre-refactor code
+        meant to do exactly this, but its guard
+        (`{n: kl for n in wgan_names if n not in val_loss_per_stage}`) could
+        never fire — `val_loss_per_stage` was pre-seeded with `0.0` for every
+        stage, so a WGAN stage contributed a flat `0.0` and the marginal KL
+        was recorded in the metrics row without ever influencing `best.pt`.
+        Runs from before this commit therefore selected their best checkpoint
+        on the non-adversarial stages alone."""
+        if math.isfinite(marginal_kl):
+            return marginal_kl
+        return abs(train_means.get("wasserstein", 0.0))
+
+    # --- state ----------------------------------------------------------
+
+    def _extra_state(self) -> dict:
+        return {
+            "critic": self.critic.state_dict(),
+            "optimizer_d": self.optimizer_d.state_dict(),
+        }
+
+    def _load_extra_state(self, sd: dict) -> None:
+        self.critic.load_state_dict(sd["critic"])
+        self.optimizer_d.load_state_dict(sd["optimizer_d"])
+
+    def _resume_extra_lr(self, lr: float) -> None:
+        resumed_critic_lr = self.critic_lr if self.critic_lr > 0 else lr
+        for group in self.optimizer_d.param_groups:
+            group["lr"] = resumed_critic_lr
+
+
+def build_stage_trainers(
+    cfg: dict,
+    models: dict[str, torch.nn.Module | None],
+    critics: dict[str, torch.nn.Module | None],
+    device: torch.device,
+    total_train_batches: int,
+) -> dict[str, StageTrainer]:
+    """One trainer per active stage — `models[name] is None` means that stage
+    is `active = false` and is simply never constructed."""
+    trainers: dict[str, StageTrainer] = {}
+    for name, is_stage2 in (("stage1", False), ("stage2", True)):
+        model = models.get(name)
+        if model is None:
+            continue
+        spec = StageSpec.from_config(cfg, name, is_stage2, max(total_train_batches, 1))
+        if build_objective(spec.generator).is_adversarial:
+            critic = critics.get(name)
+            assert critic is not None, (
+                f"{name}_model.generator='wgan' requires a critic (see giant.model.network.build_critics)"
+            )
+            trainers[name] = WGANStageTrainer(spec, model, critic, device)
+        else:
+            trainers[name] = FlowDDPMStageTrainer(spec, model, device)
+    return trainers

giant/validate.py

@@ -2,26 +2,13 @@ import numpy as np
 import torch
 from torch.utils.data import DataLoader
 
-from giant.constants import K_MAX, LOCAL_TARGET_NAMES
-from giant.sample import (
-    sample_flow,
-    sample_ddpm,
-    sample_ddim,
-    sample_secondaries,
-    sample_wgan,
-    sample_secondaries_wgan,
-)
+from giant.constants import LOCAL_TARGET_NAMES
+from giant.sample import resolve_n_sec, sample_stage1, sample_stage2
 
 _SEC_PHYS_NAMES = ["log_mass", "charge"]
 
 
-def _kw(steps: int | None) -> dict[str, int]:
-    return {} if steps is None else {"steps": steps}
-
-
-def _histogram_kl(
-    p_samples: np.ndarray, q_samples: np.ndarray, bins: int = 50, eps: float = 1e-8
-) -> float:
+def _histogram_kl(p_samples: np.ndarray, q_samples: np.ndarray, bins: int = 50, eps: float = 1e-8) -> float:
     """KL(P || Q) between two 1D samples, estimated via a shared histogram."""
     lo = min(p_samples.min(), q_samples.min())
     hi = max(p_samples.max(), q_samples.max())
@@ -44,15 +31,40 @@ def _bincount_frac(x: np.ndarray, minlength: int) -> np.ndarray:
     return counts / total if total > 0 else counts
 
 
+def _categorical_kl(real_idx: np.ndarray, gen_idx: np.ndarray, n_classes: int, eps: float = 1e-8) -> float:
+    """KL(P_real || Q_gen) between two class-index samples over `n_classes`
+    categories, estimated from bincount fractions. NaN if either side has no
+    valid samples (mirrors `_histogram_kl`'s empty-input handling)."""
+    if len(real_idx) == 0 or len(gen_idx) == 0:
+        return float("nan")
+    p = _bincount_frac(real_idx, n_classes) + eps
+    q = _bincount_frac(gen_idx, n_classes) + eps
+    p /= p.sum()
+    q /= q.sum()
+    return float(np.sum(p * np.log(p / q)))
+
+
+def _embedding_nearest_class(vectors: torch.Tensor, emb_weight: torch.Tensor) -> np.ndarray:
+    """Nearest row index (L1) of `vectors` (..., emb_dim) against `emb_weight`
+    (vocab, emb_dim) — same computation as
+    `giant.particles.decode_embedding_nearest`, but returning the raw class
+    index instead of a decoded PDG code: validate.py only needs a
+    real-vs-generated class-distribution comparison, not a rollout-usable
+    identity, so there's no need for the pdg_map inversion here."""
+    flat = vectors.reshape(-1, vectors.size(-1))
+    dist = (flat.unsqueeze(1) - emb_weight.detach().unsqueeze(0)).abs().sum(-1)
+    nearest = dist.argmin(dim=1)
+    return nearest.reshape(vectors.shape[:-1]).cpu().numpy()
+
+
 def validate_marginals(
-    model: torch.nn.Module,
+    stage1_model: torch.nn.Module,
     val_loader: DataLoader,
-    mode: str = "flow",
-    schedule=None,
     device: torch.device | None = None,
     n_batches: int | None = None,
     kl_bins: int = 50,
-    steps: int | None = None,
+    steps: int = 10,
+    ddpm_steps: int = 1000,
     sec_decoder: torch.nn.Module | None = None,
 ) -> dict[str, np.ndarray | float]:
     """Compare per-dimension marginals of generated vs. real steps.
@@ -61,51 +73,54 @@ def validate_marginals(
     normalised space. `kl_divergence[j]` is KL(real || generated) for
     dimension j, estimated from a shared histogram over both samples.
 
-    `steps` overrides the number of sampler steps (flow ODE steps or DDIM
-    substeps); `None` keeps each sampler's own default. Unused in "ddpm"
-    mode, which always runs the full schedule.
+    Stage 1 is sampled via `giant.sample.sample_stage1`, which dispatches on
+    `stage1_model.generator_kind` — `steps`/`ddpm_steps` are forwarded but
+    only one of them is actually read, depending on that dispatch.
 
-    When `sec_decoder` is given, also validates Stage 2: n_sec distribution
-    (+ classification accuracy), predicted secondary physical-identity
-    (log_mass, charge) marginals, and per-slot energy-fraction marginals —
-    restricted to each side's own valid slots (real: `n_sec`; generated: the
-    Stage-1 head's argmax), since the two need not agree on how many slots
-    are valid. Compared directly in normalised space (no denormalising —
-    KL estimated from a shared per-sample histogram is invariant to a shared
-    affine rescaling of both sides). Adds {"n_sec_real", "n_sec_pred",
-    "n_sec_accuracy", "phys_real", "phys_generated", "phys_kl",
-    "energy_fraction_kl"} to the returned dict.
+    When `sec_decoder` is given, also validates Stage 2 via
+    `giant.sample.sample_stage2`/`resolve_n_sec` (generator- and
+    one-shot-vs-autoregressive-agnostic): n_sec
+    distribution (+ classification accuracy), per-slot energy-fraction
+    marginals, and a particle-type marginal whose shape depends on
+    `sec_decoder.particle_type_cfg.target` — restricted to each side's own
+    valid slots (real: `n_sec`; generated: the resolved `n_sec_pred`), since
+    the two need not agree on how many slots are valid. Adds {"n_sec_real",
+    "n_sec_pred", "n_sec_accuracy", "energy_fraction_kl"} plus, under
+    `target = "physical"`, {"phys_real", "phys_generated", "phys_kl"}
+    (continuous log_mass/charge marginals — v0.2 behaviour), or under
+    `target` in `("onehot", "embedding")`, {"type_class_real",
+    "type_class_gen", "type_class_kl"} (categorical class-index marginal:
+    argmax for "onehot", L1-nearest conditioning-embedding row for
+    "embedding" — see `_embedding_nearest_class`). Compared directly in
+    normalised space (no denormalising — KL estimated from a shared
+    per-sample histogram/bincount is invariant to a shared affine rescaling
+    of both sides).
     """
     if device is None:
-        device = next(model.parameters()).device
-    model.eval()
+        device = next(stage1_model.parameters()).device
+    stage1_model.eval()
     if sec_decoder is not None:
         sec_decoder.eval()
 
+    k_max = sec_decoder.k_max if sec_decoder is not None else 0
+    target = sec_decoder.particle_type_cfg.target if sec_decoder is not None else "physical"
+
     all_real, all_gen = [], []
     all_n_sec_real, all_n_sec_pred = [], []
     all_phys_real, all_phys_gen = [], []
-    all_frac_real: list[list[np.ndarray]] = [[] for _ in range(K_MAX)]
-    all_frac_gen: list[list[np.ndarray]] = [[] for _ in range(K_MAX)]
+    all_type_class_real, all_type_class_gen = [], []
+    all_frac_real: list[list[np.ndarray]] = [[] for _ in range(k_max)]
+    all_frac_gen: list[list[np.ndarray]] = [[] for _ in range(k_max)]
 
     for i, batch in enumerate(val_loader):
         if n_batches is not None and i >= n_batches:
             break
-        # Batch is (cond_cont, cond_cat, target_s1, n_sec, sec_cont, proc_idx).
-        cond_cont, cond_cat, x1, n_sec, sec_cont, _proc_idx = batch
-        cond_cont = cond_cont.to(device)
-        cond_cat = cond_cat.to(device)
+        # batch is a StepBatch (giant.data.dataset).
+        x1, n_sec, sec_cont, sec_type_idx = batch.target_s1, batch.n_sec, batch.sec_cont, batch.sec_type_idx
+        cond_cont = batch.cond_cont.to(device)
+        cond_cat = batch.cond_cat.to(device)
 
-        if mode == "flow":
-            gen, n_sec_pred = sample_flow(model, cond_cont, cond_cat, **_kw(steps))
-        elif mode == "ddpm":
-            gen, n_sec_pred = sample_ddpm(model, cond_cont, cond_cat, schedule)
-        elif mode == "wgan":
-            gen, n_sec_pred = sample_wgan(model, cond_cont, cond_cat)
-        else:
-            gen, n_sec_pred = sample_ddim(
-                model, cond_cont, cond_cat, schedule, **_kw(steps)
-            )
+        gen, n_sec_pred = sample_stage1(stage1_model, cond_cont, cond_cat, steps=steps, ddpm_steps=ddpm_steps)
 
         all_real.append(x1.numpy())
         all_gen.append(gen.cpu().numpy())
@@ -113,54 +128,46 @@ def validate_marginals(
         if sec_decoder is None:
             continue
 
+        n_sec_pred = resolve_n_sec(stage1_model, sec_decoder, cond_cont, cond_cat, gen, n_sec_pred)
         n_sec_pred_np = n_sec_pred.cpu().numpy()
         n_sec_np = n_sec.numpy()
         all_n_sec_real.append(n_sec_np)
         all_n_sec_pred.append(n_sec_pred_np)
 
-        real_valid = np.arange(K_MAX)[None, :] < n_sec_np[:, None]  # (B, K_MAX)
+        real_valid = np.arange(k_max)[None, :] < n_sec_np[:, None]  # (B, k_max)
         real_frac = 1.0 / (1.0 + np.exp(-sec_cont[:, :, 0].numpy().astype(np.float64)))
-        real_phys = sec_cont[:, :, 4:6].numpy()  # (B, K_MAX, 2) [log_mass, charge]
 
-        if mode == "wgan":
-            sec_cont_pred, sec_phys_pred, sec_valid_pred = sample_secondaries_wgan(
-                sec_decoder, cond_cont, cond_cat, gen, n_sec_pred
-            )
-        else:
-            sec_cont_pred, sec_phys_pred, sec_valid_pred = sample_secondaries(
-                sec_decoder,
-                cond_cont,
-                cond_cat,
-                gen,
-                n_sec_pred,
-                steps=steps if steps is not None else 10,
-            )
-        gen_frac = 1.0 / (
-            1.0 + np.exp(-sec_cont_pred[:, :, 0].cpu().numpy().astype(np.float64))
+        sec_cont_pred, sec_type_pred, sec_valid_pred = sample_stage2(
+            sec_decoder, cond_cont, cond_cat, gen, n_sec_pred, steps=steps
         )
-        gen_phys = sec_phys_pred.cpu().numpy()
+        gen_frac = 1.0 / (1.0 + np.exp(-sec_cont_pred[:, :, 0].cpu().numpy().astype(np.float64)))
         gen_valid = sec_valid_pred.cpu().numpy()
 
-        all_phys_real.append(real_phys[real_valid])
-        all_phys_gen.append(gen_phys[gen_valid])
-        for j in range(K_MAX):
+        if target == "physical":
+            real_phys = sec_cont[:, :, 4:6].numpy()  # (B, k_max, 2) [log_mass, charge]
+            gen_phys = sec_type_pred.cpu().numpy()
+            all_phys_real.append(real_phys[real_valid])
+            all_phys_gen.append(gen_phys[gen_valid])
+        else:
+            sec_type_idx_np = sec_type_idx.numpy()
+            all_type_class_real.append(sec_type_idx_np[real_valid])
+            if target == "onehot":
+                gen_class = sec_type_pred.argmax(dim=-1).cpu().numpy()
+            else:  # "embedding"
+                emb_weight = sec_decoder.cond_enc.pdg_emb.weight
+                gen_class = _embedding_nearest_class(sec_type_pred, emb_weight)
+            all_type_class_gen.append(gen_class[gen_valid])
+
+        for j in range(k_max):
             all_frac_real[j].append(real_frac[real_valid[:, j], j])
             all_frac_gen[j].append(gen_frac[gen_valid[:, j], j])
 
     real = np.concatenate(all_real, axis=0)
     generated = np.concatenate(all_gen, axis=0)
 
-    kl_divergence = np.array(
-        [
-            _histogram_kl(real[:, j], generated[:, j], bins=kl_bins)
-            for j in range(real.shape[1])
-        ]
-    )
+    kl_divergence = np.array([_histogram_kl(real[:, j], generated[:, j], bins=kl_bins) for j in range(real.shape[1])])
 
-    header = (
-        f"{'Dim':<20} {'real_mean':>10} {'gen_mean':>10} "
-        f"{'real_std':>10} {'gen_std':>10} {'KL(real||gen)':>14}"
-    )
+    header = f"{'Dim':<20} {'real_mean':>10} {'gen_mean':>10} {'real_std':>10} {'gen_std':>10} {'KL(real||gen)':>14}"
     print(f"\n{header}")
     print("-" * len(header))
     for j, name in enumerate(LOCAL_TARGET_NAMES):
@@ -181,24 +188,9 @@ def validate_marginals(
     n_sec_real = np.concatenate(all_n_sec_real, axis=0)
     n_sec_pred_all = np.concatenate(all_n_sec_pred, axis=0)
     n_sec_accuracy = float((n_sec_real == n_sec_pred_all).mean())
-    phys_real = np.concatenate(all_phys_real, axis=0)  # (M, 2)
-    phys_gen = np.concatenate(all_phys_gen, axis=0)  # (M, 2)
 
-    if len(phys_real) > 0 and len(phys_gen) > 0:
-        phys_kl = np.array(
-            [
-                _histogram_kl(phys_real[:, j], phys_gen[:, j], bins=kl_bins)
-                for j in range(2)
-            ]
-        )
-    else:
-        phys_kl = np.full(2, np.nan)
-
-    energy_fraction_kl = np.full(K_MAX, np.nan)
-    print(
-        f"\n{'n_sec':<20} accuracy={n_sec_accuracy:.4f}  "
-        f"mean|Δ|={np.abs(n_sec_real - n_sec_pred_all).mean():.4f}"
-    )
+    energy_fraction_kl = np.full(k_max, np.nan)
+    print(f"\n{'n_sec':<20} accuracy={n_sec_accuracy:.4f}  mean|Δ|={np.abs(n_sec_real - n_sec_pred_all).mean():.4f}")
     n_sec_dist_header = f"{'n_sec value':<20} {'real_frac':>10} {'gen_frac':>10}"
     print(n_sec_dist_header)
     print("-" * len(n_sec_dist_header))
@@ -208,46 +200,70 @@ def validate_marginals(
     for v in range(max_n_sec):
         print(f"{v:<20} {real_n_sec_dist[v]:>10.4f} {gen_n_sec_dist[v]:>10.4f}")
 
-    print(
-        f"\n{'sec phys (normalised)':<20} {'real_mean':>10} {'gen_mean':>10} "
-        f"{'real_std':>10} {'gen_std':>10} {'KL(real||gen)':>14}"
-    )
-    print("-" * 68)
-    for j, name in enumerate(_SEC_PHYS_NAMES):
-        r, g = phys_real[:, j], phys_gen[:, j]
-        if len(r) == 0 or len(g) == 0:
-            continue
-        print(
-            f"{name:<20} {r.mean():>10.4f} {g.mean():>10.4f} "
-            f"{r.std():>10.4f} {g.std():>10.4f} {phys_kl[j]:>14.4f}"
-        )
-
     print(
         f"\n{'sec slot (energy frac.)':<24} {'real_mean':>10} {'gen_mean':>10} "
         f"{'real_std':>10} {'gen_std':>10} {'KL(real||gen)':>14}"
     )
     print("-" * 90)
-    for j in range(K_MAX):
+    for j in range(k_max):
         r = np.concatenate(all_frac_real[j]) if all_frac_real[j] else np.array([])
         g = np.concatenate(all_frac_gen[j]) if all_frac_gen[j] else np.array([])
         if len(r) == 0 or len(g) == 0:
             continue
         kl = _histogram_kl(r, g, bins=kl_bins)
         energy_fraction_kl[j] = kl
-        print(
-            f"{j:<24} {r.mean():>10.4f} {g.mean():>10.4f} "
-            f"{r.std():>10.4f} {g.std():>10.4f} {kl:>14.4f}"
-        )
+        print(f"{j:<24} {r.mean():>10.4f} {g.mean():>10.4f} {r.std():>10.4f} {g.std():>10.4f} {kl:>14.4f}")
 
     result.update(
         {
             "n_sec_real": n_sec_real,
             "n_sec_pred": n_sec_pred_all,
             "n_sec_accuracy": n_sec_accuracy,
-            "phys_real": phys_real,
-            "phys_generated": phys_gen,
-            "phys_kl": phys_kl,
             "energy_fraction_kl": energy_fraction_kl,
         }
     )
+
+    if target == "physical":
+        phys_real = np.concatenate(all_phys_real, axis=0)  # (M, 2)
+        phys_gen = np.concatenate(all_phys_gen, axis=0)  # (M, 2)
+
+        if len(phys_real) > 0 and len(phys_gen) > 0:
+            phys_kl = np.array([_histogram_kl(phys_real[:, j], phys_gen[:, j], bins=kl_bins) for j in range(2)])
+        else:
+            phys_kl = np.full(2, np.nan)
+
+        print(
+            f"\n{'sec phys (normalised)':<20} {'real_mean':>10} {'gen_mean':>10} "
+            f"{'real_std':>10} {'gen_std':>10} {'KL(real||gen)':>14}"
+        )
+        print("-" * 68)
+        for j, name in enumerate(_SEC_PHYS_NAMES):
+            r, g = phys_real[:, j], phys_gen[:, j]
+            if len(r) == 0 or len(g) == 0:
+                continue
+            print(
+                f"{name:<20} {r.mean():>10.4f} {g.mean():>10.4f} {r.std():>10.4f} {g.std():>10.4f} {phys_kl[j]:>14.4f}"
+            )
+
+        result.update({"phys_real": phys_real, "phys_generated": phys_gen, "phys_kl": phys_kl})
+    else:
+        type_class_real = np.concatenate(all_type_class_real, axis=0)
+        type_class_gen = np.concatenate(all_type_class_gen, axis=0)
+        n_classes = sec_decoder.type_dim if target == "onehot" else sec_decoder.cond_enc.pdg_emb.weight.size(0)
+        type_class_kl = _categorical_kl(type_class_real, type_class_gen, n_classes)
+
+        print(
+            f"\n{'sec type class (' + target + ')':<24} "
+            f"n={len(type_class_real)}/{len(type_class_gen)} "
+            f"KL(real||gen)={type_class_kl:.4f}"
+        )
+
+        result.update(
+            {
+                "type_class_real": type_class_real,
+                "type_class_gen": type_class_gen,
+                "type_class_kl": type_class_kl,
+            }
+        )
+
     return result

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "giant"
-version = "0.2.0"
+version = "0.3.1"
 description = "Geant4 step-function surrogate via conditional flow matching"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -23,6 +23,7 @@ cuda = [
 ]
 dev = [
     "pytest>=8,<10",
+    "pytest-cov>=5,<8",
     "ruff>=0.15,<1",
     "ty>=0.0.50,<0.1",
     "giant[convert,analysis,geometry,wandb]",
@@ -49,14 +50,27 @@ analysis = [
 
 [project.scripts]
 giant = "giant.cli:app"
-dwarf = "scripts.dwarf:app"
+dwarf = "giant.tools.dwarf:app"
+
+[tool.ruff]
+line-length = 120
+
+[tool.coverage.run]
+source = ["giant"]
+omit = ["*/legacy/*"]
+
+[tool.coverage.report]
+exclude_also = [
+    "if TYPE_CHECKING:",
+    "raise NotImplementedError",
+]
 
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]
-packages = ["giant", "scripts"]
+packages = ["giant"]
 
 [tool.uv]
 conflicts = [

tests/test_bump_dataset_version.py

@@ -1,7 +1,7 @@
 import os
 import subprocess
 
-from scripts import bump_dataset_version
+from giant.tools import bump_dataset_version
 
 plan_bump_gen = bump_dataset_version.plan_bump_gen
 plan_bump_schema = bump_dataset_version.plan_bump_schema
@@ -22,9 +22,7 @@ def test_git_user_name_returns_none_on_timeout(monkeypatch):
 
 
 def test_bump_gen_starts_at_gen1_when_none_exist(tmp_path):
-    dirs, log_line = plan_bump_gen(
-        tmp_path, "steps", "first generation", None, "2026-01-01"
-    )
+    dirs, log_line = plan_bump_gen(tmp_path, "steps", "first generation", None, "2026-01-01")
     assert dirs == [
         tmp_path / "raw" / "steps" / "gen1",
         tmp_path / "processed" / "steps" / "gen1" / "schema1",
@@ -56,9 +54,7 @@ def test_bump_gen_kinds_are_independent(tmp_path):
 
 def test_bump_schema_starts_at_schema1_for_a_fresh_gen(tmp_path):
     (tmp_path / "raw" / "steps" / "gen1").mkdir(parents=True)
-    dirs, log_line = plan_bump_schema(
-        tmp_path, "steps", "gen1", "added e_sec column", None, "2026-01-01"
-    )
+    dirs, log_line = plan_bump_schema(tmp_path, "steps", "gen1", "added e_sec column", None, "2026-01-01")
     assert dirs == [tmp_path / "processed" / "steps" / "gen1" / "schema1"]
     assert "`gen1`/`schema1`" in log_line
 
@@ -66,18 +62,14 @@ def test_bump_schema_starts_at_schema1_for_a_fresh_gen(tmp_path):
 def test_bump_schema_increments_within_its_gen(tmp_path):
     (tmp_path / "processed" / "steps" / "gen1" / "schema1").mkdir(parents=True)
     (tmp_path / "processed" / "steps" / "gen1" / "schema2").mkdir(parents=True)
-    dirs, _ = plan_bump_schema(
-        tmp_path, "steps", "gen1", "next schema", None, "2026-01-01"
-    )
+    dirs, _ = plan_bump_schema(tmp_path, "steps", "gen1", "next schema", None, "2026-01-01")
     assert dirs == [tmp_path / "processed" / "steps" / "gen1" / "schema3"]
 
 
 def test_bump_schema_does_not_see_other_gens_schemas(tmp_path):
     (tmp_path / "processed" / "steps" / "gen1" / "schema5").mkdir(parents=True)
     (tmp_path / "raw" / "steps" / "gen2").mkdir(parents=True)
-    dirs, _ = plan_bump_schema(
-        tmp_path, "steps", "gen2", "fresh schema for gen2", None, "2026-01-01"
-    )
+    dirs, _ = plan_bump_schema(tmp_path, "steps", "gen2", "fresh schema for gen2", None, "2026-01-01")
     assert dirs == [tmp_path / "processed" / "steps" / "gen2" / "schema1"]
 
 
@@ -91,9 +83,7 @@ def test_bump_schema_rejects_nonexistent_gen(tmp_path):
 
 def test_bump_gen_to_specific_tag(tmp_path):
     (tmp_path / "raw" / "steps" / "gen1").mkdir(parents=True)
-    dirs, log_line = plan_bump_gen(
-        tmp_path, "steps", "jump to gen5", None, "2026-01-01", target="gen5"
-    )
+    dirs, log_line = plan_bump_gen(tmp_path, "steps", "jump to gen5", None, "2026-01-01", target="gen5")
     assert dirs[0] == tmp_path / "raw" / "steps" / "gen5"
     assert "`gen5`" in log_line
 
@@ -125,9 +115,7 @@ def test_bump_schema_to_specific_tag(tmp_path):
 def test_bump_schema_rejects_invalid_to_tag(tmp_path):
     (tmp_path / "raw" / "steps" / "gen1").mkdir(parents=True)
     try:
-        plan_bump_schema(
-            tmp_path, "steps", "gen1", "bad tag", None, "2026-01-01", target="v3"
-        )
+        plan_bump_schema(tmp_path, "steps", "gen1", "bad tag", None, "2026-01-01", target="v3")
         assert False, "expected SystemExit"
     except SystemExit:
         pass
@@ -164,15 +152,7 @@ def _make_parquet(path):
 
 
 def test_update_manifest_bumps_to_specified_schema(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -194,15 +174,7 @@ def test_update_manifest_auto_detects_highest_schema(tmp_path):
     for schema in ("schema1", "schema2", "schema3"):
         d = tmp_path / "processed" / "steps" / "gen1" / schema / "pbwo4"
         d.mkdir(parents=True)
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema3"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen1" / "schema3" / "pbwo4" / "shard-000.parquet"
     parquet.touch()
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -231,15 +203,7 @@ def test_update_manifest_reports_missing_targets(tmp_path):
 
 
 def test_update_manifest_skips_already_at_target(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -254,15 +218,7 @@ def test_update_manifest_skips_already_at_target(tmp_path):
 
 
 def test_update_manifest_preserves_comments_and_blanks(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -278,15 +234,7 @@ def test_update_manifest_preserves_comments_and_blanks(tmp_path):
 
 
 def test_update_manifest_bumps_gen(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen2"
-        / "schema1"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen2" / "schema1" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -303,15 +251,7 @@ def test_update_manifest_bumps_gen(tmp_path):
 
 
 def test_update_manifest_bumps_gen_and_schema(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen2"
-        / "schema3"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen2" / "schema3" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -328,15 +268,7 @@ def test_update_manifest_bumps_gen_and_schema(tmp_path):
 
 
 def test_apply_update_manifest_writes_file(tmp_path):
-    parquet = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    parquet = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-000.parquet"
     _make_parquet(parquet)
 
     manifest_dir = tmp_path / "pools" / "pbwo4"
@@ -358,24 +290,8 @@ def test_apply_update_manifest_writes_file(tmp_path):
 
 
 def test_create_manifest_writes_relative_paths(tmp_path):
-    pq1 = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
-    pq2 = (
-        tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema2"
-        / "pbwo4"
-        / "shard-001.parquet"
-    )
+    pq1 = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-000.parquet"
+    pq2 = tmp_path / "processed" / "steps" / "gen1" / "schema2" / "pbwo4" / "shard-001.parquet"
     _make_parquet(pq1)
     _make_parquet(pq2)
 
@@ -419,9 +335,7 @@ def test_run_create_manifest_refuses_to_overwrite_existing_output(tmp_path):
     output.write_text("original contents\n")
 
     try:
-        bump_dataset_version.run_create_manifest(
-            [str(pq)], execute=True, output=str(output)
-        )
+        bump_dataset_version.run_create_manifest([str(pq)], execute=True, output=str(output))
         assert False, "expected SystemExit"
     except SystemExit:
         pass
@@ -435,9 +349,7 @@ def test_run_create_manifest_force_overwrites_existing_output(tmp_path):
     output.parent.mkdir(parents=True)
     output.write_text("original contents\n")
 
-    bump_dataset_version.run_create_manifest(
-        [str(pq)], execute=True, output=str(output), force=True
-    )
+    bump_dataset_version.run_create_manifest([str(pq)], execute=True, output=str(output), force=True)
     assert output.read_text() != "original contents\n"
 
 

tests/test_catalog.py

@@ -13,9 +13,7 @@ from tests.test_analysis_reduce import _reference_frame, _rollout_frame
 
 def _build_ctx() -> Context:
     r, t = _rollout_frame(), _reference_frame()
-    return build_context(
-        r, t, n_energy_bins=2, n_marginal_bins=10, top_k_pdg=3, sample_rows=1000
-    )
+    return build_context(r, t, n_energy_bins=2, n_marginal_bins=10, top_k_pdg=3, sample_rows=1000)
 
 
 @pytest.fixture(scope="module")
@@ -142,10 +140,7 @@ def test_chunked_matches_unchunked(ctx: Context, spec_id: str):
 
     # 4 chunks over only 2 distinct event_ids also exercises empty chunks.
     n_chunks = 4 if spec.chunkable else 1
-    parts = [
-        spec.compute_partial(Bundle.open(r, t, ctx, chunk=(k, n_chunks)))
-        for k in range(n_chunks)
-    ]
+    parts = [spec.compute_partial(Bundle.open(r, t, ctx, chunk=(k, n_chunks))) for k in range(n_chunks)]
     chunked = spec.finalize(parts, ctx)
 
     assert chunked.id == unchunked.id

tests/test_checkpoint_io.py

@@ -0,0 +1,280 @@
+"""Tests for giant.checkpoint_io.load_for_inference (issues.md Issue 5) —
+the shared bootstrap `giant predict`/`giant rollout` use to go from a
+checkpoint path to ready-to-run models."""
+
+from __future__ import annotations
+
+import copy
+
+import numpy as np
+import pytest
+import torch
+
+from giant import config as gconfig
+from giant.checkpoint_io import (
+    CheckpointCompatibilityError,
+    InferenceContext,
+    conditioning_axes,
+    load_for_inference,
+    stage_cfg,
+)
+from giant.data.loader import TopNMap
+from giant.data.setup_cache import topnmap_to_json
+from giant.data.transforms import Normalizer
+from giant.model.network import build_models
+
+PDG_MAP = {11: 0, 22: 1, -11: 2}
+MAT_MAP = {"G4_PbWO4": 0, "G4_AIR": 1}
+
+
+def _model_cfg(stage2_active: bool = True) -> dict:
+    """DEFAULT_CONFIG-derived, shrunk for speed — same pattern as
+    tests/test_network.py::_minimal_model_config. Default `conditioning`
+    (both axes "physical") needs no top-N vocab map, so this is a cheap,
+    fully self-contained happy-path config."""
+    cfg = copy.deepcopy(gconfig.DEFAULT_CONFIG)
+    cfg["conditioning"]["particle"]["emb_dim"] = 4
+    cfg["conditioning"]["material"]["emb_dim"] = 4
+    cfg["stage1_model"].update({"hidden_dim": 8, "n_res_blocks": 1})
+    cfg["stage2_model"].update({"hidden_dim": 8, "n_res_blocks": 1, "k_max": 3})
+    cfg["stage2_model"]["active"] = stage2_active
+    return {
+        "pdg_vocab": len(PDG_MAP),
+        "mat_vocab": len(MAT_MAP),
+        "conditioning": cfg["conditioning"],
+        "stage1_model": cfg["stage1_model"],
+        "stage2_model": cfg["stage2_model"],
+    }
+
+
+def _norms() -> tuple[Normalizer, Normalizer, Normalizer]:
+    rng = np.random.default_rng(0)
+    cond = Normalizer().fit(rng.standard_normal((100, 15)).astype(np.float32))
+    tgt = Normalizer().fit(rng.standard_normal((100, 9)).astype(np.float32))
+    sec_phys = Normalizer().fit(rng.standard_normal((100, 2)).astype(np.float32))
+    return cond, tgt, sec_phys
+
+
+def _write_checkpoint(tmp_path, model_cfg=None, ema: bool = False, **ckpt_overrides):
+    cfg = model_cfg if model_cfg is not None else _model_cfg()
+    built = build_models(cfg)
+    stage1, stage2 = built["stage1"], built["stage2"]
+    cond, tgt, sec_phys = _norms()
+    ckpt: dict = {
+        "model_config": cfg,
+        "model": stage1.state_dict() if stage1 is not None else {},
+        "sec_decoder": stage2.state_dict() if stage2 is not None else {},
+        "pdg_map": PDG_MAP,
+        "mat_map": MAT_MAP,
+        "normalizer": {"cond": cond.to_dict(), "target": tgt.to_dict(), "sec_phys": sec_phys.to_dict()},
+        "epoch": 3,
+        "best_val_loss": 0.5,
+    }
+    if ema:
+        ckpt["model_ema"] = stage1.state_dict() if stage1 is not None else {}
+        ckpt["sec_decoder_ema"] = stage2.state_dict() if stage2 is not None else {}
+    # DEFAULT_CONFIG's stage2_model.particle_type.target defaults to
+    # "onehot", and giant train's pipeline (gitea #29) now always writes a
+    # sec_type_topn_map in that case — default one in here too, unless a
+    # test explicitly overrides it, so fixtures represent a real, loadable
+    # checkpoint by default rather than exercising the "missing" guard by
+    # accident.
+    particle_type_target = cfg.get("stage2_model", {}).get("particle_type", {}).get("target", "onehot")
+    if particle_type_target == "onehot" and "sec_type_topn_map" not in ckpt_overrides:
+        default_sec_type_topn = TopNMap(class_map=dict(zip(PDG_MAP, range(len(PDG_MAP)))), other_members={})
+        ckpt["sec_type_topn_map"] = topnmap_to_json(default_sec_type_topn)
+    ckpt.update(ckpt_overrides)
+    path = tmp_path / "ckpt.pt"
+    torch.save(ckpt, path)
+    return path
+
+
+def _onehot_model_cfg() -> dict:
+    cfg = _model_cfg()
+    cfg["conditioning"]["particle"]["type"] = "onehot"
+    return cfg
+
+
+# ---------------------------------------------------------------------------
+# Happy path
+# ---------------------------------------------------------------------------
+
+
+def test_happy_path_returns_populated_context(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path)
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+    assert isinstance(ctx, InferenceContext)
+    assert ctx.stage1 is not None and ctx.stage2 is not None
+    assert not ctx.stage1.training
+    assert not ctx.stage2.training
+    assert next(ctx.stage1.parameters()).device == torch.device("cpu")
+    assert ctx.pdg_map == PDG_MAP
+    assert ctx.mat_map == MAT_MAP
+    assert all(isinstance(k, int) for k in ctx.pdg_map)
+    assert all(isinstance(k, str) for k in ctx.mat_map)
+    assert ctx.particle_conditioning == "physical"
+    assert ctx.material_conditioning == "physical"
+    assert ctx.k_max == 3
+    assert ctx.epoch == 3
+    assert ctx.best_val_loss == 0.5
+    assert ctx.model_config["stage1_model"]["hidden_dim"] == 8
+
+
+def test_happy_path_normalizer_values_round_trip(tmp_path):
+    cond, tgt, sec_phys = _norms()
+    checkpoint = _write_checkpoint(tmp_path)
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+    assert ctx.cond_norm.mean is not None and cond.mean is not None
+    assert ctx.tgt_norm.mean is not None and tgt.mean is not None
+    assert ctx.sec_phys_norm.mean is not None and sec_phys.mean is not None
+    np.testing.assert_allclose(ctx.cond_norm.mean, cond.mean)
+    np.testing.assert_allclose(ctx.tgt_norm.mean, tgt.mean)
+    np.testing.assert_allclose(ctx.sec_phys_norm.mean, sec_phys.mean)
+
+
+# ---------------------------------------------------------------------------
+# Guards
+# ---------------------------------------------------------------------------
+
+
+def test_missing_model_config_raises(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path)
+    ckpt = torch.load(checkpoint, weights_only=False)
+    del ckpt["model_config"]
+    torch.save(ckpt, checkpoint)
+
+    with pytest.raises(CheckpointCompatibilityError, match="no model_config"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+
+def test_missing_sec_decoder_raises(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path)
+    ckpt = torch.load(checkpoint, weights_only=False)
+    del ckpt["sec_decoder"]
+    torch.save(ckpt, checkpoint)
+
+    with pytest.raises(CheckpointCompatibilityError, match="no sec_decoder"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+
+def test_missing_sec_phys_normalizer_raises(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path)
+    ckpt = torch.load(checkpoint, weights_only=False)
+    del ckpt["normalizer"]["sec_phys"]
+    torch.save(ckpt, checkpoint)
+
+    with pytest.raises(CheckpointCompatibilityError, match="no normalizer.sec_phys"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+
+def test_onehot_particle_conditioning_without_topn_map_raises(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path, model_cfg=_onehot_model_cfg())
+
+    with pytest.raises(CheckpointCompatibilityError, match="pdg_topn_map"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+
+def test_onehot_particle_conditioning_with_topn_map_succeeds(tmp_path):
+    topn = TopNMap(class_map={11: 0, 22: 1}, other_members={})
+    checkpoint = _write_checkpoint(
+        tmp_path,
+        model_cfg=_onehot_model_cfg(),
+        pdg_topn_map=topnmap_to_json(topn),
+    )
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict")
+    assert ctx.particle_conditioning == "onehot"
+    assert ctx.pdg_topn_map is not None
+    assert ctx.pdg_topn_map.class_map == {11: 0, 22: 1}
+
+
+def test_onehot_particle_type_target_without_sec_type_topn_map_raises(tmp_path):
+    """DEFAULT_CONFIG's stage2_model.particle_type.target="onehot" needs a
+    sec_type_topn_map (gitea #29) — a checkpoint with neither key at all
+    (not even the pre-#29 pdg_topn_map to fall back to) must fail loudly."""
+    checkpoint = _write_checkpoint(tmp_path, sec_type_topn_map=None)
+    ckpt = torch.load(checkpoint, weights_only=False)
+    del ckpt["sec_type_topn_map"]
+    torch.save(ckpt, checkpoint)
+
+    with pytest.raises(CheckpointCompatibilityError, match="sec_type_topn_map"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict")
+
+
+def test_pre_gitea_29_checkpoint_falls_back_to_pdg_topn_map_for_sec_type(tmp_path):
+    """A checkpoint written before gitea #29 has no sec_type_topn_map key at
+    all — conditioning and secondary-type onehot maps were always the same
+    map, saved once under pdg_topn_map. load_for_inference must reproduce
+    that exact pre-#29 behavior for such a checkpoint."""
+    topn = TopNMap(class_map={11: 0, 22: 1, -11: 2}, other_members={})
+    checkpoint = _write_checkpoint(
+        tmp_path,
+        model_cfg=_onehot_model_cfg(),
+        pdg_topn_map=topnmap_to_json(topn),
+        sec_type_topn_map=None,
+    )
+    ckpt = torch.load(checkpoint, weights_only=False)
+    del ckpt["sec_type_topn_map"]
+    torch.save(ckpt, checkpoint)
+
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict")
+    assert ctx.sec_type_topn_map is not None
+    assert ctx.sec_type_topn_map.class_map == {11: 0, 22: 1, -11: 2}
+
+
+def test_ema_weights_requested_but_missing_raises(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path, ema=False)
+
+    with pytest.raises(CheckpointCompatibilityError, match="no EMA weights"):
+        load_for_inference(checkpoint, torch.device("cpu"), "predict", weights="ema")
+
+
+def test_ema_weights_requested_and_present_succeeds(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path, ema=True)
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict", weights="ema")
+    assert ctx.stage1 is not None and ctx.stage2 is not None
+
+
+@pytest.mark.parametrize("command_name", ["predict", "rollout"])
+def test_inactive_stage_with_require_stage2_raises_with_command_name(tmp_path, command_name):
+    checkpoint = _write_checkpoint(tmp_path, model_cfg=_model_cfg(stage2_active=False))
+
+    with pytest.raises(CheckpointCompatibilityError, match=f"{command_name} needs both"):
+        load_for_inference(checkpoint, torch.device("cpu"), command_name)
+
+
+def test_inactive_stage_with_require_stage2_false_succeeds_with_stage2_none(tmp_path):
+    checkpoint = _write_checkpoint(tmp_path, model_cfg=_model_cfg(stage2_active=False))
+
+    ctx = load_for_inference(checkpoint, torch.device("cpu"), "predict", require_stage2=False)
+    assert ctx.stage1 is not None
+    assert ctx.stage2 is None
+
+
+# ---------------------------------------------------------------------------
+# conditioning_axes / stage_cfg
+# ---------------------------------------------------------------------------
+
+
+def test_conditioning_axes_v02_flat_string_applies_to_both_axes():
+    assert conditioning_axes({"conditioning": "embedding"}) == ("embedding", "embedding")
+
+
+def test_conditioning_axes_v03_nested_dict_independent_per_axis():
+    model_cfg = {"conditioning": {"particle": {"type": "onehot"}, "material": {"type": "physical"}}}
+    assert conditioning_axes(model_cfg) == ("onehot", "physical")
+
+
+def test_conditioning_axes_missing_key_uses_default():
+    assert conditioning_axes({}, default="embedding") == ("embedding", "embedding")
+
+
+def test_stage_cfg_new_shape_returns_subdict():
+    model_cfg = {"stage2_model": {"k_max": 7}}
+    assert stage_cfg(model_cfg, "stage2") == {"k_max": 7}
+
+
+def test_stage_cfg_v02_flat_shape_returns_empty_dict():
+    model_cfg = {"hidden_dim": 32, "n_blocks": 4}
+    assert stage_cfg(model_cfg, "stage2") == {}

tests/test_cli_new_run.py

@@ -37,13 +37,14 @@ def test_writes_config_with_overrides_applied(tmp_path: Path):
     with open(config_path, "rb") as f:
         cfg = tomllib.load(f)
 
-    assert cfg["train"]["mode"] == "ddpm"
+    assert cfg["stage1_model"]["generator"] == "ddpm"
+    assert cfg["stage2_model"]["generator"] == "ddpm"
     assert cfg["train"]["lr"] == 0.0005
-    assert cfg["model"]["hidden_dim"] == 128
-    assert cfg["model"]["n_blocks"] == 4
+    assert cfg["stage1_model"]["hidden_dim"] == 128
+    assert cfg["stage1_model"]["n_res_blocks"] == 4
     # untouched defaults still present
     assert cfg["train"]["epochs"] == 100
-    assert "router" in cfg["model"]
+    assert "router" in cfg["stage1_model"]
 
     assert str(out_dir) in result.output
     assert "<data.parquet>" in result.output
@@ -100,9 +101,7 @@ def test_force_guard_refuses_to_clobber_existing_checkpoints(tmp_path: Path):
     assert "already has last.pt" in result.output
     assert not (out_dir / "config.toml").exists()
 
-    result = runner.invoke(
-        app, ["new-run", "--out", str(out_dir), "--mode", "ddpm", "--force"]
-    )
+    result = runner.invoke(app, ["new-run", "--out", str(out_dir), "--mode", "ddpm", "--force"])
     assert result.exit_code == 0, result.output
     assert (out_dir / "config.toml").exists()
 

tests/test_cli_predict.py

@@ -1,13 +1,18 @@
 import uuid
 
+import torch
 import yaml
+from typer.testing import CliRunner
 
 from giant.cli import (
     _CEPH_PREDICTIONS,
     _resolve_prediction_output,
     _write_prediction_ref,
+    app,
 )
 
+runner = CliRunner()
+
 
 # ---------------------------------------------------------------------------
 # _resolve_prediction_output
@@ -116,9 +121,7 @@ def test_ref_yaml_includes_comment_when_provided(tmp_path):
     dataset = tmp_path / "full.manifest"
     pred_uuid = str(uuid.uuid4())
 
-    ref_path = _write_prediction_ref(
-        checkpoint, pred_uuid, out, dataset, comment="baseline sweep run 3"
-    )
+    ref_path = _write_prediction_ref(checkpoint, pred_uuid, out, dataset, comment="baseline sweep run 3")
     data = yaml.safe_load(ref_path.read_text())
 
     assert data["comment"] == "baseline sweep run 3"
@@ -133,9 +136,7 @@ def test_ref_timestamp_is_iso_format(tmp_path):
     checkpoint.touch()
 
     pred_uuid = str(uuid.uuid4())
-    ref_path = _write_prediction_ref(
-        checkpoint, pred_uuid, tmp_path / "p.parquet", tmp_path / "d"
-    )
+    ref_path = _write_prediction_ref(checkpoint, pred_uuid, tmp_path / "p.parquet", tmp_path / "d")
     data = yaml.safe_load(ref_path.read_text())
 
     # Must parse without error and be timezone-aware (UTC).
@@ -150,9 +151,24 @@ def test_ref_checkpoint_path_is_absolute(tmp_path):
     checkpoint.touch()
 
     pred_uuid = str(uuid.uuid4())
-    ref_path = _write_prediction_ref(
-        checkpoint, pred_uuid, tmp_path / "p.parquet", tmp_path / "d"
-    )
+    ref_path = _write_prediction_ref(checkpoint, pred_uuid, tmp_path / "p.parquet", tmp_path / "d")
     data = yaml.safe_load(ref_path.read_text())
 
     assert data["checkpoint"].startswith("/")
+
+
+# ---------------------------------------------------------------------------
+# Bootstrap failure surfaces via the CLI (issues.md Issue 5 — confirms
+# CheckpointCompatibilityError -> typer.Exit(1) actually wires up end-to-end,
+# not just at the giant.checkpoint_io unit level).
+# ---------------------------------------------------------------------------
+
+
+def test_predict_exits_1_on_checkpoint_missing_model_config(tmp_path):
+    checkpoint = tmp_path / "bad.pt"
+    torch.save({"sec_decoder": {}, "normalizer": {"sec_phys": {}}}, checkpoint)
+
+    result = runner.invoke(app, ["predict", "dummy.parquet", "--checkpoint", str(checkpoint)])
+
+    assert result.exit_code == 1
+    assert "checkpoint has no model_config" in result.output

tests/test_cli_rollout.py

@@ -0,0 +1,33 @@
+"""Thin CLI smoke coverage for `giant rollout` (issues.md Issue 5) — confirms
+the CheckpointCompatibilityError raised by giant.checkpoint_io.load_for_inference
+surfaces as a clean typer.Exit(1) with the expected message, end-to-end
+through the CLI, not just at the giant.checkpoint_io unit level."""
+
+from __future__ import annotations
+
+import torch
+from typer.testing import CliRunner
+
+from giant.cli import app
+
+runner = CliRunner()
+
+
+def test_rollout_exits_1_on_checkpoint_missing_model_config(tmp_path):
+    checkpoint = tmp_path / "bad.pt"
+    torch.save({"sec_decoder": {}, "normalizer": {"sec_phys": {}}}, checkpoint)
+
+    result = runner.invoke(
+        app,
+        [
+            "rollout",
+            "dummy.parquet",
+            "--checkpoint",
+            str(checkpoint),
+            "--geometry",
+            "dummy_geometry.pkl",
+        ],
+    )
+
+    assert result.exit_code == 1
+    assert "checkpoint has no model_config" in result.output

tests/test_cli_train_overrides.py

@@ -0,0 +1,177 @@
+"""Tests for `giant train`'s stage-prefixed CLI flags: --stage1-*/--stage2-*
+must independently override each stage's config block, and must take precedence
+over the older shared flags (--mode/--hidden-dim/--n-critic/... ) that still
+apply the same value to both stages for backward compatibility."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from typer.testing import CliRunner
+
+import giant.cli as cli
+
+runner = CliRunner()
+
+
+def _invoke_and_capture_cfg(monkeypatch, tmp_path: Path, args: list[str]) -> dict:
+    captured: dict = {}
+
+    def _fake_run_train_job(*, data, cfg, out_dir, **kwargs):
+        captured["cfg"] = cfg
+
+    monkeypatch.setattr(cli, "run_train_job", _fake_run_train_job)
+
+    result = runner.invoke(
+        cli.app,
+        ["train", "dummy.parquet", "--out", str(tmp_path / "run")] + args,
+    )
+    assert result.exit_code == 0, result.output
+    return captured["cfg"]
+
+
+def test_stage_prefixed_generator_overrides_shared_mode(monkeypatch, tmp_path):
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        ["--mode", "wgan", "--stage1-generator", "flow"],
+    )
+    assert cfg["stage1_model"]["generator"] == "flow"
+    assert cfg["stage2_model"]["generator"] == "wgan"
+
+
+def test_stage2_only_knobs(monkeypatch, tmp_path):
+    # --stage2-stage1-context is exercised separately at the overrides-dict
+    # level (test_overrides_from_flags_stage2_only_knobs in test_config.py):
+    # its only non-default value, "sampled", is rejected by validate_config
+    # (issues.md Issue 1), so it can't appear in a full CLI invocation here.
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        [
+            "--stage2-decoder",
+            "one_shot",
+            "--stage2-k-max",
+            "8",
+            "--stage2-hidden-dim",
+            "32",
+            "--stage2-context-dim",
+            "16",
+        ],
+    )
+    assert cfg["stage2_model"]["decoder"] == "one_shot"
+    assert cfg["stage2_model"]["k_max"] == 8
+    assert cfg["stage2_model"]["hidden_dim"] == 32
+    assert cfg["stage2_model"]["context_dim"] == 16
+    # untouched stage1 defaults
+    assert cfg["stage1_model"]["hidden_dim"] == 256
+
+
+def test_stage1_hidden_dim_flag_overrides_legacy_hidden_dim_flag(monkeypatch, tmp_path):
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        ["--hidden-dim", "64", "--stage1-hidden-dim", "128"],
+    )
+    assert cfg["stage1_model"]["hidden_dim"] == 128
+
+
+def test_wgan_knobs_split_per_stage(monkeypatch, tmp_path):
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        [
+            "--mode",
+            "wgan",
+            "--n-critic",
+            "5",
+            "--stage1-n-critic",
+            "3",
+            "--stage2-gp-weight",
+            "2.5",
+        ],
+    )
+    assert cfg["stage1_model"]["wgan"]["n_critic"] == 3
+    assert cfg["stage1_model"]["wgan"]["gp_weight"] == 10.0
+    assert cfg["stage2_model"]["wgan"]["n_critic"] == 5
+    assert cfg["stage2_model"]["wgan"]["gp_weight"] == 2.5
+
+
+def test_batch_size_invalid_string_errors(monkeypatch, tmp_path):
+    monkeypatch.setattr(cli, "run_train_job", lambda *a, **kw: None)
+    result = runner.invoke(
+        cli.app,
+        ["train", "dummy.parquet", "--out", str(tmp_path / "run"), "--batch-size", "not-a-number"],
+    )
+    assert result.exit_code == 1
+    assert "--batch-size must be an integer or 'auto'" in result.output
+
+
+def test_out_dir_resolution_prefers_explicit_out_over_resume(monkeypatch, tmp_path):
+    captured: dict = {}
+
+    def _fake_run_train_job(*, data, cfg, out_dir, **kwargs):
+        captured["out_dir"] = out_dir
+
+    monkeypatch.setattr(cli, "run_train_job", _fake_run_train_job)
+
+    resume_dir = tmp_path / "resumed_run"
+    resume_dir.mkdir()
+    (resume_dir / "last.pt").touch()
+    explicit_out = tmp_path / "explicit_run"
+
+    result = runner.invoke(
+        cli.app,
+        ["train", "dummy.parquet", "--out", str(explicit_out), "--resume", str(resume_dir / "last.pt")],
+    )
+    assert result.exit_code == 0, result.output
+    assert captured["out_dir"] == explicit_out
+
+
+def test_out_dir_resolution_falls_back_to_resume_parent(monkeypatch, tmp_path):
+    captured: dict = {}
+
+    def _fake_run_train_job(*, data, cfg, out_dir, **kwargs):
+        captured["out_dir"] = out_dir
+
+    monkeypatch.setattr(cli, "run_train_job", _fake_run_train_job)
+
+    resume_dir = tmp_path / "resumed_run"
+    resume_dir.mkdir()
+    (resume_dir / "last.pt").touch()
+
+    result = runner.invoke(cli.app, ["train", "dummy.parquet", "--resume", str(resume_dir / "last.pt")])
+    assert result.exit_code == 0, result.output
+    assert captured["out_dir"] == resume_dir
+
+
+def test_out_dir_resolution_defaults_when_neither_out_nor_resume_given(monkeypatch, tmp_path):
+    captured: dict = {}
+
+    def _fake_run_train_job(*, data, cfg, out_dir, **kwargs):
+        captured["out_dir"] = out_dir
+
+    monkeypatch.setattr(cli, "run_train_job", _fake_run_train_job)
+    monkeypatch.chdir(tmp_path)
+
+    result = runner.invoke(cli.app, ["train", "dummy.parquet"])
+    assert result.exit_code == 0, result.output
+    assert captured["out_dir"] == Path("checkpoints") / cli.gconfig.default_out_dir_name(cli.gconfig.DEFAULT_CONFIG)
+
+
+def test_batch_size_auto_estimates_and_echoes(monkeypatch, tmp_path):
+    captured: dict = {}
+
+    def _fake_run_train_job(*, data, cfg, out_dir, num_workers, **kwargs):
+        captured["batch_size"] = cfg["train"]["batch_size"]
+
+    monkeypatch.setattr(cli, "run_train_job", _fake_run_train_job)
+    monkeypatch.setattr(cli.gconfig, "estimate_batch_size", lambda hidden_dim, n_blocks, device: 123)
+
+    result = runner.invoke(
+        cli.app,
+        ["train", "dummy.parquet", "--out", str(tmp_path / "run"), "--batch-size", "auto"],
+    )
+    assert result.exit_code == 0, result.output
+    assert captured["batch_size"] == 123
+    assert "batch_size: 123 (auto-estimated from free GPU memory)" in result.output

tests/test_cond_layout.py

@@ -0,0 +1,87 @@
+import pytest
+from giant.cond_layout import AXIS_TYPES, CondLayout
+from giant.constants import COND_DIM, COND_DIM_BASE, MATERIAL_PHYS_DIM, PARTICLE_PHYS_DIM
+
+# ── cond_cat column layout ───────────────────────────────────────────────────
+
+
+def test_topn_cols_neither_onehot():
+    layout = CondLayout.from_types("physical", "embedding")
+    assert (layout.particle_topn_col, layout.material_topn_col) == (None, None)
+    assert layout.cat_dim == 2
+
+
+def test_topn_cols_particle_only():
+    layout = CondLayout.from_types("onehot", "physical")
+    assert (layout.particle_topn_col, layout.material_topn_col) == (2, None)
+    assert layout.cat_dim == 3
+
+
+def test_topn_cols_material_only():
+    layout = CondLayout.from_types("physical", "onehot")
+    assert (layout.particle_topn_col, layout.material_topn_col) == (None, 2)
+    assert layout.cat_dim == 3
+
+
+def test_topn_cols_both_onehot_particle_then_material():
+    layout = CondLayout.from_types("onehot", "onehot")
+    assert (layout.particle_topn_col, layout.material_topn_col) == (2, 3)
+    assert layout.cat_dim == 4
+
+
+def test_dense_vocab_cols_are_mode_independent():
+    """Columns 0/1 are always the dense pdg/material index — giant.model.routers
+    reads them without knowing the conditioning mode."""
+    assert (CondLayout.PDG_COL, CondLayout.MAT_COL) == (0, 1)
+    for particle in AXIS_TYPES:
+        for material in AXIS_TYPES:
+            layout = CondLayout.from_types(particle, material)
+            assert layout.particle_topn_col not in (layout.PDG_COL, layout.MAT_COL)
+            assert layout.material_topn_col not in (layout.PDG_COL, layout.MAT_COL)
+
+
+# ── cond_cont slice layout ───────────────────────────────────────────────────
+
+
+def test_cont_slices_tile_cond_cont_exactly():
+    """base / particle_phys / material_phys must partition cond_cont with no
+    gap and no overlap — a gap or overlap is exactly the silent
+    mis-indexing this object exists to prevent."""
+    layout = CondLayout.from_types("physical", "physical")
+    covered = list(range(*layout.base.indices(COND_DIM)))
+    covered += list(range(*layout.particle_phys.indices(COND_DIM)))
+    covered += list(range(*layout.material_phys.indices(COND_DIM)))
+    assert covered == list(range(COND_DIM))
+
+
+def test_cont_slice_widths_match_constants():
+    layout = CondLayout.from_types("embedding", "embedding")
+    assert layout.base == slice(0, COND_DIM_BASE)
+    assert layout.particle_phys.stop - layout.particle_phys.start == PARTICLE_PHYS_DIM
+    assert layout.material_phys.stop - layout.material_phys.start == MATERIAL_PHYS_DIM
+    assert layout.cont_dim == COND_DIM
+
+
+def test_cont_slices_are_mode_independent():
+    """cond_cont is COND_DIM wide in every mode — a non-"physical" axis gets
+    its block zero-filled rather than dropped, so the slices never move."""
+    physical = CondLayout.from_types("physical", "physical")
+    for particle in AXIS_TYPES:
+        for material in AXIS_TYPES:
+            layout = CondLayout.from_types(particle, material)
+            assert layout.base == physical.base
+            assert layout.particle_phys == physical.particle_phys
+            assert layout.material_phys == physical.material_phys
+
+
+# ── validation ───────────────────────────────────────────────────────────────
+
+
+def test_unknown_particle_type_raises():
+    with pytest.raises(ValueError, match="unknown conditioning.particle.type 'bogus'"):
+        CondLayout.from_types("bogus", "physical")
+
+
+def test_unknown_material_type_raises():
+    with pytest.raises(ValueError, match="unknown conditioning.material.type 'bogus'"):
+        CondLayout.from_types("physical", "bogus")

tests/test_condor.py

@@ -62,9 +62,7 @@ def _fake_venv(repo_dir: Path) -> None:
     giant.chmod(0o755)
 
 
-def _prep(
-    rollout_yaml: Path, run_dir: str | Path | None = None, chunks: int = 1
-) -> Path:
+def _prep(rollout_yaml: Path, run_dir: str | Path | None = None, chunks: int = 1) -> Path:
     """``prep`` with small test-sized context bins/sampling."""
     return prep(
         rollout_yaml,
@@ -224,16 +222,12 @@ def test_write_submit_description(tmp_path: Path):
     assert "--chunk" in body and "--run-dir" in body
 
 
-def test_write_submit_requires_synced_venv(
-    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
-):
+def test_write_submit_requires_synced_venv(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
     run_dir = _prep(_write_inputs(tmp_path))
     cfg = SubmitConfig(run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path)
     # No `giant` next to the (fake) active interpreter, so this falls through
     # to repo_dir/.venv/bin/giant, which _write_inputs/_prep also didn't create.
-    monkeypatch.setattr(
-        sys, "executable", str(tmp_path / "not-a-venv" / "bin" / "python")
-    )
+    monkeypatch.setattr(sys, "executable", str(tmp_path / "not-a-venv" / "bin" / "python"))
     with pytest.raises(FileNotFoundError, match="uv sync"):
         write_submit(cfg)
 
@@ -241,9 +235,7 @@ def test_write_submit_requires_synced_venv(
 def test_write_submit_remote_flag(tmp_path: Path):
     run_dir = _prep(_write_inputs(tmp_path))
     _fake_venv(tmp_path)
-    cfg = SubmitConfig(
-        run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, remote=True
-    )
+    cfg = SubmitConfig(run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, remote=True)
     txt = write_submit(cfg).read_text()
     assert "+RemoteJob = True" in txt
     assert "ProvidesETPResources" not in txt
@@ -253,9 +245,7 @@ def test_write_submit_chunks_respect_chunkable(tmp_path: Path):
     assert get_spec("router_gating").chunkable is False
     run_dir = _prep(_write_inputs(tmp_path), chunks=4)
     _fake_venv(tmp_path)
-    cfg = SubmitConfig(
-        run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=4
-    )
+    cfg = SubmitConfig(run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=4)
     write_submit(cfg)
     jobs = [line.split(",") for line in (run_dir / "jobs.txt").read_text().split()]
     counts: dict[str, int] = {}
@@ -272,9 +262,7 @@ def test_write_submit_rejects_n_chunks_mismatch_with_run_meta(tmp_path: Path):
     _job_walltimes instead of a clear error here."""
     run_dir = _prep(_write_inputs(tmp_path), chunks=2)
     _fake_venv(tmp_path)
-    cfg = SubmitConfig(
-        run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=4
-    )
+    cfg = SubmitConfig(run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=4)
     with pytest.raises(ValueError, match="n_chunks"):
         write_submit(cfg)
 
@@ -297,16 +285,9 @@ def test_write_submit_walltime_grows_with_chunk_rows(tmp_path: Path):
     run_dir = _prep(_write_inputs(tmp_path), chunks=2)
     meta = RunMeta.load(run_dir / "run_meta.json")
     _fake_venv(tmp_path)
-    cfg = SubmitConfig(
-        run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=2
-    )
+    cfg = SubmitConfig(run_dir=run_dir, accounting_group="cms", repo_dir=tmp_path, n_chunks=2)
     write_submit(cfg)
-    jobs = {
-        (i, int(k)): int(w)
-        for i, k, w in (
-            line.split(",") for line in (run_dir / "jobs.txt").read_text().split()
-        )
-    }
+    jobs = {(i, int(k)): int(w) for i, k, w in (line.split(",") for line in (run_dir / "jobs.txt").read_text().split())}
     for chunk in range(2):
         expected = estimate_runtime_s("marginal_edep", meta.rows_per_chunk[chunk])
         assert jobs[("marginal_edep", chunk)] == expected

tests/test_config_consumed_keys.py

@@ -0,0 +1,156 @@
+"""Consumed-keys audit (issues.md Issue 5).
+
+`validate_config_keys` (`giant/config.py`) only checks that a config key is
+*declared* — present somewhere in `DEFAULT_CONFIG`, which is generated from
+the frozen dataclasses. It says nothing about whether anything actually
+*reads* the value once parsed. Issues 1, 2 and 4 are three keys that slipped
+through exactly that gap: declared, round-tripped, silently ignored. This
+module walks every leaf path in `DEFAULT_CONFIG` and asserts each is either
+genuinely consumed by the model-building/training/rollout code, or explicitly
+recorded in `_KNOWN_UNUSED` with a reason.
+
+"Consumed" is approximated by static analysis rather than true call-graph
+reachability: for each leaf path's field name, does it appear anywhere in a
+fixed whitelist of source files as a real attribute access, a dict-key-shaped
+string constant, or a function/constructor parameter name (the last of these
+because `Router` subclasses receive their config via `**kwargs` filtered by
+signature — see `giant.model.routers.build_router`)? Docstrings are excluded
+from the string-constant scan so prose mentioning a dotted config path in
+passing can't masquerade as a read of it. This whitelist-based approach is
+deliberately narrower than "anywhere in `giant/`": scanning the whole package
+produces false negatives from unrelated identifier collisions (e.g.
+`giant/analysis/router_gating.py`'s `_top1_shares(..., order: list, ...)`
+parameter would otherwise make `stage2_model.autoregressive.order` read as
+"consumed").
+"""
+
+import ast
+from pathlib import Path
+
+from giant.config import DEFAULT_CONFIG
+
+_REPO_ROOT = Path(__file__).resolve().parents[1]
+
+# Files that legitimately consume model_config / training config at
+# build/train/rollout time. Not `giant/cli.py` (a CLI flag existing is not
+# consumption — that's precisely how Issue 1 slipped through), not
+# `giant/config.py` itself (declaring/parsing a field is not reading it), and
+# not `giant/model/_legacy.py` (the protected v0.2 migration surface, which
+# intentionally re-derives old flat keys under old names).
+_CONSUMER_ROOTS = ("giant/model", "giant/training")
+_CONSUMER_FILES = (
+    "giant/sample.py",
+    "giant/pipeline.py",
+    "giant/rollout.py",
+    "giant/checkpoint_io.py",
+    "giant/particles.py",
+    "giant/materials.py",
+)
+_EXCLUDED_FILES = ("giant/model/_legacy.py",)
+
+# Leaf DEFAULT_CONFIG paths that are declared but not (yet) read anywhere in
+# the consumer whitelist above. Each entry must name the issue that tracks
+# it. If a key here starts showing up as consumed, the fix landed and this
+# entry is stale — see test_known_unused_allow_list_has_no_stale_entries.
+_KNOWN_UNUSED = {
+    "stage2_model.stage1_context": (
+        "issues.md Issue 1 — trainers.py hardcodes stage1_ctx to the "
+        "ground-truth stage-1 output; 'sampled' is now rejected loudly by "
+        "validate_config (not silently accepted), but the key still isn't "
+        "read by any build/train consumer file since only 'truth' can pass "
+        "validation — see Issue 16 for the real implementation"
+    ),
+    "stage2_model.autoregressive.order": (
+        "gitea #30 — validate_config now checks order is 'energy_desc', but "
+        "nothing in the build/train/rollout consumer whitelist reads the "
+        "value itself since it's still single-valued"
+    ),
+}
+
+# "lambda" is a Python keyword, so the dataclasses expose the dict key
+# "lambda" as the field `lambda_weight` (giant/config.py:49-50).
+_FIELD_NAME_OVERRIDES = {"lambda": "lambda_weight"}
+
+
+def _leaf_paths(node: dict, prefix: str = "") -> list[str]:
+    paths = []
+    for key, value in node.items():
+        if prefix == "" and key == "meta":
+            continue
+        path = f"{prefix}.{key}" if prefix else key
+        if isinstance(value, dict):
+            paths.extend(_leaf_paths(value, path))
+        else:
+            paths.append(path)
+    return paths
+
+
+def _field_name(leaf_path: str) -> str:
+    name = leaf_path.rsplit(".", 1)[-1]
+    return _FIELD_NAME_OVERRIDES.get(name, name)
+
+
+def _is_docstring_expr(expr: ast.Expr) -> bool:
+    return isinstance(expr.value, ast.Constant) and isinstance(expr.value.value, str)
+
+
+def _collect_names(source: str, filename: str) -> set[str]:
+    tree = ast.parse(source, filename=filename)
+    docstring_ids = set()
+    for node in ast.walk(tree):
+        if isinstance(node, (ast.Module, ast.ClassDef, ast.FunctionDef, ast.AsyncFunctionDef)):
+            body = getattr(node, "body", [])
+            if body and isinstance(body[0], ast.Expr) and _is_docstring_expr(body[0]):
+                docstring_ids.add(id(body[0].value))
+
+    names: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Attribute):
+            names.add(node.attr)
+        elif isinstance(node, ast.Constant) and isinstance(node.value, str) and id(node) not in docstring_ids:
+            names.add(node.value)
+        elif isinstance(node, ast.arg):
+            names.add(node.arg)
+        elif isinstance(node, ast.keyword) and node.arg is not None:
+            names.add(node.arg)
+    return names
+
+
+def _consumer_files() -> list[Path]:
+    files: set[Path] = {_REPO_ROOT / f for f in _CONSUMER_FILES}
+    for root in _CONSUMER_ROOTS:
+        files |= set((_REPO_ROOT / root).rglob("*.py"))
+    files -= {_REPO_ROOT / f for f in _EXCLUDED_FILES}
+    return sorted(files)
+
+
+def _consumed_names() -> set[str]:
+    names: set[str] = set()
+    for path in _consumer_files():
+        names |= _collect_names(path.read_text(), str(path))
+    return names
+
+
+def test_every_config_key_is_consumed_or_allow_listed():
+    consumed = _consumed_names()
+    unconsumed = {p for p in _leaf_paths(DEFAULT_CONFIG) if _field_name(p) not in consumed}
+    unexplained = unconsumed - _KNOWN_UNUSED.keys()
+    assert not unexplained, (
+        f"config key(s) {sorted(unexplained)} are declared in DEFAULT_CONFIG "
+        "but not read anywhere in the build/train/rollout consumer files "
+        f"({[str(f.relative_to(_REPO_ROOT)) for f in _consumer_files()]}) — "
+        "either wire the key up, or add it to _KNOWN_UNUSED with a reason "
+        "(see issues.md Issue 5)"
+    )
+
+
+def test_known_unused_allow_list_has_no_stale_entries():
+    consumed = _consumed_names()
+    all_paths = set(_leaf_paths(DEFAULT_CONFIG))
+    stale = {p for p in _KNOWN_UNUSED if p not in all_paths or _field_name(p) in consumed}
+    assert not stale, (
+        f"_KNOWN_UNUSED entry/entries {sorted(stale)} no longer belong on the "
+        "allow-list — either the key was removed from DEFAULT_CONFIG, or it "
+        "is now consumed (the underlying issue was fixed). Remove the stale "
+        "entry/entries."
+    )

tests/test_create_root_files.py

@@ -4,7 +4,7 @@ from pathlib import Path
 
 import pytest
 
-from scripts import create_root_files
+from giant.tools import create_root_files
 
 parse_detector_spec = create_root_files.parse_detector_spec
 next_shard_index = create_root_files.next_shard_index
@@ -16,9 +16,7 @@ SimJob = create_root_files.SimJob
 PlanError = create_root_files.PlanError
 
 
-def _write_fake_executable(
-    path: Path, *, output_count: int = 1, exit_code: int = 0, sleep: float = 0.0
-) -> Path:
+def _write_fake_executable(path: Path, *, output_count: int = 1, exit_code: int = 0, sleep: float = 0.0) -> Path:
     """Stand-in for run_pbwo4/run_sampling: writes *output_count* .root files
     into its own cwd (so callers can verify each job gets an isolated workdir
     and that the workdir ends up holding *only* the .root output, matching
@@ -89,17 +87,13 @@ def test_next_shard_index_continues_past_existing(tmp_path):
 
 def test_plan_jobs_rejects_missing_gen(tmp_path):
     with pytest.raises(PlanError):
-        plan_jobs(
-            ["pbwo4"], num_files=2, dataset_root=tmp_path, kind="steps", gen="gen1"
-        )
+        plan_jobs(["pbwo4"], num_files=2, dataset_root=tmp_path, kind="steps", gen="gen1")
 
 
 def test_plan_jobs_rejects_malformed_gen(tmp_path):
     (tmp_path / "raw" / "steps" / "gen1").mkdir(parents=True)
     with pytest.raises(PlanError):
-        plan_jobs(
-            ["pbwo4"], num_files=2, dataset_root=tmp_path, kind="steps", gen="notagen"
-        )
+        plan_jobs(["pbwo4"], num_files=2, dataset_root=tmp_path, kind="steps", gen="notagen")
 
 
 def test_plan_jobs_continues_from_existing_shards(tmp_path):
@@ -108,9 +102,7 @@ def test_plan_jobs_continues_from_existing_shards(tmp_path):
     (gen_dir / "pbwo4" / "shard-000.root").touch()
     (gen_dir / "pbwo4" / "shard-001.root").touch()
 
-    jobs = plan_jobs(
-        ["pbwo4"], num_files=3, dataset_root=tmp_path, kind="steps", gen="gen1"
-    )
+    jobs = plan_jobs(["pbwo4"], num_files=3, dataset_root=tmp_path, kind="steps", gen="gen1")
 
     assert [j.shard_index for j in jobs] == [2, 3, 4]
     assert all(j.detector == "pbwo4" and j.config is None for j in jobs)
@@ -320,9 +312,7 @@ def test_run_all_caps_concurrency(tmp_path):
     assert {d.name for d in dests} == {f"shard-{i:03d}.root" for i in range(6)}
 
     intervals = [json.loads(d.read_text()) for d in dests]
-    events = sorted(
-        [(p["start"], 1) for p in intervals] + [(p["end"], -1) for p in intervals]
-    )
+    events = sorted([(p["start"], 1) for p in intervals] + [(p["end"], -1) for p in intervals])
     concurrent = 0
     peak = 0
     for _, delta in events:

tests/test_dataset.py

@@ -95,7 +95,7 @@ def _dummy_normalizer(width):
 
 def test_streaming_dataset_offsets_colliding_event_ids_across_files(tmp_path):
     """Two files that each restart event_id from 0 (one Geant4 job per file,
-    see scripts/steps_to_parquet.py) must not have their same-numbered events
+    see giant/tools/steps_to_parquet.py) must not have their same-numbered events
     collapsed together: every row from every file must show up in exactly one
     of train/val, and the number of distinct events must be the sum across
     files, not the union of raw ids."""
@@ -126,7 +126,8 @@ def test_streaming_dataset_offsets_colliding_event_ids_across_files(tmp_path):
             target_normalizer=tgt_norm,
             batch_size=4,
             shuffle=False,
-            conditioning="embedding",
+            particle_conditioning="embedding",
+            material_conditioning="embedding",
         )
         return sum(len(batch[0]) for batch in ds)
 

tests/test_dwarf.py

@@ -3,15 +3,15 @@ from typer.testing import CliRunner
 from giant import cli as giant_cli
 from giant.config import Conditioning
 from giant.data import setup_cache
-from scripts import dwarf
-from scripts.dwarf import app
+from giant.tools import dwarf
+from giant.tools.dwarf import app
 from test_pipeline import _make_synthetic_steps
 
 runner = CliRunner()
 
 
 def test_conditioning_enum_shared_across_both_clis():
-    """giant.cli and scripts.dwarf must use the one giant.config.Conditioning
+    """giant.cli and giant.tools.dwarf must use the one giant.config.Conditioning
     enum, not independently redefined copies that could silently drift apart
     on valid --conditioning values."""
     assert dwarf.Conditioning is Conditioning
@@ -51,9 +51,7 @@ def test_convert_rejects_output_with_multiple_files(tmp_path):
 def test_convert_rejects_output_with_parallel_jobs(tmp_path):
     root_file = tmp_path / "shard.root"
     root_file.touch()
-    result = runner.invoke(
-        app, ["convert", str(root_file), "--output", "out.parquet", "--jobs", "2"]
-    )
+    result = runner.invoke(app, ["convert", str(root_file), "--output", "out.parquet", "--jobs", "2"])
     assert result.exit_code != 0
     assert "--output cannot be combined with --jobs > 1" in result.output
 
@@ -103,7 +101,7 @@ def test_warm_cache_writes_sidecar(tmp_path):
     assert loaded is not None
     assert loaded.vocab is not None
     assert loaded.event_index is not None
-    assert "valfrac=0.1_seed=0_cond=physical" in loaded.normalizers
+    assert "valfrac=0.1_seed=0_pcond=physical_mcond=physical" in loaded.normalizers
 
 
 def test_warm_cache_second_run_hits_cache(tmp_path):
@@ -167,5 +165,5 @@ def test_warm_cache_different_val_fraction_is_separate_entry(tmp_path):
     assert "fitting normalizer (streaming)" in result.output
     loaded = setup_cache.load(data, [data])
     assert loaded is not None
-    assert "valfrac=0.1_seed=0_cond=physical" in loaded.normalizers
-    assert "valfrac=0.3_seed=0_cond=physical" in loaded.normalizers
+    assert "valfrac=0.1_seed=0_pcond=physical_mcond=physical" in loaded.normalizers
+    assert "valfrac=0.3_seed=0_pcond=physical_mcond=physical" in loaded.normalizers

tests/test_flow.py

@@ -1,12 +1,24 @@
 import torch
+from giant.config import ConditioningAxisConfig
 from giant.constants import COND_DIM
-from giant.model.network import DenoisingMLP
+from giant.model.network import Stage1Model
 from giant.model.schedule import CosineSchedule, flow_matching_loss
 from giant.sample import sample_flow, sample_ddim
 
+PARTICLE_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+MATERIAL_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+
 
 def _small_model():
-    return DenoisingMLP(pdg_vocab=3, mat_vocab=2, hidden_dim=32, n_blocks=2)
+    return Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=PARTICLE_CFG,
+        material_cfg=MATERIAL_CFG,
+        hidden_dim=32,
+        n_res_blocks=2,
+        n_sec_head_k_max=15,
+    )
 
 
 def _batch(B=8):
@@ -41,7 +53,7 @@ def test_sample_flow_shape():
     cond_cat = torch.zeros(B, 2, dtype=torch.long)
     sample, n_sec = sample_flow(_small_model(), cond_cont, cond_cat, steps=5)
     assert sample.shape == (B, 9)
-    assert n_sec.shape == (B,)
+    assert n_sec is not None and n_sec.shape == (B,)
 
 
 def test_ddpm_loss_nonneg():
@@ -58,4 +70,4 @@ def test_sample_ddim_shape():
     cond_cat = torch.zeros(B, 2, dtype=torch.long)
     sample, n_sec = sample_ddim(_small_model(), cond_cont, cond_cat, schedule, steps=5)
     assert sample.shape == (B, 9)
-    assert n_sec.shape == (B,)
+    assert n_sec is not None and n_sec.shape == (B,)

tests/test_geometry.py

@@ -4,6 +4,7 @@ from pathlib import Path
 from unittest.mock import patch
 
 import numpy as np
+import pandas as pd
 import pytest
 
 from giant import geometry as g
@@ -12,6 +13,70 @@ from giant import geometry as g
 pytest.importorskip("sklearn")
 
 
+def _steps_frame(n=5, with_post=True):
+    rng = np.random.default_rng(0)
+    data = {
+        "pre_x": rng.uniform(-10, 10, n),
+        "pre_y": rng.uniform(-10, 10, n),
+        "pre_z": rng.uniform(-10, 10, n),
+        "material": ["G4_AIR"] * n,
+        "layer_id": np.arange(n, dtype=np.int64),
+    }
+    if with_post:
+        data["post_x"] = rng.uniform(-10, 10, n)
+        data["post_y"] = rng.uniform(-10, 10, n)
+        data["post_z"] = rng.uniform(-10, 10, n)
+    return pd.DataFrame(data)
+
+
+def test_iter_point_batches_missing_columns_raises(tmp_path):
+    path = tmp_path / "steps.parquet"
+    pd.DataFrame({"pre_x": [0.0]}).to_parquet(path)
+    with pytest.raises(ValueError, match="missing columns"):
+        next(g._iter_point_batches(path))
+
+
+def test_iter_point_batches_without_post_columns_yields_pre_only(tmp_path):
+    path = tmp_path / "steps.parquet"
+    df = _steps_frame(n=5, with_post=False)
+    df.to_parquet(path)
+
+    (pos, mat, lay) = next(g._iter_point_batches(path))
+
+    assert pos.shape == (5, 3)
+    np.testing.assert_allclose(pos, df[["pre_x", "pre_y", "pre_z"]].to_numpy(dtype=np.float32))
+    assert list(mat) == ["G4_AIR"] * 5
+    np.testing.assert_array_equal(lay, np.arange(5))
+
+
+def test_iter_point_batches_with_post_columns_doubles_and_concatenates_points(
+    tmp_path,
+):
+    path = tmp_path / "steps.parquet"
+    df = _steps_frame(n=5, with_post=True)
+    df.to_parquet(path)
+
+    (pos, mat, lay) = next(g._iter_point_batches(path))
+
+    # Every step contributes both its pre_pos and post_pos, sharing the
+    # step's material/layer_id label — so batches double in length.
+    assert pos.shape == (10, 3)
+    np.testing.assert_allclose(pos[:5], df[["pre_x", "pre_y", "pre_z"]].to_numpy(dtype=np.float32))
+    np.testing.assert_allclose(pos[5:], df[["post_x", "post_y", "post_z"]].to_numpy(dtype=np.float32))
+    assert list(mat) == ["G4_AIR"] * 10
+    np.testing.assert_array_equal(lay, np.concatenate([np.arange(5), np.arange(5)]))
+
+
+def test_iter_point_batches_respects_batch_size(tmp_path):
+    path = tmp_path / "steps.parquet"
+    df = _steps_frame(n=10, with_post=False)
+    df.to_parquet(path, row_group_size=10)
+
+    batches = list(g._iter_point_batches(path, batch_size=4))
+
+    assert [len(pos) for pos, _, _ in batches] == [4, 4, 2]
+
+
 def _box_batch(n, rng):
     """A labelled point cloud: inside a 100mm box -> PbWO4/0, else AIR/-1."""
     pos = rng.uniform(-200, 200, (n, 3)).astype(np.float32)
@@ -137,9 +202,7 @@ def test_slab_classes_discovered():
 
 def test_slab_query_labels_by_depth():
     orc = _build_slab()
-    pos = np.array(
-        [[0.0, 0.0, 50.0], [0.0, 0.0, 105.0], [0.0, 0.0, 150.0]]
-    )  # layer 0, gap, layer 1
+    pos = np.array([[0.0, 0.0, 50.0], [0.0, 0.0, 105.0], [0.0, 0.0, 150.0]])  # layer 0, gap, layer 1
     material, layer_id, escaped = orc.query(pos)
     assert list(material) == ["G4_PbWO4", "G4_AIR", "G4_W"]
     assert list(layer_id) == [0, -1, 1]
@@ -168,9 +231,7 @@ def test_slab_save_load_roundtrip(tmp_path):
     orc.save(p)
     loaded = g.GeometryOracle.load(p)
 
-    pos = np.array(
-        [[0.0, 0.0, 50.0], [0.0, 0.0, 105.0], [0.0, 0.0, 150.0], [0.0, 0.0, 1e5]]
-    )
+    pos = np.array([[0.0, 0.0, 50.0], [0.0, 0.0, 105.0], [0.0, 0.0, 150.0], [0.0, 0.0, 1e5]])
     m0, l0, e0 = orc.query(pos)
     m1, l1, e1 = loaded.query(pos)
     assert (m0 == m1).all() and (l0 == l1).all() and (e0 == e1).all()

tests/test_layers.py

@@ -0,0 +1,42 @@
+import pytest
+import torch
+
+from giant.model.layers import build_mlp_head
+
+
+def test_build_mlp_head_depth_1_is_bare_linear():
+    head = build_mlp_head(8, 4, hidden=16, depth=1)
+    assert len(head) == 1
+    assert isinstance(head[0], torch.nn.Linear)
+    assert head[0].in_features == 8
+    assert head[0].out_features == 4
+    out = head(torch.randn(3, 8))
+    assert out.shape == (3, 4)
+
+
+def test_build_mlp_head_depth_2_matches_pre_gitea_36_shape():
+    head = build_mlp_head(8, 4, hidden=16, depth=2)
+    assert len(head) == 3
+    assert isinstance(head[0], torch.nn.Linear)
+    assert head[0].in_features == 8
+    assert head[0].out_features == 16
+    assert isinstance(head[1], torch.nn.SiLU)
+    assert isinstance(head[2], torch.nn.Linear)
+    assert head[2].in_features == 16
+    assert head[2].out_features == 4
+    out = head(torch.randn(5, 8))
+    assert out.shape == (5, 4)
+
+
+def test_build_mlp_head_depth_3_has_extra_hidden_layer():
+    head = build_mlp_head(8, 4, hidden=16, depth=3)
+    assert len(head) == 5
+    widths = [(m.in_features, m.out_features) for m in head if isinstance(m, torch.nn.Linear)]
+    assert widths == [(8, 16), (16, 16), (16, 4)]
+    out = head(torch.randn(2, 8))
+    assert out.shape == (2, 4)
+
+
+def test_build_mlp_head_depth_0_raises():
+    with pytest.raises(ValueError, match="depth"):
+        build_mlp_head(8, 4, hidden=16, depth=0)

tests/test_loader.py

@@ -6,7 +6,9 @@ from giant.data.loader import (
     EVENT_ID_FILE_STRIDE,
     build_index_maps,
     build_index_maps_from_files,
+    build_pdg_topn_map_from_files,
     build_process_map_from_files,
+    build_topn_map_from_files,
     event_id_offset,
     find_parquet_files,
     iter_cond_chunks,
@@ -178,6 +180,63 @@ def test_build_process_map_from_files_three_files_partial_overlap(tmp_path):
     assert proc_map["compt"] == 2
 
 
+# ── build_topn_map_from_files / build_pdg_topn_map_from_files ──────────────
+
+
+def test_build_topn_map_from_files_keeps_most_frequent(tmp_path):
+    materials = ["G4_AIR"] * 5 + ["PbWO4"] * 3 + ["G4_Fe"] * 2 + ["G4_Pb"] * 1
+    path = tmp_path / "a.parquet"
+    pd.DataFrame({"material": materials}).to_parquet(path)
+
+    m = build_topn_map_from_files([path], "material", n_classes=3, cast=str)
+
+    assert m.class_map["G4_AIR"] == 0
+    assert m.class_map["PbWO4"] == 1
+    assert m.class_map["G4_Fe"] == 2  # "other" (n_classes - 1)
+    assert m.class_map["G4_Pb"] == 2
+    assert m.other_members == {"G4_Fe": 2, "G4_Pb": 1}
+
+
+def test_build_topn_map_from_files_fewer_values_than_n_classes(tmp_path):
+    path = tmp_path / "a.parquet"
+    pd.DataFrame({"material": ["G4_AIR", "PbWO4"]}).to_parquet(path)
+
+    m = build_topn_map_from_files([path], "material", n_classes=5, cast=str)
+
+    assert m.class_map == {"G4_AIR": 0, "PbWO4": 1}
+    assert m.other_members == {}
+
+
+def test_build_pdg_topn_map_from_files_pools_primary_and_secondary_pdg(tmp_path):
+    """A species that's rare as a primary but common as a secondary must
+    still rank by its pooled (primary + secondary) count, not just its
+    primary-role count alone — the whole point of pooling both roles."""
+    path = tmp_path / "a.parquet"
+    # primary pdg: mostly 11 (electron), one lone 22 (photon)
+    pdg = [11] * 5 + [22] * 1
+    # secondaries: 22 (photon) appears often as a secondary despite being
+    # rare as a primary above
+    sec_pdg_list = [[22, 22]] * 5 + [[]] * 1
+    pd.DataFrame({"pdg": pdg, "sec_pdg_list": sec_pdg_list}).to_parquet(path)
+
+    m = build_pdg_topn_map_from_files([path], n_classes=3)
+
+    # pooled: 11 -> 5, 22 -> 1 (primary) + 10 (secondary) = 11
+    assert m.class_map[22] == 0
+    assert m.class_map[11] == 1
+
+
+def test_build_pdg_topn_map_from_files_missing_sec_pdg_list_column(tmp_path):
+    """Files predating the parent->child join have no sec_pdg_list column —
+    must not raise, just count the primary pdg column alone."""
+    path = tmp_path / "a.parquet"
+    pd.DataFrame({"pdg": [11, 11, 22]}).to_parquet(path)
+
+    m = build_pdg_topn_map_from_files([path], n_classes=3)
+
+    assert m.class_map == {11: 0, 22: 1}
+
+
 # ── build_index_maps (in-memory) ────────────────────────────────────────────
 
 
@@ -256,9 +315,7 @@ def test_build_index_maps_from_files_ordering_independent_of_file_order(tmp_path
 
 def test_build_index_maps_from_files_numeric_sort_for_nuclear_codes(tmp_path):
     path = tmp_path / "a.parquet"
-    pd.DataFrame({"pdg": [22, 1000060120, 11], "material": ["X", "X", "X"]}).to_parquet(
-        path
-    )
+    pd.DataFrame({"pdg": [22, 1000060120, 11], "material": ["X", "X", "X"]}).to_parquet(path)
 
     pdg_map, _ = build_index_maps_from_files([path])
     assert list(pdg_map.keys()) == [11, 22, 1000060120]
@@ -339,9 +396,7 @@ def test_load_event_ids_applies_offset(tmp_path):
     path = tmp_path / "a.parquet"
     pd.DataFrame({"event_id": [0, 1, 2]}).to_parquet(path)
     offset = event_id_offset(1)
-    np.testing.assert_array_equal(
-        load_event_ids(path, offset=offset), [offset, offset + 1, offset + 2]
-    )
+    np.testing.assert_array_equal(load_event_ids(path, offset=offset), [offset, offset + 1, offset + 2])
 
 
 def test_load_event_ids_raises_when_event_id_reaches_stride(tmp_path):

tests/test_materials.py

@@ -23,11 +23,7 @@ def test_get_material_properties_unfilled_entry_raises():
 
 
 def test_get_material_properties_returns_filled_entry_from_injected_table():
-    table = {
-        "G4_Pb": MaterialProperties(
-            z_eff=82.0, a_eff=207.2, density=11.35, x0=0.5612, lambda_int=17.59
-        )
-    }
+    table = {"G4_Pb": MaterialProperties(z_eff=82.0, a_eff=207.2, density=11.35, x0=0.5612, lambda_int=17.59)}
     props = get_material_properties("G4_Pb", table)
     assert props.z_eff == 82.0
     assert props.a_eff == 207.2

tests/test_migration_v02_v03.py

@@ -0,0 +1,283 @@
+"""Migration acceptance test for v0.3.0 step 2: "load a v0.2 checkpoint
+through migrate_config + the new build_models, and diff its outputs against
+v0.2 code on the same input batch — bit-identical, or the refactor has
+changed something it should not have."
+
+No `/ceph` access on this machine (see CLAUDE.md's Compute environment
+section), so a real trained checkpoint can't be used here — a separate
+portal-machine follow-up with a real checkpoint is planned instead. This
+test is the synthetic stand-in: build a v0.2-shaped
+model from the frozen `tests/legacy/network_v02_snapshot.py` classes with
+fixed-seed random weights (playing the role of "a v0.2 checkpoint"), migrate
+its config and remap its state dict onto the new `build_models` output, and
+assert the two produce bit-identical output on the same random input batch.
+"""
+
+import torch
+
+from giant.constants import COND_DIM, SEC_SLOT_DIM, X_DIM
+from giant.model import network as net
+from tests.legacy import network_v02_snapshot as legacy
+
+PDG_VOCAB = 12
+MAT_VOCAB = 4
+HIDDEN_DIM = 32
+N_BLOCKS = 2
+EMB_DIM = 8
+K = 6  # small k_max for a fast test
+BATCH = 5
+
+
+def _legacy_model_config(mode: str, conditioning: str) -> dict:
+    return {
+        "pdg_vocab": PDG_VOCAB,
+        "mat_vocab": MAT_VOCAB,
+        "hidden_dim": HIDDEN_DIM,
+        "n_blocks": N_BLOCKS,
+        "emb_dim": EMB_DIM,
+        "dropout": 0.0,
+        "k_max": K,
+        "conditioning": conditioning,
+        "router": {"enabled": False},
+        "mode": mode,
+        "noise_dim": 16,
+    }
+
+
+def _random_batch(seed: int):
+    g = torch.Generator().manual_seed(seed)
+    cond_cont = torch.randn(BATCH, COND_DIM, generator=g)
+    cond_cat = torch.randint(0, min(PDG_VOCAB, MAT_VOCAB), (BATCH, 2), generator=g)
+    x1 = torch.randn(BATCH, X_DIM, generator=g)
+    x2 = torch.randn(BATCH, K * SEC_SLOT_DIM, generator=g)
+    t = torch.rand(BATCH, generator=g)
+    return cond_cont, cond_cat, x1, x2, t
+
+
+def _assert_bit_identical(a: torch.Tensor, b: torch.Tensor, label: str) -> None:
+    assert a.shape == b.shape, f"{label}: shape mismatch {a.shape} vs {b.shape}"
+    assert torch.equal(a, b), f"{label}: outputs diverged, max abs diff = {(a - b).abs().max().item()}"
+
+
+def _run_migration_check(mode: str, conditioning: str) -> None:
+    torch.manual_seed(0)
+    legacy_cfg = _legacy_model_config(mode, conditioning)
+
+    if mode == "wgan":
+        old_stage1 = legacy.WGANGenerator(
+            pdg_vocab=PDG_VOCAB,
+            mat_vocab=MAT_VOCAB,
+            hidden_dim=HIDDEN_DIM,
+            n_blocks=N_BLOCKS,
+            emb_dim=EMB_DIM,
+            noise_dim=16,
+            dropout=0.0,
+            k_max=K,
+            conditioning=conditioning,
+        )
+        old_stage2 = legacy.WGANSecondaryGenerator(
+            pdg_vocab=PDG_VOCAB,
+            mat_vocab=MAT_VOCAB,
+            hidden_dim=HIDDEN_DIM,
+            n_blocks=N_BLOCKS,
+            emb_dim=EMB_DIM,
+            sec_dim=K * SEC_SLOT_DIM,
+            noise_dim=16,
+            dropout=0.0,
+            conditioning=conditioning,
+        )
+    else:
+        old_stage1 = legacy.DenoisingMLP(
+            pdg_vocab=PDG_VOCAB,
+            mat_vocab=MAT_VOCAB,
+            hidden_dim=HIDDEN_DIM,
+            n_blocks=N_BLOCKS,
+            emb_dim=EMB_DIM,
+            dropout=0.0,
+            k_max=K,
+            conditioning=conditioning,
+        )
+        old_stage2 = legacy.SecondaryDecoder(
+            pdg_vocab=PDG_VOCAB,
+            mat_vocab=MAT_VOCAB,
+            hidden_dim=HIDDEN_DIM,
+            n_blocks=N_BLOCKS,
+            emb_dim=EMB_DIM,
+            sec_dim=K * SEC_SLOT_DIM,
+            dropout=0.0,
+            conditioning=conditioning,
+        )
+    old_stage1.eval()
+    old_stage2.eval()
+
+    cond_cont, cond_cat, x1, x2, t = _random_batch(seed=123)
+    z1 = torch.randn(BATCH, 16, generator=torch.Generator().manual_seed(456))
+    z2 = torch.randn(BATCH, 16, generator=torch.Generator().manual_seed(789))
+
+    with torch.no_grad():
+        if mode == "wgan":
+            old_out1 = old_stage1(z1, cond_cont, cond_cat)
+        else:
+            old_out1 = old_stage1(x1, t, cond_cont, cond_cat)
+        old_n_sec = old_stage1.predict_n_sec(cond_cont, cond_cat)
+        if mode == "wgan":
+            old_out2 = old_stage2(z2, cond_cont, cond_cat, old_out1)
+        else:
+            old_out2 = old_stage2(x2, t, cond_cont, cond_cat, old_out1)
+
+    # --- migrate: config + state dict, through the new build_models ---
+    new_models = net.build_models(legacy_cfg)
+    new_stage1, new_stage2 = new_models["stage1"], new_models["stage2"]
+    assert isinstance(new_stage1, net.Stage1Model)
+    assert isinstance(new_stage2, net.Stage2OneShot)
+    # n_sec.owner="stage1": n_sec lives on stage1, not stage2, for a
+    # migrated v0.2 checkpoint.
+    assert new_stage1.n_sec_head is not None
+    assert new_stage2.n_sec_head is None
+
+    remapped1, remapped2 = net.migrate_legacy_state_dict(old_stage1.state_dict(), old_stage2.state_dict())
+    missing1, unexpected1 = new_stage1.load_state_dict(remapped1, strict=True)
+    missing2, unexpected2 = new_stage2.load_state_dict(remapped2, strict=True)
+    assert not missing1 and not unexpected1
+    assert not missing2 and not unexpected2
+    new_stage1.eval()
+    new_stage2.eval()
+
+    with torch.no_grad():
+        if mode == "wgan":
+            new_out1 = new_stage1(z1, cond_cont, cond_cat)
+        else:
+            new_out1 = new_stage1(x1, cond_cont, cond_cat, t=t)
+        new_n_sec = new_stage1.predict_n_sec(cond_cont, cond_cat)
+        if mode == "wgan":
+            new_out2 = new_stage2(z2, cond_cont, cond_cat, new_out1)
+        else:
+            new_out2 = new_stage2(x2, cond_cont, cond_cat, new_out1, t=t)
+
+    _assert_bit_identical(old_out1, new_out1, f"stage1 output ({mode}, {conditioning})")
+    _assert_bit_identical(old_n_sec, new_n_sec, f"n_sec logits ({mode}, {conditioning})")
+    _assert_bit_identical(old_out2, new_out2, f"stage2 output ({mode}, {conditioning})")
+
+
+def test_migration_flow_embedding():
+    _run_migration_check(mode="flow", conditioning="embedding")
+
+
+def test_migration_flow_physical():
+    _run_migration_check(mode="flow", conditioning="physical")
+
+
+def test_migration_wgan_embedding():
+    _run_migration_check(mode="wgan", conditioning="embedding")
+
+
+def test_migration_wgan_physical():
+    _run_migration_check(mode="wgan", conditioning="physical")
+
+
+def test_migrate_legacy_model_config_shape():
+    """_migrate_legacy_model_config produces the nested shape build_models
+    expects, with the n_sec.owner marker set so build_models routes the
+    n_sec head back onto stage 1."""
+    legacy_cfg = _legacy_model_config(mode="flow", conditioning="physical")
+    migrated = net._migrate_legacy_model_config(legacy_cfg)
+    assert migrated["pdg_vocab"] == PDG_VOCAB
+    assert migrated["mat_vocab"] == MAT_VOCAB
+    assert migrated["conditioning"]["particle"]["type"] == "physical"
+    assert migrated["conditioning"]["particle"]["n_layers"] == 2
+    assert migrated["conditioning"]["material"]["n_layers"] == 2
+    assert migrated["stage1_model"]["hidden_dim"] == HIDDEN_DIM
+    assert migrated["stage2_model"]["n_sec"]["owner"] == "stage1"
+    assert migrated["stage2_model"]["decoder"] == "one_shot"
+
+
+def test_migrate_legacy_model_config_nonzero_expert_dims_raises():
+    """Regression: a v0.2 checkpoint's model_config carrying a non-default
+    expert_hidden_dim/expert_n_blocks must fail loudly through this path too
+    — not just giant.config.migrate_config's parallel TOML-load path.
+    Silently dropping these keys (build_router's kwarg filtering) would
+    resize the experts instead of refusing."""
+    legacy_cfg = _legacy_model_config(mode="flow", conditioning="physical")
+    legacy_cfg["router"] = {
+        "enabled": True,
+        "expert_hidden_dim": 128,
+        "expert_n_blocks": 0,
+    }
+    try:
+        net._migrate_legacy_model_config(legacy_cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "expert_hidden_dim" in str(e)
+
+
+def test_migrate_legacy_model_config_zero_expert_dims_dropped_silently():
+    legacy_cfg = _legacy_model_config(mode="flow", conditioning="physical")
+    legacy_cfg["router"] = {
+        "enabled": True,
+        "expert_hidden_dim": 0,
+        "expert_n_blocks": 0,
+    }
+    migrated = net._migrate_legacy_model_config(legacy_cfg)
+    assert "expert_hidden_dim" not in migrated["stage1_model"]["router"]
+    assert "expert_n_blocks" not in migrated["stage1_model"]["router"]
+    assert "expert_hidden_dim" not in migrated["stage2_model"]["router"]
+    assert "expert_n_blocks" not in migrated["stage2_model"]["router"]
+
+
+def test_build_models_with_legacy_config_nonzero_expert_dims_raises():
+    """The same check must also fire through the actual caller,
+    build_models, not just the internal helper directly."""
+    legacy_cfg = _legacy_model_config(mode="flow", conditioning="physical")
+    legacy_cfg["router"] = {
+        "enabled": True,
+        "expert_hidden_dim": 128,
+        "expert_n_blocks": 0,
+    }
+    try:
+        net.build_models(legacy_cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "expert_hidden_dim" in str(e)
+
+
+def test_build_models_accepts_new_nested_shape_unchanged():
+    """A dict that already has a 'stage1_model' key (the new shape) is
+    passed through build_models without going through the legacy migration
+    path at all."""
+    cfg = {
+        "pdg_vocab": PDG_VOCAB,
+        "mat_vocab": MAT_VOCAB,
+        "conditioning": {
+            "out_dim": 32,
+            "particle": {"type": "physical", "emb_dim": EMB_DIM, "n_layers": 1},
+            "material": {"type": "physical", "emb_dim": EMB_DIM, "n_layers": 1},
+        },
+        "stage1_model": {
+            "active": True,
+            "generator": "flow",
+            "hidden_dim": HIDDEN_DIM,
+            "n_res_blocks": N_BLOCKS,
+            "dropout": 0.0,
+            "flow": {"time_dim": 16},
+            "router": {"enabled": False},
+        },
+        "stage2_model": {
+            "active": True,
+            "decoder": "one_shot",
+            "generator": "flow",
+            "hidden_dim": HIDDEN_DIM,
+            "n_res_blocks": N_BLOCKS,
+            "dropout": 0.0,
+            "k_max": K,
+            "context_dim": 16,
+            "n_sec": {"mode": "head"},
+            "flow": {"time_dim": 16},
+            "router": {"enabled": False, "tie_to_stage1": False},
+        },
+    }
+    models = net.build_models(cfg)
+    assert isinstance(models["stage1"], net.Stage1Model)
+    assert isinstance(models["stage2"], net.Stage2OneShot)
+    # Fresh v0.3.0 config, n_sec.owner defaults to "stage2": n_sec lives on stage 2.
+    assert models["stage1"].n_sec_head is None
+    assert models["stage2"].n_sec_head is not None

tests/test_objectives.py

@@ -0,0 +1,251 @@
+"""Tests for `giant/model/objectives.py` — the generator/objective registry
+(gitea #32) that replaced bare `generator in ("flow", "ddpm", "wgan")`
+string checks scattered across models.py/sample.py/builders.py/
+stage2_inputs.py/trainers.py."""
+
+import pytest
+import torch
+
+from giant.config import ConditioningAxisConfig
+from giant.constants import COND_DIM, CONT_SLOT_DIM, PARTICLE_PHYS_DIM, SEC_SLOT_DIM, X_DIM
+from giant.model.network import (
+    OBJECTIVE_REGISTRY,
+    DdpmObjective,
+    FlowObjective,
+    Stage1Model,
+    Stage2Autoregressive,
+    Stage2OneShot,
+    WganObjective,
+    build_objective,
+)
+from giant.model.schedule import CosineSchedule, flow_matching_loss, flow_matching_loss_secondary
+
+_PHYS_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+
+
+def _cond(B: int, pdg: int = 3, mat: int = 2) -> tuple[torch.Tensor, torch.Tensor]:
+    cond_cont = torch.randn(B, COND_DIM)
+    cond_cat = torch.stack([torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1)
+    return cond_cont, cond_cat
+
+
+# ── registry ─────────────────────────────────────────────────────────────
+
+
+def test_registry_has_exactly_the_three_known_objectives():
+    assert set(OBJECTIVE_REGISTRY) == {"flow", "ddpm", "wgan"}
+
+
+def test_build_objective_returns_correct_concrete_type():
+    assert isinstance(build_objective("flow"), FlowObjective)
+    assert isinstance(build_objective("ddpm"), DdpmObjective)
+    assert isinstance(build_objective("wgan"), WganObjective)
+
+
+def test_build_objective_unknown_name_raises():
+    with pytest.raises(ValueError, match="unknown generator/objective"):
+        build_objective("bogus")
+
+
+def test_build_objective_filters_kwargs_by_signature():
+    # FlowObjective takes no constructor args — n_steps (a DdpmObjective-only
+    # kwarg) must be silently dropped, not raise a TypeError.
+    build_objective("flow", n_steps=500)
+    ddpm = build_objective("ddpm", n_steps=250)
+    assert isinstance(ddpm, DdpmObjective)
+    assert ddpm.n_steps == 250
+
+
+# ── flags ────────────────────────────────────────────────────────────────
+
+
+def test_flow_objective_flags():
+    obj = build_objective("flow")
+    assert obj.needs_time is True
+    assert obj.is_adversarial is False
+    assert obj.folds_type_slice is False
+    assert obj.supports_stage2_decoder is True
+
+
+def test_ddpm_objective_flags():
+    obj = build_objective("ddpm")
+    assert obj.needs_time is True
+    assert obj.is_adversarial is False
+    assert obj.folds_type_slice is False
+    assert obj.supports_stage2_decoder is False
+
+
+def test_wgan_objective_flags():
+    obj = build_objective("wgan")
+    assert obj.needs_time is False
+    assert obj.is_adversarial is True
+    assert obj.folds_type_slice is True
+    assert obj.supports_stage2_decoder is True
+
+
+# ── trunk_in_dim ─────────────────────────────────────────────────────────
+
+
+def test_trunk_in_dim_flow_and_ddpm_pass_through_out_dim():
+    assert build_objective("flow").trunk_in_dim(out_dim=9, noise_dim=8) == 9
+    assert build_objective("ddpm").trunk_in_dim(out_dim=9, noise_dim=8) == 9
+
+
+def test_trunk_in_dim_wgan_uses_noise_dim():
+    assert build_objective("wgan").trunk_in_dim(out_dim=9, noise_dim=8) == 8
+
+
+# ── ddpm schedule ────────────────────────────────────────────────────────
+
+
+def test_ddpm_build_schedule_has_requested_length():
+    schedule = build_objective("ddpm").build_schedule(n_steps=17, device=torch.device("cpu"))
+    assert isinstance(schedule, CosineSchedule)
+    assert schedule.T == 17
+
+
+def test_flow_and_wgan_build_schedule_is_none():
+    assert build_objective("flow").build_schedule(100, torch.device("cpu")) is None
+    assert build_objective("wgan").build_schedule(100, torch.device("cpu")) is None
+
+
+# ── stage1_loss parity ──────────────────────────────────────────────────
+
+
+def test_flow_objective_stage1_loss_matches_direct_call():
+    torch.manual_seed(0)
+    model = Stage1Model(
+        pdg_vocab=3, mat_vocab=2, particle_cfg=_PHYS_CFG, material_cfg=_PHYS_CFG, hidden_dim=16, n_res_blocks=1
+    )
+    cond_cont, cond_cat = _cond(4)
+    x1 = torch.randn(4, X_DIM)
+
+    torch.manual_seed(1)
+    expected = flow_matching_loss(model, x1, cond_cont, cond_cat)
+    torch.manual_seed(1)
+    actual = build_objective("flow").stage1_loss(model, x1, cond_cont, cond_cat)
+    assert torch.allclose(actual, expected)
+
+
+def test_ddpm_objective_stage1_loss_matches_direct_call():
+    torch.manual_seed(0)
+    model = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="ddpm",
+    )
+    cond_cont, cond_cat = _cond(4)
+    x1 = torch.randn(4, X_DIM)
+    objective = build_objective("ddpm", n_steps=50)
+    schedule = objective.build_schedule(50, torch.device("cpu"))
+    assert isinstance(schedule, CosineSchedule)
+
+    torch.manual_seed(1)
+    expected = schedule.loss(model, x1, cond_cont, cond_cat)
+    torch.manual_seed(1)
+    actual = objective.stage1_loss(model, x1, cond_cont, cond_cat, schedule=schedule)
+    assert torch.allclose(actual, expected)
+
+
+def test_ddpm_objective_stage1_loss_requires_a_schedule():
+    model = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="ddpm",
+    )
+    cond_cont, cond_cat = _cond(4)
+    with pytest.raises(AssertionError):
+        build_objective("ddpm").stage1_loss(model, torch.randn(4, X_DIM), cond_cont, cond_cat, schedule=None)
+
+
+# ── stage2_loss dispatch ─────────────────────────────────────────────────
+
+
+def test_flow_objective_stage2_loss_one_shot_matches_direct_call():
+    torch.manual_seed(0)
+    B, k_max = 4, 5
+    sec_dim = k_max * SEC_SLOT_DIM
+    model = Stage2OneShot(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="flow",
+        sec_dim=sec_dim,
+        k_max=k_max,
+    )
+    cond_cont, cond_cat = _cond(B)
+    stage1_ctx = torch.randn(B, X_DIM)
+    x1_s2 = torch.randn(B, sec_dim)
+    sec_mask = torch.ones(B, k_max, dtype=torch.bool)
+
+    torch.manual_seed(1)
+    expected = flow_matching_loss_secondary(model, x1_s2, cond_cont, cond_cat, stage1_ctx, sec_mask, type_dim=None)
+    torch.manual_seed(1)
+    actual = build_objective("flow").stage2_loss(
+        model, x1_s2, cond_cont, cond_cat, stage1_ctx, sec_mask, type_dim=None, ar_inputs=None
+    )
+    assert torch.allclose(actual, expected)
+
+
+def test_flow_objective_stage2_loss_dispatches_to_ar_when_ar_inputs_given():
+    torch.manual_seed(0)
+    B, k_max = 4, 5
+    model = Stage2Autoregressive(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="flow",
+        k_max=k_max,
+    )
+    cond_cont, cond_cat = _cond(B)
+    stage1_ctx = torch.randn(B, X_DIM)
+    token_dim = CONT_SLOT_DIM + PARTICLE_PHYS_DIM
+    x1_s2 = torch.randn(B, k_max, token_dim)
+    sec_mask = torch.ones(B, k_max, dtype=torch.bool)
+    ar_inputs = {
+        "history_feat": torch.randn(B, k_max, token_dim),
+        "has_prev": torch.ones(B, k_max, dtype=torch.bool),
+        "remaining_frac": torch.rand(B, k_max),
+        "slot_idx": torch.linspace(0, 1, k_max).unsqueeze(0).expand(B, -1),
+    }
+
+    loss = build_objective("flow").stage2_loss(
+        model, x1_s2, cond_cont, cond_cat, stage1_ctx, sec_mask, type_dim=None, ar_inputs=ar_inputs
+    )
+    assert loss.dim() == 0
+    assert torch.isfinite(loss)
+
+
+def test_ddpm_objective_stage2_loss_not_implemented():
+    dummy_model = torch.nn.Module()
+    dummy = torch.zeros(1)
+    with pytest.raises(NotImplementedError):
+        build_objective("ddpm").stage2_loss(
+            dummy_model, dummy, dummy, dummy, dummy, torch.ones(1, 1, dtype=torch.bool), type_dim=None
+        )
+
+
+def test_wgan_objective_has_no_loss_methods():
+    dummy_model = torch.nn.Module()
+    dummy = torch.zeros(1)
+    objective = build_objective("wgan")
+    with pytest.raises(NotImplementedError):
+        objective.stage1_loss(dummy_model, dummy, dummy, dummy)
+    with pytest.raises(NotImplementedError):
+        objective.stage2_loss(
+            dummy_model, dummy, dummy, dummy, dummy, torch.ones(1, 1, dtype=torch.bool), type_dim=None
+        )

tests/test_particles.py

@@ -1,7 +1,11 @@
 import numpy as np
 import pytest
 
+from giant.data.loader import TopNMap
 from giant.particles import (
+    decode_embedding_nearest,
+    decode_topn_class,
+    invert_dense_map,
     nearest_known_pdg,
     particle_mass_charge,
     particle_phys_array,
@@ -45,9 +49,7 @@ def test_ground_state_nucleus_resolved_via_particle_package():
     """He-4 (Z=2, A=4) is a common nuclide in `particle`'s ground-state table."""
     mass, charge = particle_mass_charge(1000020040)
     assert charge == pytest.approx(2.0)
-    assert mass == pytest.approx(
-        4 * 931.494, rel=0.05
-    )  # near A*amu, binding-energy-corrected
+    assert mass == pytest.approx(4 * 931.494, rel=0.05)  # near A*amu, binding-energy-corrected
 
 
 def test_nuclear_isomer_falls_back_to_z_a_decode():
@@ -126,3 +128,99 @@ def test_nearest_known_pdg_shape():
     )
     assert result.shape == (n,)
     assert set(result.tolist()) <= set(candidates)
+
+
+# ── invert_dense_map ─────────────────────────────────────────────────────
+
+
+def test_invert_dense_map_round_trips():
+    pdg_map = {22: 0, 11: 1, -11: 2, 2212: 3}
+    inv = invert_dense_map(pdg_map)
+    for pdg, idx in pdg_map.items():
+        assert inv[idx] == pdg
+
+
+# ── decode_topn_class ────────────────────────────────────────────────────
+
+
+def _topn_fixture():
+    # n_classes=4: photon/electron/positron get their own class (0,1,2),
+    # everything else (proton, neutron) falls into "other" (class 3).
+    class_map = {22: 0, 11: 1, -11: 2, 2212: 3, 2112: 3}
+    other_members = {2212: 7, 2112: 3}
+    return TopNMap(class_map=class_map, other_members=other_members), 4
+
+
+def test_decode_topn_class_known_classes_are_exact():
+    topn_map, n_classes = _topn_fixture()
+    out = decode_topn_class(np.array([0, 1, 2]), topn_map, n_classes)
+    np.testing.assert_array_equal(out, [22, 11, -11])
+
+
+def test_decode_topn_class_other_modal_picks_most_frequent():
+    topn_map, n_classes = _topn_fixture()
+    out = decode_topn_class(np.array([3, 3]), topn_map, n_classes, other_policy="modal")
+    assert (out == 2212).all()  # count 7 > 3
+
+
+def test_decode_topn_class_other_drop_returns_zero_sentinel():
+    topn_map, n_classes = _topn_fixture()
+    out = decode_topn_class(np.array([3]), topn_map, n_classes, other_policy="drop")
+    assert out[0] == 0
+
+
+def test_decode_topn_class_other_sample_stays_within_members():
+    topn_map, n_classes = _topn_fixture()
+    rng = np.random.default_rng(0)
+    out = decode_topn_class(np.full(50, 3), topn_map, n_classes, other_policy="sample", rng=rng)
+    assert set(out.tolist()) <= {2212, 2112}
+
+
+def test_decode_topn_class_unknown_other_policy_raises():
+    topn_map, n_classes = _topn_fixture()
+    with pytest.raises(ValueError):
+        decode_topn_class(np.array([3]), topn_map, n_classes, other_policy="bogus")
+
+
+def test_decode_topn_class_empty_other_members_raises():
+    class_map = {22: 0, 11: 1}
+    topn_map = TopNMap(class_map=class_map, other_members={})
+    with pytest.raises(ValueError):
+        decode_topn_class(np.array([1]), topn_map, 2, other_policy="sample")
+
+
+def test_decode_topn_class_preserves_shape():
+    topn_map, n_classes = _topn_fixture()
+    idx = np.array([[0, 1], [2, 3]])
+    out = decode_topn_class(idx, topn_map, n_classes, other_policy="modal")
+    assert out.shape == (2, 2)
+
+
+# ── decode_embedding_nearest ─────────────────────────────────────────────
+
+
+def test_decode_embedding_nearest_exact_row_recovers_pdg():
+    emb_weight = np.array([[1.0, 0.0], [0.0, 1.0], [-1.0, -1.0]])
+    idx_to_pdg = {0: 22, 1: 11, 2: 2212}
+    vectors = np.array([[0.0, 1.0], [-1.0, -1.0]])  # exact rows 1, 2
+    pdg, dist = decode_embedding_nearest(vectors, emb_weight, idx_to_pdg)
+    np.testing.assert_array_equal(pdg, [11, 2212])
+    np.testing.assert_allclose(dist, [0.0, 0.0], atol=1e-8)
+
+
+def test_decode_embedding_nearest_off_manifold_snaps_to_closest_row():
+    emb_weight = np.array([[1.0, 0.0], [0.0, 1.0]])
+    idx_to_pdg = {0: 22, 1: 11}
+    vectors = np.array([[0.9, 0.2]])  # closer to row 0
+    pdg, dist = decode_embedding_nearest(vectors, emb_weight, idx_to_pdg)
+    assert pdg[0] == 22
+    assert dist[0] > 0.0
+
+
+def test_decode_embedding_nearest_preserves_leading_shape():
+    emb_weight = np.array([[1.0, 0.0], [0.0, 1.0]])
+    idx_to_pdg = {0: 22, 1: 11}
+    vectors = np.random.default_rng(0).standard_normal((3, 4, 2))
+    pdg, dist = decode_embedding_nearest(vectors, emb_weight, idx_to_pdg)
+    assert pdg.shape == (3, 4)
+    assert dist.shape == (3, 4)

tests/test_phase2.py

@@ -4,44 +4,65 @@ import numpy as np
 import pytest
 import torch
 
-from giant.constants import COND_DIM, K_MAX, PARTICLE_PHYS_DIM, SEC_DIM, X_DIM
-from giant.model.network import DenoisingMLP, SecondaryDecoder
-from giant.model.schedule import flow_matching_loss_secondary
+from giant.config import ConditioningAxisConfig
+from giant.constants import (
+    COND_DIM,
+    CONT_SLOT_DIM,
+    K_MAX,
+    PARTICLE_PHYS_DIM,
+    SEC_DIM,
+    X_DIM,
+)
+from giant.model.network import Stage1Model, Stage2Autoregressive, Stage2OneShot
+from giant.model.schedule import (
+    flow_matching_loss_secondary,
+    flow_matching_loss_secondary_ar,
+)
 from giant.sample import sample_secondaries
 
 
 # ── helpers ──────────────────────────────────────────────────────────────────
 
 
+def _particle_material_cfg(conditioning: str) -> tuple[ConditioningAxisConfig, ConditioningAxisConfig]:
+    cfg = ConditioningAxisConfig(type=conditioning, emb_dim=16, n_layers=1)
+    return cfg, cfg
+
+
 def _stage1(pdg=3, mat=2, conditioning="embedding"):
-    return DenoisingMLP(
+    particle_cfg, material_cfg = _particle_material_cfg(conditioning)
+    return Stage1Model(
         pdg_vocab=pdg,
         mat_vocab=mat,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
         hidden_dim=32,
-        n_blocks=2,
-        conditioning=conditioning,
+        n_res_blocks=2,
+        n_sec_head_k_max=K_MAX,
     )
 
 
 def _sec_decoder(pdg=3, mat=2, conditioning="embedding"):
-    return SecondaryDecoder(
+    particle_cfg, material_cfg = _particle_material_cfg(conditioning)
+    return Stage2OneShot(
         pdg_vocab=pdg,
         mat_vocab=mat,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
         hidden_dim=32,
-        n_blocks=2,
-        conditioning=conditioning,
+        n_res_blocks=2,
+        generator="flow",
+        time_dim=16,
     )
 
 
 def _cond(B=8, pdg=3, mat=2):
     cond_cont = torch.randn(B, COND_DIM)
-    cond_cat = torch.stack(
-        [torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1
-    )
+    cond_cat = torch.stack([torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1)
     return cond_cont, cond_cat
 
 
-# ── DenoisingMLP Phase-2 additions ───────────────────────────────────────────
+# ── Stage1Model Phase-2 additions ────────────────────────────────────────────
 
 
 def test_predict_n_sec_shape():
@@ -82,7 +103,7 @@ def test_condition_encoder_embedding_mode_has_embedding_tables():
     assert not hasattr(model.cond_enc, "particle_mlp")
 
 
-# ── SecondaryDecoder ──────────────────────────────────────────────────────────
+# ── Stage2OneShot ─────────────────────────────────────────────────────────────
 
 
 @pytest.mark.parametrize("conditioning", ["embedding", "physical"])
@@ -93,7 +114,7 @@ def test_sec_decoder_output_shape(conditioning):
     t = torch.rand(B)
     cond_cont, cond_cat = _cond(B)
     stage1_out = torch.randn(B, X_DIM)
-    out = decoder(x_t, t, cond_cont, cond_cat, stage1_out)
+    out = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t)
     assert out.shape == (B, SEC_DIM)
 
 
@@ -104,7 +125,7 @@ def test_sec_decoder_no_nan():
     t = torch.rand(B)
     cond_cont, cond_cat = _cond(B)
     stage1_out = torch.randn(B, X_DIM)
-    out = decoder(x_t, t, cond_cont, cond_cat, stage1_out)
+    out = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t)
     assert torch.isfinite(out).all()
 
 
@@ -115,7 +136,9 @@ def test_sec_decoder_gradients():
     t = torch.rand(B)
     cond_cont, cond_cat = _cond(B)
     stage1_out = torch.randn(B, X_DIM)
-    decoder(x_t, t, cond_cont, cond_cat, stage1_out).sum().backward()
+    flow_out = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t).sum()
+    nsec_out = decoder.predict_n_sec(cond_cont, cond_cat, stage1_out).sum()
+    (flow_out + nsec_out).backward()
     for name, p in decoder.named_parameters():
         assert p.grad is not None, f"no grad for {name}"
 
@@ -130,9 +153,7 @@ def test_flow_matching_loss_secondary_scalar():
     cond_cont, cond_cat = _cond(B, pdg, mat)
     stage1_out = torch.randn(B, X_DIM)
     sec_mask = torch.ones(B, K_MAX, dtype=torch.bool)
-    loss = flow_matching_loss_secondary(
-        decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask
-    )
+    loss = flow_matching_loss_secondary(decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask)
     assert loss.shape == ()
     assert loss.item() >= 0.0
 
@@ -145,9 +166,7 @@ def test_flow_matching_loss_secondary_mask_zeros_padding():
     cond_cont, cond_cat = _cond(B, pdg, mat)
     stage1_out = torch.randn(B, X_DIM)
     sec_mask = torch.zeros(B, K_MAX, dtype=torch.bool)
-    loss = flow_matching_loss_secondary(
-        decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask
-    )
+    loss = flow_matching_loss_secondary(decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask)
     assert loss.item() == pytest.approx(0.0, abs=1e-6)
 
 
@@ -158,8 +177,102 @@ def test_flow_matching_loss_secondary_has_grad():
     cond_cont, cond_cat = _cond(B, pdg, mat)
     stage1_out = torch.randn(B, X_DIM)
     sec_mask = torch.ones(B, K_MAX, dtype=torch.bool)
-    flow_matching_loss_secondary(
-        decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask
+    flow_matching_loss_secondary(decoder, x1, cond_cont, cond_cat, stage1_out, sec_mask).backward()
+    assert any(p.grad is not None for p in decoder.parameters())
+
+
+# ── masked flow matching loss — autoregressive (v0.3.0 step 5) ─────────────
+
+
+def _sec_decoder_ar(pdg=3, mat=2, k_max=K_MAX):
+    particle_cfg, material_cfg = _particle_material_cfg("embedding")
+    return Stage2Autoregressive(
+        pdg_vocab=pdg,
+        mat_vocab=mat,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        generator="flow",
+        time_dim=16,
+        k_max=k_max,
+    )
+
+
+def _ar_history_inputs(B, K, hist_dim):
+    history_feat = torch.randn(B, K, hist_dim)
+    has_prev = (torch.arange(K) >= 1).unsqueeze(0).expand(B, -1)
+    remaining_frac = torch.rand(B, K)
+    slot_idx = torch.linspace(0, 1, K).unsqueeze(0).expand(B, -1)
+    return history_feat, has_prev, remaining_frac, slot_idx
+
+
+def test_flow_matching_loss_secondary_ar_scalar():
+    B, K, pdg, mat = 8, K_MAX, 3, 2
+    decoder = _sec_decoder_ar(pdg, mat, k_max=K)
+    x1 = torch.randn(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    cond_cont, cond_cat = _cond(B, pdg, mat)
+    stage1_out = torch.randn(B, X_DIM)
+    history_feat, has_prev, remaining_frac, slot_idx = _ar_history_inputs(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    sec_mask = torch.ones(B, K, dtype=torch.bool)
+    loss = flow_matching_loss_secondary_ar(
+        decoder,
+        x1,
+        cond_cont,
+        cond_cat,
+        stage1_out,
+        history_feat,
+        has_prev,
+        remaining_frac,
+        slot_idx,
+        sec_mask,
+    )
+    assert loss.shape == ()
+    assert loss.item() >= 0.0
+
+
+def test_flow_matching_loss_secondary_ar_mask_zeros_padding():
+    B, K, pdg, mat = 4, K_MAX, 3, 2
+    decoder = _sec_decoder_ar(pdg, mat, k_max=K)
+    x1 = torch.randn(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    cond_cont, cond_cat = _cond(B, pdg, mat)
+    stage1_out = torch.randn(B, X_DIM)
+    history_feat, has_prev, remaining_frac, slot_idx = _ar_history_inputs(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    sec_mask = torch.zeros(B, K, dtype=torch.bool)
+    loss = flow_matching_loss_secondary_ar(
+        decoder,
+        x1,
+        cond_cont,
+        cond_cat,
+        stage1_out,
+        history_feat,
+        has_prev,
+        remaining_frac,
+        slot_idx,
+        sec_mask,
+    )
+    assert loss.item() == pytest.approx(0.0, abs=1e-6)
+
+
+def test_flow_matching_loss_secondary_ar_has_grad():
+    B, K, pdg, mat = 4, K_MAX, 3, 2
+    decoder = _sec_decoder_ar(pdg, mat, k_max=K)
+    x1 = torch.randn(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    cond_cont, cond_cat = _cond(B, pdg, mat)
+    stage1_out = torch.randn(B, X_DIM)
+    history_feat, has_prev, remaining_frac, slot_idx = _ar_history_inputs(B, K, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    sec_mask = torch.ones(B, K, dtype=torch.bool)
+    flow_matching_loss_secondary_ar(
+        decoder,
+        x1,
+        cond_cont,
+        cond_cat,
+        stage1_out,
+        history_feat,
+        has_prev,
+        remaining_frac,
+        slot_idx,
+        sec_mask,
     ).backward()
     assert any(p.grad is not None for p in decoder.parameters())
 
@@ -173,9 +286,7 @@ def test_sample_secondaries_shapes():
     cond_cont, cond_cat = _cond(B, pdg, mat)
     stage1_out = torch.randn(B, X_DIM)
     n_sec_pred = torch.randint(0, K_MAX + 1, (B,))
-    sec_cont, sec_phys, sec_valid = sample_secondaries(
-        decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=3
-    )
+    sec_cont, sec_phys, sec_valid = sample_secondaries(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=3)
     assert sec_cont.shape == (B, K_MAX, 4)
     assert sec_phys.shape == (B, K_MAX, PARTICLE_PHYS_DIM)
     assert sec_valid.shape == (B, K_MAX)
@@ -188,9 +299,7 @@ def test_sample_secondaries_valid_mask_matches_n_sec():
     cond_cont, cond_cat = _cond(B, pdg, mat)
     stage1_out = torch.randn(B, X_DIM)
     n_sec_pred = torch.tensor([0, 1, 3, K_MAX])
-    _, _, sec_valid = sample_secondaries(
-        decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2
-    )
+    _, _, sec_valid = sample_secondaries(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2)
     for i, n in enumerate(n_sec_pred.tolist()):
         assert sec_valid[i, :n].all()
         assert not sec_valid[i, n:].any()
@@ -224,9 +333,7 @@ def test_encode_secondaries_energy_conservation():
     pre_dir = rng.standard_normal((N, 3)).astype(np.float32)
     pre_dir /= np.linalg.norm(pre_dir, axis=1, keepdims=True)
 
-    sec_cont = encode_secondaries(
-        sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list
-    )
+    sec_cont = encode_secondaries(sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list)
     assert sec_cont.shape == (N, K_MAX, 6)
     assert np.isfinite(sec_cont).all()
 
@@ -273,9 +380,7 @@ def test_encode_secondaries_stick_logits_match_naive_reference():
             logit = np.clip(logit, -_STICK_LOGIT_CLIP, _STICK_LOGIT_CLIP)
             expected[row, i] = logit
 
-    np.testing.assert_allclose(
-        stick_logits[sec_valid], expected[sec_valid], rtol=1e-4, atol=1e-4
-    )
+    np.testing.assert_allclose(stick_logits[sec_valid], expected[sec_valid], rtol=1e-4, atol=1e-4)
 
 
 def test_encode_secondaries_direction_encoding():
@@ -387,9 +492,7 @@ def test_encode_secondaries_physical_columns_match_ground_truth_pdg():
     sec_valid[0, 0] = True
     pre_dir = np.array([[0.0, 0.0, 1.0]], dtype=np.float32)
 
-    sec_cont = encode_secondaries(
-        sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list
-    )
+    sec_cont = encode_secondaries(sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list)
     mass, charge = particle_mass_charge(11)
     assert sec_cont[0, 0, 4] == pytest.approx(log_transform(np.array([mass]))[0])
     assert sec_cont[0, 0, 5] == pytest.approx(charge)
@@ -423,9 +526,7 @@ def test_decode_secondaries_valid_slots_sum_to_e_sec():
     e_sec = rng.uniform(0.0, 50.0, size=N).astype(np.float32)
     pre_dir = np.tile([0.0, 0.0, 1.0], (N, 1)).astype(np.float32)
 
-    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(
-        sec_cont, n_sec, e_sec, pre_dir
-    )
+    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(sec_cont, n_sec, e_sec, pre_dir)
 
     valid_sum = (sec_E * sec_valid).sum(axis=1)
     has_secondaries = n_sec > 0
@@ -448,9 +549,7 @@ def test_decode_secondaries_zero_n_sec_has_zero_energy():
     e_sec = rng.uniform(1.0, 10.0, size=N).astype(np.float32)
     pre_dir = np.tile([0.0, 0.0, 1.0], (N, 1)).astype(np.float32)
 
-    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(
-        sec_cont, n_sec, e_sec, pre_dir
-    )
+    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(sec_cont, n_sec, e_sec, pre_dir)
 
     assert not sec_valid.any()
     np.testing.assert_allclose(sec_E, 0.0)
@@ -470,9 +569,7 @@ def test_decode_secondaries_degenerate_row_falls_back_to_even_split():
     e_sec = np.array([0.0, 4.0, 9.0, 30.0], dtype=np.float32)
     pre_dir = np.tile([0.0, 0.0, 1.0], (N, 1)).astype(np.float32)
 
-    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(
-        sec_cont, n_sec, e_sec, pre_dir
-    )
+    sec_E, _sec_dir, _mass, _charge, sec_valid = decode_secondaries(sec_cont, n_sec, e_sec, pre_dir)
 
     for i, k in enumerate(n_sec):
         if k == 0:
@@ -496,12 +593,8 @@ def test_decode_secondaries_rescale_preserves_relative_shares():
     n_sec = np.array([4])
     pre_dir = np.array([[0.0, 0.0, 1.0]], dtype=np.float32)
 
-    sec_E_small, _, _, _, sec_valid = decode_secondaries(
-        sec_cont, n_sec, np.array([5.0], dtype=np.float32), pre_dir
-    )
-    sec_E_large, _, _, _, _ = decode_secondaries(
-        sec_cont, n_sec, np.array([50.0], dtype=np.float32), pre_dir
-    )
+    sec_E_small, _, _, _, sec_valid = decode_secondaries(sec_cont, n_sec, np.array([5.0], dtype=np.float32), pre_dir)
+    sec_E_large, _, _, _, _ = decode_secondaries(sec_cont, n_sec, np.array([50.0], dtype=np.float32), pre_dir)
 
     ratio_small = sec_E_small[0, :4] / sec_E_small[0, 0]
     ratio_large = sec_E_large[0, :4] / sec_E_large[0, 0]
@@ -523,20 +616,14 @@ def test_decode_secondaries_mass_charge_round_trip_with_normalizer():
     sec_valid[0, 0] = True
     pre_dir = np.array([[0.0, 0.0, 1.0]], dtype=np.float32)
 
-    sec_cont = encode_secondaries(
-        sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list
-    )
+    sec_cont = encode_secondaries(sec_E_list, sec_dir_list, sec_valid, e_sec, pre_dir, sec_pdg_list=sec_pdg_list)
     norm = Normalizer()
     norm.mean = np.array([-2.0, 0.5], dtype=np.float32)
     norm.std = np.array([3.0, 1.5], dtype=np.float32)
     sec_cont_normed = sec_cont.copy()
-    sec_cont_normed[:, :, 4:6] = norm.transform(
-        sec_cont[:, :, 4:6].reshape(-1, 2)
-    ).reshape(N, K_MAX, 2)
+    sec_cont_normed[:, :, 4:6] = norm.transform(sec_cont[:, :, 4:6].reshape(-1, 2)).reshape(N, K_MAX, 2)
 
     n_sec = np.array([1])
-    _, _, sec_mass, sec_charge, _ = decode_secondaries(
-        sec_cont_normed, n_sec, e_sec, pre_dir, sec_phys_normalizer=norm
-    )
+    _, _, sec_mass, sec_charge, _ = decode_secondaries(sec_cont_normed, n_sec, e_sec, pre_dir, sec_phys_normalizer=norm)
     assert sec_mass[0, 0] == pytest.approx(938.27208943, abs=1e-2)
     assert sec_charge[0, 0] == pytest.approx(1.0, abs=1e-4)

tests/test_pipeline.py

@@ -6,8 +6,10 @@ import pytest
 import torch
 
 from giant import config as gconfig
+from giant.constants import COND_DIM
 from giant.data import setup_cache
-from giant.pipeline import run_train_job
+from giant.data.transforms import Normalizer
+from giant.pipeline import _seed_energy_router, run_train_job
 
 
 def _unit(v):
@@ -46,9 +48,7 @@ def _make_synthetic_steps(path, n_events=20, seed=0):
             pre_dir = np.array([0.0, 0.0, 1.0])
             post_dir = _unit(rng.normal(size=3))
             post_pos = pre_pos + step_length * pre_dir
-            sec_energies = (
-                list(rng.dirichlet(np.ones(n_sec)) * e_sec) if n_sec > 0 else []
-            )
+            sec_energies = list(rng.dirichlet(np.ones(n_sec)) * e_sec) if n_sec > 0 else []
             sec_pdgs = [pdgs[(row_idx + j) % 2] for j in range(n_sec)]
             sec_dirs = [_unit(rng.normal(size=3)) for _ in range(n_sec)]
             rows.append(
@@ -99,10 +99,19 @@ def _tiny_cfg(**train_overrides):
             "warmup_epochs": 0,
             "validate_every": 0,
             "max_val_batches": 1,
+            "wandb": False,
         }
     )
     cfg["train"].update(train_overrides)
-    cfg["model"].update({"hidden_dim": 8, "n_blocks": 1, "emb_dim": 4, "dropout": 0.0})
+    cfg["stage1_model"].update({"hidden_dim": 8, "n_res_blocks": 1, "dropout": 0.0})
+    cfg["stage2_model"].update(
+        # decoder="autoregressive" is DEFAULT_CONFIG's default (v0.3.0 step 5)
+        # and left as-is here on purpose, so this pipeline-level fixture
+        # exercises the real default end-to-end against actual data.
+        {"hidden_dim": 8, "n_res_blocks": 1, "dropout": 0.0}
+    )
+    cfg["conditioning"]["particle"]["emb_dim"] = 4
+    cfg["conditioning"]["material"]["emb_dim"] = 4
     return cfg
 
 
@@ -143,17 +152,106 @@ def test_run_train_job_second_run_hits_cache(tmp_path, data, monkeypatch):
     assert "normalizer: cache hit" in joined
 
 
-def test_run_train_job_warns_when_num_workers_exceeds_shared_quota(
-    tmp_path, data, monkeypatch
-):
+def test_run_train_job_builds_caches_and_persists_sec_type_topn_map(tmp_path, data):
+    """DEFAULT_CONFIG's stage2_model.particle_type.target defaults to
+    "onehot" while conditioning.particle.type stays "physical" — a plain
+    _tiny_cfg() run must build the secondary-type-only pdg top-N map (gitea
+    #29: no longer shared with any conditioning-side onehot map), cache it in
+    the setup-cache sidecar, and persist it into the checkpoint's
+    sec_type_topn_map key, with no extra config needed. pdg_topn_map
+    (conditioning-only) stays unbuilt since conditioning.particle.type is
+    "physical" here."""
+    echo1 = _run(data, tmp_path / "out1")
+    assert any("building pdg top-N map" in m for m in echo1)
+
+    loaded = setup_cache.load(data, [data])
+    assert loaded is not None
+    # stage2_model.particle_type.n_classes = 0 -> conditioning.particle.emb_dim = 4
+    key = setup_cache.topn_key("pdg", 4)
+    assert key in loaded.topn_maps
+    assert set(loaded.topn_maps[key].class_map.keys()) >= {11, 22}
+
+    ckpt = torch.load(tmp_path / "out1" / "last.pt", weights_only=False)
+    assert ckpt.get("pdg_topn_map") is None
+    assert "sec_type_topn_map" in ckpt
+    assert set(ckpt["sec_type_topn_map"]["class_map"].keys()) >= {"11", "22"}
+
+    echo2 = _run(data, tmp_path / "out2")
+    assert any("pdg top-N map: cache hit" in m for m in echo2)
+
+
+def test_run_train_job_independent_cond_and_sec_type_topn_maps(tmp_path, data):
+    """conditioning.particle.type="onehot" and
+    stage2_model.particle_type.target="onehot" with different class counts
+    (gitea #29's fix: stage2_model.particle_type.n_classes decouples the two)
+    build two distinct top-N maps, cached under their own (axis, n_classes)
+    key and persisted under two distinct checkpoint keys — no longer forced
+    to share conditioning.particle.emb_dim."""
+    cfg = _tiny_cfg()
+    cfg["conditioning"]["particle"]["type"] = "onehot"  # emb_dim = 4, from _tiny_cfg
+    cfg["stage2_model"]["particle_type"]["n_classes"] = 3
+    echo = _run(data, tmp_path / "out", cfg=cfg)
+    assert any("mapped to 4 classes" in m for m in echo)
+    assert any("mapped to 3 classes" in m for m in echo)
+
+    loaded = setup_cache.load(data, [data])
+    assert loaded is not None
+    cond_key = setup_cache.topn_key("pdg", 4)
+    type_key = setup_cache.topn_key("pdg", 3)
+    assert cond_key in loaded.topn_maps
+    assert type_key in loaded.topn_maps
+
+    ckpt = torch.load(tmp_path / "out" / "last.pt", weights_only=False)
+    assert ckpt.get("pdg_topn_map") is not None
+    assert ckpt.get("sec_type_topn_map") is not None
+
+
+def test_run_train_job_builds_caches_and_persists_material_topn_map(tmp_path, data):
+    """conditioning.material.type="onehot" is an independent axis from the
+    pdg one above, with its own build/cache-hit branch in run_setup_stage —
+    exercise both here the same way the pdg test above does."""
+    cfg = _tiny_cfg()
+    cfg["conditioning"]["material"]["type"] = "onehot"
+    echo1 = _run(data, tmp_path / "out1", cfg=cfg)
+    assert any("building material top-N map" in m for m in echo1)
+
+    loaded = setup_cache.load(data, [data])
+    assert loaded is not None
+    key = setup_cache.topn_key("material", 4)  # conditioning.material.emb_dim = 4
+    assert key in loaded.topn_maps
+    assert set(loaded.topn_maps[key].class_map.keys()) >= {"G4_AIR", "G4_Fe"}
+
+    ckpt = torch.load(tmp_path / "out1" / "last.pt", weights_only=False)
+    assert "mat_topn_map" in ckpt
+    assert set(ckpt["mat_topn_map"]["class_map"].keys()) >= {"G4_AIR", "G4_Fe"}
+
+    echo2 = _run(data, tmp_path / "out2", cfg=cfg)
+    assert any("material top-N map: cache hit" in m for m in echo2)
+
+
+def test_run_train_job_no_topn_map_for_physical_target(tmp_path, data):
+    cfg = _tiny_cfg()
+    cfg["stage2_model"]["particle_type"] = {"target": "physical", "lambda": 1.0}
+    echo = _run(data, tmp_path / "out", cfg=cfg)
+    assert not any("top-N map" in m for m in echo)
+
+    loaded = setup_cache.load(data, [data])
+    assert loaded is not None
+    assert loaded.topn_maps == {}
+
+
+@pytest.mark.filterwarnings("ignore::DeprecationWarning:multiprocessing.popen_fork")
+def test_run_train_job_warns_when_num_workers_exceeds_shared_quota(tmp_path, data, monkeypatch):
+    # num_workers>0 makes DataLoader actually fork worker subprocesses
+    # (unlike every other test here, which runs with num_workers=0) — pytest
+    # itself is multi-threaded, hence Python's fork-safety warning below.
     monkeypatch.setattr("giant.pipeline.os.cpu_count", lambda: 8)  # quota = 2
     echo = _run(data, tmp_path / "out", num_workers=3)
     assert any("num-workers=3" in m and "exceeds" in m for m in echo)
 
 
-def test_run_train_job_no_warning_when_num_workers_within_shared_quota(
-    tmp_path, data, monkeypatch
-):
+@pytest.mark.filterwarnings("ignore::DeprecationWarning:multiprocessing.popen_fork")
+def test_run_train_job_no_warning_when_num_workers_within_shared_quota(tmp_path, data, monkeypatch):
     monkeypatch.setattr("giant.pipeline.os.cpu_count", lambda: 8)  # quota = 2
     echo = _run(data, tmp_path / "out", num_workers=2)
     assert not any("exceeds" in m for m in echo)
@@ -193,6 +291,70 @@ def test_run_train_job_new_val_fraction_is_partial_hit(tmp_path, data, monkeypat
     assert "fitting normalizer (streaming)" in joined
 
 
+def test_run_train_job_custom_k_max_end_to_end(tmp_path, data):
+    """Regression: stage2_model.k_max other
+    than the K_MAX module constant's default (15) must not produce a shape
+    mismatch between the data pipeline (loader.py/transforms.py padding) and
+    the model (network.py's trunks, sized from this same config value)."""
+    cfg = _tiny_cfg()
+    cfg["stage2_model"]["k_max"] = 3
+    _run(data, tmp_path / "out", cfg=cfg)
+    ckpt = torch.load(tmp_path / "out" / "last.pt", weights_only=False)
+    assert ckpt["model_config"]["stage2_model"]["k_max"] == 3
+
+
+def test_run_train_job_mixed_particle_material_conditioning_end_to_end(tmp_path, data):
+    """Regression: conditioning.particle.type
+    and conditioning.material.type are configured independently and may mix
+    freely — e.g. particle "embedding" with
+    material "physical" — end-to-end through the real data pipeline, not
+    just accepted by validate_config."""
+    cfg = _tiny_cfg()
+    cfg["conditioning"]["particle"]["type"] = "embedding"
+    cfg["conditioning"]["material"]["type"] = "physical"
+    _run(data, tmp_path / "out", cfg=cfg)
+    ckpt = torch.load(tmp_path / "out" / "last.pt", weights_only=False)
+    cond_cfg = ckpt["model_config"]["conditioning"]
+    assert cond_cfg["particle"]["type"] == "embedding"
+    assert cond_cfg["material"]["type"] == "physical"
+
+    cond_norm = Normalizer.from_dict(ckpt["normalizer"]["cond"])
+    # Particle block ([COND_DIM_BASE:COND_DIM_BASE+PARTICLE_PHYS_DIM]) stays
+    # unfitted (mean=0/std=1) since "embedding" never computes real values
+    # for it; the material block is fit for real under "physical".
+    from giant.constants import COND_DIM_BASE, PARTICLE_PHYS_DIM
+
+    assert cond_norm.mean is not None and cond_norm.std is not None
+    np.testing.assert_allclose(cond_norm.mean[COND_DIM_BASE : COND_DIM_BASE + PARTICLE_PHYS_DIM], 0.0)
+    np.testing.assert_allclose(cond_norm.std[COND_DIM_BASE : COND_DIM_BASE + PARTICLE_PHYS_DIM], 1.0)
+    material_std = cond_norm.std[COND_DIM_BASE + PARTICLE_PHYS_DIM :]
+    assert np.all(material_std > 0) and not np.allclose(material_std, 1.0)
+
+
+def test_run_train_job_share_stages_end_to_end(tmp_path, data):
+    """Regression: conditioning.share_stages
+    = true must actually train (not raise NotImplementedError), and the
+    resulting checkpoint's two stages must reload into a single shared
+    ConditionEncoder instance rather than two independent ones."""
+    from giant.model.network import build_models
+
+    cfg = _tiny_cfg()
+    cfg["conditioning"]["share_stages"] = True
+    _run(data, tmp_path / "out", cfg=cfg)
+    ckpt = torch.load(tmp_path / "out" / "last.pt", weights_only=False)
+    assert ckpt["model_config"]["conditioning"]["share_stages"] is True
+
+    built = build_models(ckpt["model_config"])
+    stage1, stage2 = built["stage1"], built["stage2"]
+    assert stage1 is not None and stage2 is not None
+    assert stage1.cond_enc is stage2.cond_enc
+
+    stage1.load_state_dict(ckpt["model"])
+    stage2.load_state_dict(ckpt["sec_decoder"])
+    for p1, p2 in zip(stage1.cond_enc.parameters(), stage2.cond_enc.parameters()):
+        assert torch.equal(p1, p2)
+
+
 def test_run_train_job_matches_uncached_output(tmp_path, data):
     _run(data, tmp_path / "uncached", cache_setup=False)
     _run(data, tmp_path / "cached1", cache_setup=True)
@@ -202,11 +364,63 @@ def test_run_train_job_matches_uncached_output(tmp_path, data):
     cached = torch.load(tmp_path / "cached2" / "last.pt", weights_only=False)
 
     for key in ("cond", "target", "sec_phys"):
-        np.testing.assert_allclose(
-            uncached["normalizer"][key]["mean"], cached["normalizer"][key]["mean"]
-        )
-        np.testing.assert_allclose(
-            uncached["normalizer"][key]["std"], cached["normalizer"][key]["std"]
-        )
+        np.testing.assert_allclose(uncached["normalizer"][key]["mean"], cached["normalizer"][key]["mean"])
+        np.testing.assert_allclose(uncached["normalizer"][key]["std"], cached["normalizer"][key]["std"])
     assert uncached["pdg_map"] == cached["pdg_map"]
     assert uncached["mat_map"] == cached["mat_map"]
+
+
+def _fitted_cond_norm(seed=0):
+    rng = np.random.default_rng(seed)
+    return Normalizer().fit(rng.normal(size=(64, COND_DIM)).astype(np.float32))
+
+
+@pytest.mark.parametrize(
+    "router_cfg",
+    [
+        {"enabled": False, "type": "energy", "n_experts": 4},
+        {"enabled": True, "type": "pdg", "n_experts": 4},
+    ],
+)
+def test_seed_energy_router_noop_when_not_an_enabled_energy_router(router_cfg):
+    cond_norm = _fitted_cond_norm()
+    echoed = []
+    _seed_energy_router(router_cfg, cond_norm, np.array([1.0, 2.0]), 3, echoed.append)
+    assert "centers_init" not in router_cfg
+    assert echoed == []
+
+
+def test_seed_energy_router_falls_back_to_default_and_warns_when_no_samples():
+    router_cfg = {"enabled": True, "type": "energy", "n_experts": 4}
+    cond_norm = _fitted_cond_norm()
+    echoed = []
+    _seed_energy_router(router_cfg, cond_norm, np.empty(0), energy_idx=3, echo=echoed.append)
+    assert "centers_init" not in router_cfg
+    assert len(echoed) == 1
+    assert "falls back to default centers" in echoed[0]
+
+
+def test_seed_energy_router_seeds_centers_from_data_quantiles():
+    router_cfg = {"enabled": True, "type": "energy", "n_experts": 4}
+    cond_norm = _fitted_cond_norm()
+    energy_idx = 3
+    # A grid of "raw" quantile values as setup_cache.energy_quantiles_from_sample
+    # would produce them: monotonically increasing, in the same (log-energy)
+    # units as the conditioning column being normalized against.
+    energy_quantiles = np.linspace(1.0, 10.0, 33).astype(np.float32)
+    echoed = []
+    _seed_energy_router(router_cfg, cond_norm, energy_quantiles, energy_idx, echoed.append)
+
+    assert "centers_init" in router_cfg
+    centers = np.asarray(router_cfg["centers_init"], dtype=np.float32)
+    assert centers.shape == (router_cfg["n_experts"],)
+
+    levels = np.linspace(0.0, 1.0, router_cfg["n_experts"])
+    raw_centers = setup_cache.energy_quantile_at(energy_quantiles, levels)
+    expected = (raw_centers - cond_norm.mean[energy_idx]) / cond_norm.std[energy_idx]
+    np.testing.assert_allclose(centers, expected, rtol=1e-5)
+    # Quantile levels are increasing, and the normalizer's std is positive, so
+    # the seeded centers must preserve that order rather than e.g. reversing it.
+    assert np.all(np.diff(centers) > 0)
+    assert len(echoed) == 1
+    assert "seeded EnergyRouter centers" in echoed[0]

tests/test_render.py

@@ -4,10 +4,12 @@ from __future__ import annotations
 
 from pathlib import Path
 
+import numpy as np
 import pytest
 
 pytest.importorskip("plotstyle")
 
+from giant.analysis import render as render_mod  # noqa: E402
 from giant.analysis.reduced import Reduced  # noqa: E402
 
 
@@ -19,6 +21,152 @@ def _try_render(reduced: list[Reduced], out: Path):
     return render_all(out / "reduced", out / "plots")
 
 
+def test_render_router_diagnostics_and_edge_cases(tmp_path: Path):
+    reduced = [
+        Reduced(
+            "rg",
+            "router",
+            "router_gating",
+            "Router gating",
+            "pre-step energy [MeV]",
+            {
+                "n_experts": 2,
+                "log_x": True,
+                "router_type": "energy",
+                "rollout": {
+                    "centers": [1.0, 10.0, 100.0],
+                    "means": [[0.6, 0.4], [0.5, 0.5], [0.4, 0.6]],
+                },
+                "reference": {
+                    "centers": [1.0, 10.0, 100.0],
+                    "means": [[0.55, 0.45], [0.5, 0.5], [0.45, 0.55]],
+                },
+            },
+        ),
+        Reduced(
+            "rs",
+            "router",
+            "router_share",
+            "Router share",
+            "species",
+            {
+                "categories": ["e-", "gamma"],
+                "n_experts": 2,
+                "router_type": "energy",
+                "rollout": {"e-": [0.7, 0.3], "gamma": [0.2, 0.8]},
+                "reference": {"e-": [0.6, 0.4], "gamma": [0.3, 0.7]},
+            },
+        ),
+        Reduced(
+            "ru",
+            "router",
+            "unavailable",
+            "Router unavailable",
+            "x",
+            {"note": "router diagnostics unavailable: no router in this run"},
+        ),
+        Reduced(
+            "g4",
+            "marginals",
+            "grouped_hist",
+            "Grouped (4)",
+            "x",
+            {
+                "edges": [0, 1, 2],
+                "groups": {lbl: {"rollout": [1, 2], "reference": [2, 1]} for lbl in ("a", "b", "c", "d")},
+                "log_y": True,
+            },
+        ),
+        Reduced(
+            "sl",
+            "species",
+            "single_hist",
+            "Single (log-x)",
+            "x",
+            {"edges": [1, 10, 100], "rollout": [5, 1], "log_x": True, "log_y": True},
+        ),
+    ]
+    try:
+        pdfs = _try_render(reduced, tmp_path)
+    except RuntimeError as e:  # LaTeX missing at render time
+        pytest.skip(f"LaTeX rendering unavailable: {e}")
+    assert len(pdfs) == len(reduced)
+    assert all(p.exists() for p in pdfs)
+
+
+def test_render_all_run_gallery_invokes_subprocess(tmp_path: Path, monkeypatch):
+    # render_mod.subprocess *is* the stdlib subprocess module, so a blanket
+    # patch of .run would also swallow the real subprocess.run calls
+    # matplotlib's texmanager makes to compile LaTeX during savefig — only
+    # intercept the "gallery generate" call itself and pass everything else
+    # (LaTeX included) through to the real subprocess.run.
+    calls = []
+    real_run = render_mod.subprocess.run
+
+    def fake_run(*a, **k):
+        if a and a[0] and a[0][0] == "gallery":
+            calls.append((a, k))
+            return None
+        return real_run(*a, **k)
+
+    monkeypatch.setattr(render_mod.subprocess, "run", fake_run)
+    reduced = [
+        Reduced(
+            "s",
+            "species",
+            "single_hist",
+            "Single",
+            "x",
+            {"edges": [0, 1, 2], "rollout": [5, 1]},
+        )
+    ]
+    for r in reduced:
+        r.save(tmp_path / "reduced" / f"{r.id}.json")
+    try:
+        render_mod.render_all(tmp_path / "reduced", tmp_path / "plots", run_gallery=True)
+    except RuntimeError as e:
+        pytest.skip(f"LaTeX rendering unavailable: {e}")
+
+    assert len(calls) == 1
+    args, kwargs = calls[0]
+    assert args[0] == ["gallery", "generate", "--source", str(tmp_path / "plots")]
+    assert kwargs == {"check": True}
+
+
+def test_render_run_glues_condor_run_meta_into_render_all(tmp_path: Path, monkeypatch):
+    from giant.analysis import condor as condor_mod
+
+    run_dir = tmp_path / "run"
+    (run_dir / "reduced").mkdir(parents=True)
+
+    merge_calls = []
+    monkeypatch.setattr(condor_mod, "merge_all", lambda rd: merge_calls.append(Path(rd)))
+    meta = condor_mod.RunMeta(
+        rollout="rollout.parquet",
+        reference="reference.parquet",
+        run_dir=str(run_dir),
+        title="my-run",
+        plot_meta={"checkpoint": "ckpt/best.pt"},
+    )
+    monkeypatch.setattr(condor_mod.RunMeta, "load", classmethod(lambda cls, p: meta))
+
+    Reduced("s", "species", "single_hist", "Single", "x", {"edges": [0, 1], "rollout": [1]}).save(
+        run_dir / "reduced" / "s.json"
+    )
+
+    try:
+        pdfs = render_mod.render_run(run_dir)
+    except RuntimeError as e:
+        pytest.skip(f"LaTeX rendering unavailable: {e}")
+
+    assert merge_calls == [run_dir]
+    assert len(pdfs) == 1
+    plot_meta = (run_dir / "plots" / "species" / "s.yaml").read_text()
+    assert "checkpoint" in plot_meta
+    root_meta = (run_dir / "plots" / "metadata.yaml").read_text()
+    assert "my-run" in root_meta
+
+
 def test_render_one_of_each_kind(tmp_path: Path):
     reduced = [
         Reduced(
@@ -90,3 +238,133 @@ def test_render_one_of_each_kind(tmp_path: Path):
     assert len(pdfs) == len(reduced)
     assert all(p.exists() for p in pdfs)
     assert (tmp_path / "plots" / "metadata.yaml").exists()
+
+
+# ── pure-function helpers: no matplotlib figure needed ──────────────────
+
+
+def test_density_zero_total_returns_counts_unchanged():
+    counts = np.array([0.0, 0.0, 0.0])
+    out = render_mod._density(counts, np.array([0.0, 1.0, 2.0, 3.0]))
+    np.testing.assert_array_equal(out, counts)
+
+
+def test_density_normalizes_by_total_and_bin_width():
+    counts = [1, 3]
+    edges = np.array([0.0, 2.0, 4.0])  # bin width 2
+    out = render_mod._density(counts, edges)
+    np.testing.assert_allclose(out, np.array([1, 3]) / (4 * 2))
+
+
+def test_router_summary_disabled_is_off():
+    assert render_mod._router_summary({"enabled": False, "type": "energy"}) == "off"
+    assert render_mod._router_summary({}) == "off"
+
+
+def test_router_summary_enabled_formats_type_and_n_experts():
+    cfg = {"enabled": True, "type": "energy", "n_experts": 8}
+    assert render_mod._router_summary(cfg) == "energy×8"
+
+
+def test_figure_params_v2_basics_and_router_and_epoch():
+    mc = {
+        "stage1_model": {
+            "hidden_dim": 256,
+            "n_res_blocks": 4,
+            "generator": "flow",
+            "router": {"enabled": True, "type": "energy", "n_experts": 4},
+        },
+        "conditioning": {"particle": {"type": "physical"}},
+    }
+    run_meta = {"training_epoch": 12, "best_val_loss": 0.123456, "steps": 10}
+    params = render_mod._figure_params(run_meta | {"model_config": mc})
+    assert params == {
+        "hidden_dim": 256,
+        "n_res_blocks": 4,
+        "mode": "flow",
+        "conditioning": "physical",
+        "router": "energy×4",
+        "epoch": 12,
+        "best_val_loss": 0.1235,
+        "steps": 10,
+    }
+
+
+def test_figure_params_v2_wgan_reports_noise_dim_not_steps():
+    mc = {
+        "stage1_model": {
+            "generator": "wgan",
+            "wgan": {"noise_dim": 32},
+        },
+    }
+    run_meta = {"model_config": mc, "steps": 10}
+    params = render_mod._figure_params(run_meta)
+    assert params["mode"] == "wgan"
+    assert params["noise_dim"] == 32
+    assert "steps" not in params
+
+
+def test_figure_params_v2_reports_mode_s2_only_when_it_differs():
+    same = {
+        "stage1_model": {"generator": "flow"},
+        "stage2_model": {"generator": "flow"},
+    }
+    assert "mode_s2" not in render_mod._figure_params({"model_config": same})
+
+    mixed = {
+        "stage1_model": {"generator": "flow"},
+        "stage2_model": {"generator": "wgan"},
+    }
+    params = render_mod._figure_params({"model_config": mixed})
+    assert params["mode_s2"] == "wgan"
+
+
+def test_figure_params_old_shape_basics():
+    run_meta = {
+        "model_config": {
+            "hidden_dim": 128,
+            "n_blocks": 3,
+            "mode": "ddpm",
+            "conditioning": "embedding",
+            "router": {"enabled": False},
+        },
+        "training_epoch": 5,
+        "best_val_loss": 0.5,
+        "steps": 20,
+    }
+    params = render_mod._figure_params(run_meta)
+    assert params == {
+        "hidden_dim": 128,
+        "n_blocks": 3,
+        "mode": "ddpm",
+        "conditioning": "embedding",
+        "router": "off",
+        "epoch": 5,
+        "best_val_loss": 0.5,
+        "steps": 20,
+    }
+
+
+def test_figure_params_old_shape_wgan_reports_noise_dim_not_steps():
+    run_meta = {
+        "model_config": {"mode": "wgan", "noise_dim": 16},
+        "steps": 20,
+    }
+    params = render_mod._figure_params(run_meta)
+    assert params["noise_dim"] == 16
+    assert "steps" not in params
+
+
+def test_plot_metadata_includes_note_and_run_meta_parameters():
+    r = Reduced("u", "router", "unavailable", "Unavailable", "x", {"note": "no router data"})
+    meta = render_mod._plot_metadata(r, {"title": "run-1", "checkpoint": "ckpt.pt"})
+    assert meta["note"] == "no router data"
+    assert meta["parameters"] == {"checkpoint": "ckpt.pt"}
+    assert "title" not in meta["parameters"]
+
+
+def test_plot_metadata_omits_parameters_when_run_meta_empty():
+    r = Reduced("s", "species", "single_hist", "Single", "x", {"edges": [0, 1]})
+    meta = render_mod._plot_metadata(r, {})
+    assert "parameters" not in meta
+    assert "note" not in meta

tests/test_rollout.py

@@ -8,10 +8,17 @@ import numpy as np
 import pytest
 import torch
 
-from giant.constants import TERM_ESCAPED, TERM_MAX_STEPS, TERM_UNKNOWN_PDG
+from giant.config import ConditioningAxisConfig, ParticleTypeConfig
+from giant.constants import TERM_ESCAPED, TERM_MAX_STEPS, TERM_UNKNOWN_PDG, K_MAX
+from giant.data.loader import TopNMap
 from giant.data.transforms import Normalizer
-from giant.model.network import DenoisingMLP, SecondaryDecoder
-from giant.rollout import make_seed_frontier, rollout
+from giant.model.network import (
+    Stage1Model,
+    Stage2Autoregressive,
+    Stage2OneShot,
+    stage2_trunk_sec_dim,
+)
+from giant.rollout import L1DistCollector, make_seed_frontier, rollout
 
 pytest.importorskip("sklearn")
 from giant import geometry as g  # noqa: E402
@@ -21,11 +28,26 @@ MAT_MAP = {"G4_AIR": 0, "G4_PbWO4": 1}
 
 
 def _models(conditioning="embedding"):
-    s1 = DenoisingMLP(
-        pdg_vocab=3, mat_vocab=2, hidden_dim=32, n_blocks=2, conditioning=conditioning
+    particle_cfg = ConditioningAxisConfig(type=conditioning, emb_dim=16, n_layers=1)
+    material_cfg = ConditioningAxisConfig(type=conditioning, emb_dim=16, n_layers=1)
+    s1 = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        n_sec_head_k_max=K_MAX,
     )
-    s2 = SecondaryDecoder(
-        pdg_vocab=3, mat_vocab=2, hidden_dim=32, n_blocks=2, conditioning=conditioning
+    s2 = Stage2OneShot(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        generator="flow",
+        time_dim=16,
     )
     return s1.eval(), s2.eval()
 
@@ -90,7 +112,8 @@ def _run(
         batch_size=128,
         max_tracks_per_event=max_tracks_per_event,
         escape_threshold=escape_threshold,
-        conditioning=conditioning,
+        particle_conditioning=conditioning,
+        material_conditioning=conditioning,
     )
 
 
@@ -99,12 +122,8 @@ def fake_material_props(monkeypatch):
     import giant.materials as gm
 
     fake = {
-        "G4_AIR": gm.MaterialProperties(
-            z_eff=7.3, a_eff=14.4, density=1.2e-3, x0=3.0e4, lambda_int=7.0e5
-        ),
-        "G4_PbWO4": gm.MaterialProperties(
-            z_eff=75.6, a_eff=205.3, density=8.28, x0=0.89, lambda_int=20.7
-        ),
+        "G4_AIR": gm.MaterialProperties(z_eff=7.3, a_eff=14.4, density=1.2e-3, x0=3.0e4, lambda_int=7.0e5),
+        "G4_PbWO4": gm.MaterialProperties(z_eff=75.6, a_eff=205.3, density=8.28, x0=0.89, lambda_int=20.7),
     }
     monkeypatch.setattr(gm, "MATERIAL_PROPERTIES", fake)
     return fake
@@ -151,7 +170,7 @@ def test_seed_frontier_embedding_mode_skips_unresolvable_pdg_lookup():
     frontier construction — mass/charge are simply zero-filled, unused."""
     seeds = _seeds(3)
     seeds["pdg"] = np.full(3, 999999999, dtype=np.int64)
-    fr, _counts = make_seed_frontier(**seeds, conditioning="embedding")
+    fr, _counts = make_seed_frontier(**seeds, particle_conditioning="embedding")
     np.testing.assert_array_equal(fr["mass"], 0.0)
     np.testing.assert_array_equal(fr["charge"], 0.0)
 
@@ -325,3 +344,421 @@ def test_on_chunk_never_buffers_full_records():
     assert rec.termination_reason_counts == {"natural_end": 1}
     with pytest.raises(AssertionError):
         rec.to_dict()
+
+
+# ── v0.3.0 step 6: per-stage generators, AR decoder, particle_type.target ───
+
+# emb_dim=3: "other" (class idx 2) is shared by -11 and 13 (muon), matching
+# the real shape build_pdg_topn_map_from_files produces — see
+# decode_topn_class's docstring.
+PDG_TOPN_MAP = TopNMap(
+    class_map={22: 0, 11: 1, -11: 2, 13: 2},
+    other_members={-11: 5, 13: 1},
+)
+
+
+def _models_v3(
+    conditioning="physical",
+    decoder="one_shot",
+    target="physical",
+    generator1="flow",
+    generator2="flow",
+    k_max=6,
+    emb_dim=4,
+    stage2_has_n_sec_head=True,
+):
+    particle_cfg = ConditioningAxisConfig(type=conditioning, emb_dim=emb_dim, n_layers=1)
+    material_cfg = ConditioningAxisConfig(type=conditioning, emb_dim=emb_dim, n_layers=1)
+    # A fresh v0.3.0 Stage1Model — no n_sec_head_k_max, unlike _models() above
+    # (n_sec ownership moves to stage 2 by default).
+    s1 = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        generator=generator1,
+        noise_dim=8,
+    )
+    particle_type_cfg = ParticleTypeConfig(target=target)
+    # Explicit kwargs rather than a shared **common dict: a dict() call whose
+    # values have heterogeneous types (str/int/dict/bool) widens under static
+    # analysis to dict[str, <big union>], which then makes every constructor
+    # keyword not itself part of that union (router, cond_enc, ...) look like
+    # a type mismatch to `ty` even though every actual value passed is fine.
+    if decoder == "one_shot":
+        sec_dim = stage2_trunk_sec_dim(particle_type_cfg, generator2, k_max, emb_dim)
+        s2 = Stage2OneShot(
+            pdg_vocab=3,
+            mat_vocab=2,
+            particle_cfg=particle_cfg,
+            material_cfg=material_cfg,
+            hidden_dim=32,
+            n_res_blocks=2,
+            generator=generator2,
+            time_dim=16,
+            noise_dim=8,
+            k_max=k_max,
+            particle_type_cfg=particle_type_cfg,
+            build_n_sec_head=stage2_has_n_sec_head,
+            sec_dim=sec_dim,
+        )
+    else:
+        s2 = Stage2Autoregressive(
+            pdg_vocab=3,
+            mat_vocab=2,
+            particle_cfg=particle_cfg,
+            material_cfg=material_cfg,
+            hidden_dim=32,
+            n_res_blocks=2,
+            generator=generator2,
+            time_dim=16,
+            noise_dim=8,
+            k_max=k_max,
+            particle_type_cfg=particle_type_cfg,
+            build_n_sec_head=stage2_has_n_sec_head,
+        )
+    return s1.eval(), s2.eval()
+
+
+def _run_v3(
+    s1,
+    s2,
+    escape_threshold=1e9,
+    energy_cutoff=1.0,
+    max_steps=15,
+    max_tracks_per_event=100,
+    seeds=None,
+    conditioning="physical",
+    sec_type_topn_map=None,
+    other_policy="sample",
+    seed=0,
+    stage1_ddpm_steps=1000,
+    l1_dist_collector=None,
+):
+    torch.manual_seed(0)
+    np.random.seed(0)
+    cond, tgt, sec_phys = _norms()
+    return rollout(
+        s1,
+        s2,
+        _oracle(),
+        seeds or _seeds(),
+        cond,
+        tgt,
+        sec_phys,
+        PDG_MAP,
+        MAT_MAP,
+        energy_cutoff=energy_cutoff,
+        max_steps=max_steps,
+        steps=3,
+        batch_size=128,
+        max_tracks_per_event=max_tracks_per_event,
+        escape_threshold=escape_threshold,
+        particle_conditioning=conditioning,
+        material_conditioning=conditioning,
+        sec_type_topn_map=sec_type_topn_map,
+        other_policy=other_policy,
+        seed=seed,
+        stage1_ddpm_steps=stage1_ddpm_steps,
+        l1_dist_collector=l1_dist_collector,
+    )
+
+
+def test_rollout_stage2_owns_n_sec_when_stage1_has_no_head(fake_material_props):
+    """A fresh v0.3.0 Stage1Model has no n_sec_head — n_sec must come from
+    Stage2's own head instead, and the run must still complete and
+    conserve energy."""
+    s1, s2 = _models_v3()
+    rec = _run_v3(s1, s2)
+    assert len(rec["event_id"]) > 0
+    seeds = _seeds()
+    for i, ev in enumerate(seeds["event_id"]):
+        m = rec["event_id"] == ev
+        dep = rec["edep"][m].sum()
+        leak = rec["pre_E"][m & (rec["termination_reason"] == TERM_ESCAPED)].sum()
+        assert dep + leak == pytest.approx(seeds["pre_E"][i], rel=1e-4)
+
+
+def test_resolve_n_sec_raises_when_neither_stage_owns_head(fake_material_props):
+    """Neither stage owning n_sec_head only happens for a
+    stage2_model.n_sec.mode other than "head" — not a valid rollout-capable
+    checkpoint, and must fail with a clear error rather than crash deep
+    inside predict_n_sec."""
+    s1, s2 = _models_v3(stage2_has_n_sec_head=False)
+    with pytest.raises(RuntimeError, match="n_sec_head"):
+        _run_v3(s1, s2)
+
+
+@pytest.mark.parametrize("decoder", ["one_shot", "autoregressive"])
+@pytest.mark.parametrize("generator2", ["flow", "wgan"])
+def test_rollout_physical_target_decoder_generator_matrix(fake_material_props, decoder, generator2):
+    """Every (decoder, stage2 generator) combination under
+    particle_type.target="physical" must run to completion and conserve
+    energy."""
+    s1, s2 = _models_v3(decoder=decoder, generator2=generator2)
+    rec = _run_v3(s1, s2)
+    assert len(rec["event_id"]) > 0
+    seeds = _seeds()
+    for i, ev in enumerate(seeds["event_id"]):
+        m = rec["event_id"] == ev
+        dep = rec["edep"][m].sum()
+        leak = rec["pre_E"][m & (rec["termination_reason"] == TERM_ESCAPED)].sum()
+        assert dep + leak == pytest.approx(seeds["pre_E"][i], rel=1e-4)
+
+
+def test_sample_stage1_dispatches_ddpm_by_generator_kind():
+    """stage1_model.generator="ddpm" must be dispatched to sample_ddpm
+    (previously _step_chunk silently fell through to the flow ODE sampler
+    regardless of the checkpoint's actual generator — see giant.sample.sample_stage1).
+    A short T avoids the reverse-diffusion numerical blowup an untrained,
+    random-weight network produces over many steps; that instability is a
+    property of sampling from an untrained net, not of the dispatch logic
+    under test here, so a full oracle-driven rollout isn't needed."""
+    from giant.sample import sample_stage1
+
+    s1, _ = _models_v3(generator1="ddpm")
+    cond_cont = torch.randn(6, 15)
+    cond_cat = torch.zeros(6, 2, dtype=torch.long)
+    sample, n_sec = sample_stage1(s1, cond_cont, cond_cat, steps=10, ddpm_steps=5)
+    assert sample.shape == (6, 9)
+    assert n_sec is None  # fresh v0.3.0 Stage1Model owns no n_sec_head
+
+
+@pytest.mark.parametrize("decoder", ["one_shot", "autoregressive"])
+def test_rollout_onehot_target_end_to_end(fake_material_props, decoder):
+    """particle_type.target="onehot" resolves a concrete PDG via
+    decode_topn_class (argmax + other_policy), and that PDG's real physics
+    (giant.particles.particle_phys_array) become the secondary's identity —
+    unlike "physical", not just a reporting label."""
+    s1, s2 = _models_v3(decoder=decoder, target="onehot", emb_dim=3)
+    rec = _run_v3(s1, s2, sec_type_topn_map=PDG_TOPN_MAP, other_policy="modal")
+    assert len(rec["event_id"]) > 0
+    # Every spawned secondary's nominal pdg must be one decode_topn_class can
+    # actually produce (the topn map's known classes + its "other" members).
+    possible = set(PDG_TOPN_MAP.class_map.keys()) | set(PDG_TOPN_MAP.other_members.keys())
+    secondary_pdgs = set(rec["pdg"][rec["generation"] > 0].tolist())
+    assert secondary_pdgs <= possible
+
+
+def test_rollout_onehot_target_missing_topn_map_raises(fake_material_props):
+    s1, s2 = _models_v3(target="onehot", emb_dim=3)
+    with pytest.raises(RuntimeError, match="sec_type_topn_map"):
+        _run_v3(s1, s2, sec_type_topn_map=None)
+
+
+# --- conditioning.{particle,material}.type = "onehot" — a separate axis from
+# stage2_model.particle_type.target above: this is what feeds cond_cat's
+# extra top-N columns for ConditionEncoder's own "onehot" mode, not the
+# secondary-species decode. ---------------------------------------------
+
+COND_PDG_TOPN_MAP = TopNMap(class_map=dict(PDG_MAP), other_members={})
+COND_MAT_TOPN_MAP = TopNMap(class_map={"G4_AIR": 0, "G4_PbWO4": 1}, other_members={})
+
+
+def _onehot_conditioning_models():
+    particle_cfg = ConditioningAxisConfig(type="onehot", emb_dim=len(PDG_MAP), n_layers=1)
+    material_cfg = ConditioningAxisConfig(type="onehot", emb_dim=len(MAT_MAP), n_layers=1)
+    s1 = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+    )
+    s2 = Stage2OneShot(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        sec_dim=stage2_trunk_sec_dim(ParticleTypeConfig(target="physical"), "flow", K_MAX, 3),
+        generator="flow",
+        time_dim=16,
+    )
+    return s1.eval(), s2.eval()
+
+
+def _run_onehot_conditioning(pdg_topn_map=COND_PDG_TOPN_MAP, mat_topn_map=COND_MAT_TOPN_MAP):
+    s1, s2 = _onehot_conditioning_models()
+    cond, tgt, sec_phys = _norms()
+    return rollout(
+        s1,
+        s2,
+        _oracle(),
+        _seeds(),
+        cond,
+        tgt,
+        sec_phys,
+        PDG_MAP,
+        MAT_MAP,
+        energy_cutoff=1.0,
+        max_steps=30,
+        steps=4,
+        batch_size=128,
+        max_tracks_per_event=300,
+        escape_threshold=1e9,
+        particle_conditioning="onehot",
+        material_conditioning="onehot",
+        pdg_topn_map=pdg_topn_map,
+        mat_topn_map=mat_topn_map,
+    )
+
+
+def test_rollout_conditioning_onehot_end_to_end(fake_material_props):
+    rec = _run_onehot_conditioning()
+    assert len(rec["event_id"]) > 0
+    assert set(rec["event_id"].tolist()) == set(range(6))
+
+
+def test_rollout_conditioning_onehot_particle_missing_topn_map_raises(
+    fake_material_props,
+):
+    with pytest.raises(RuntimeError, match="pdg_topn_map"):
+        _run_onehot_conditioning(pdg_topn_map=None)
+
+
+def test_rollout_conditioning_onehot_material_missing_topn_map_raises(
+    fake_material_props,
+):
+    with pytest.raises(RuntimeError, match="mat_topn_map"):
+        _run_onehot_conditioning(mat_topn_map=None)
+
+
+SEC_TYPE_TOPN_MAP_DIFFERENT_N = TopNMap(class_map={22: 0, 11: 1, -11: 2, 13: 3}, other_members={2112: 3, 2212: 1})
+
+
+def _run_conditioning_and_type_onehot_different_n_classes():
+    """Both conditioning.particle.type="onehot" and
+    stage2_model.particle_type.target="onehot" active at once, with
+    stage2_model.particle_type.n_classes deliberately different from
+    conditioning.particle.emb_dim (gitea #29)."""
+    cond_emb_dim = len(PDG_MAP)  # 3
+    type_n_classes = 5  # deliberately different from cond_emb_dim
+    particle_cfg = ConditioningAxisConfig(type="onehot", emb_dim=cond_emb_dim, n_layers=1)
+    material_cfg = ConditioningAxisConfig(type="onehot", emb_dim=len(MAT_MAP), n_layers=1)
+    particle_type_cfg = ParticleTypeConfig(target="onehot", n_classes=type_n_classes)
+
+    s1 = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+    ).eval()
+    sec_dim = stage2_trunk_sec_dim(particle_type_cfg, "flow", K_MAX, type_n_classes)
+    s2 = Stage2OneShot(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        sec_dim=sec_dim,
+        generator="flow",
+        time_dim=16,
+        k_max=K_MAX,
+        particle_type_cfg=particle_type_cfg,
+    ).eval()
+    # Sanity: the model's own type_dim followed n_classes, not cond_emb_dim.
+    assert s2.type_dim == type_n_classes
+
+    cond, tgt, sec_phys = _norms()
+    return rollout(
+        s1,
+        s2,
+        _oracle(),
+        _seeds(),
+        cond,
+        tgt,
+        sec_phys,
+        PDG_MAP,
+        MAT_MAP,
+        energy_cutoff=1.0,
+        max_steps=15,
+        steps=3,
+        batch_size=128,
+        max_tracks_per_event=100,
+        escape_threshold=1e9,
+        particle_conditioning="onehot",
+        material_conditioning="onehot",
+        pdg_topn_map=COND_PDG_TOPN_MAP,
+        mat_topn_map=COND_MAT_TOPN_MAP,
+        sec_type_topn_map=SEC_TYPE_TOPN_MAP_DIFFERENT_N,
+        other_policy="modal",
+    )
+
+
+def test_rollout_conditioning_and_type_onehot_with_different_n_classes(fake_material_props):
+    """gitea #29 end-to-end: conditioning.particle.type="onehot" and
+    stage2_model.particle_type.target="onehot" now use independently sized
+    top-N maps (stage2_model.particle_type.n_classes != conditioning.particle
+    .emb_dim), and rollout must decode secondaries using the type-side map,
+    not silently reuse the conditioning-side one (the pre-#29 bug)."""
+    rec = _run_conditioning_and_type_onehot_different_n_classes()
+    assert len(rec["event_id"]) > 0
+    possible = set(SEC_TYPE_TOPN_MAP_DIFFERENT_N.class_map.keys()) | set(
+        SEC_TYPE_TOPN_MAP_DIFFERENT_N.other_members.keys()
+    )
+    secondary_pdgs = set(rec["pdg"][rec["generation"] > 0].tolist())
+    assert secondary_pdgs <= possible
+
+
+@pytest.mark.parametrize("decoder", ["one_shot", "autoregressive"])
+def test_rollout_embedding_target_end_to_end(decoder):
+    """particle_type.target="embedding" L1-snaps to the nearest row of the
+    conditioning's own particle embedding table, so every resolved PDG must
+    be a real member of the dense training vocab (pdg_map) — unlike
+    "onehot", there is no "other" bucket to fall outside of."""
+    s1, s2 = _models_v3(conditioning="embedding", decoder=decoder, target="embedding", emb_dim=4)
+    rec = _run_v3(s1, s2, conditioning="embedding")
+    assert len(rec["event_id"]) > 0
+    secondary_pdgs = set(rec["pdg"][rec["generation"] > 0].tolist())
+    assert secondary_pdgs <= set(PDG_MAP.keys())
+
+
+def test_l1_dist_collector_populated_only_for_embedding_target():
+    """The L1-distance diagnostic only makes sense under
+    particle_type.target="embedding" — a physical-target run must leave the
+    collector empty rather than silently accumulating garbage."""
+    s1, s2 = _models_v3(target="physical")
+    collector = L1DistCollector()
+    _run_v3(s1, s2, l1_dist_collector=collector)
+    assert collector.n == 0
+    assert collector.summary() is None
+
+
+def test_l1_dist_collector_accumulates_for_embedding_target():
+    s1, s2 = _models_v3(conditioning="embedding", target="embedding", emb_dim=4)
+    collector = L1DistCollector()
+    rec = _run_v3(s1, s2, conditioning="embedding", l1_dist_collector=collector)
+    n_secondaries = int((rec["generation"] > 0).sum())
+    assert n_secondaries > 0  # sanity: the tiny model does spawn secondaries
+    summary = collector.summary()
+    assert summary is not None
+    assert summary["n"] == collector.n > 0
+    assert summary["min"] <= summary["mean"] <= summary["max"]
+    assert summary["std"] >= 0.0
+    assert len(summary["hist_edges"]) == len(summary["hist_counts"]) + 1
+    assert sum(summary["hist_counts"]) <= summary["n"]  # some may fall outside [lo, hi)
+
+
+def test_l1_dist_collector_add_ignores_invalid_slots():
+    collector = L1DistCollector()
+    dist = np.array([[1.0, 5.0, 9.0]])
+    valid = np.array([[True, False, True]])
+    collector.add(dist, valid)
+    assert collector.n == 2
+    assert collector.minimum == 1.0
+    assert collector.maximum == 9.0
+
+
+def test_l1_dist_collector_add_empty_is_noop():
+    collector = L1DistCollector()
+    collector.add(np.zeros((0, 3)), np.zeros((0, 3), dtype=bool))
+    assert collector.n == 0
+    assert collector.summary() is None

tests/test_router.py

@@ -3,50 +3,66 @@
 import pytest
 import torch
 
+from giant.config import ConditioningAxisConfig
 from giant.constants import COND_DIM, K_MAX, SEC_DIM, X_DIM
 from giant.model.network import (
+    BLOCK_REGISTRY,
+    TRUNK_REGISTRY,
+    AdaLNResBlock,
     ComposedRouter,
-    DenoisingMLP,
     EnergyRouter,
+    ExpertTrunk,
+    FilmResBlock,
     PdgRouter,
     ProcessRouter,
     ROUTER_REGISTRY,
-    RoutedDenoisingMLP,
-    RoutedSecondaryDecoder,
-    SecondaryDecoder,
+    ResBlock,
+    RoutedTrunk,
+    Stage1Model,
+    Stage2OneShot,
+    build_block,
     build_composed_router,
+    build_expert_body,
     build_models,
     build_router,
 )
 
+PARTICLE_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+MATERIAL_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+
 
 def _cond(B=8, pdg=3, mat=2):
     cond_cont = torch.randn(B, COND_DIM)
-    cond_cat = torch.stack(
-        [torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1
-    )
+    cond_cat = torch.stack([torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1)
     return cond_cont, cond_cat
 
 
 def _routed_stage1(n_experts=4, pdg=3, mat=2, **router_kwargs):
     router = build_router("energy", n_experts, **router_kwargs)
-    return RoutedDenoisingMLP(
+    return Stage1Model(
         pdg_vocab=pdg,
         mat_vocab=mat,
+        particle_cfg=PARTICLE_CFG,
+        material_cfg=MATERIAL_CFG,
+        hidden_dim=16,
+        n_res_blocks=2,
         router=router,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
+        n_sec_head_k_max=K_MAX,
     )
 
 
 def _routed_sec_decoder(n_experts=4, pdg=3, mat=2, **router_kwargs):
     router = build_router("energy", n_experts, **router_kwargs)
-    return RoutedSecondaryDecoder(
+    return Stage2OneShot(
         pdg_vocab=pdg,
         mat_vocab=mat,
+        particle_cfg=PARTICLE_CFG,
+        material_cfg=MATERIAL_CFG,
+        hidden_dim=16,
+        n_res_blocks=2,
+        generator="flow",
+        time_dim=16,
         router=router,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
     )
 
 
@@ -68,9 +84,7 @@ def test_energy_router_gate_partition_of_unity():
 def test_energy_router_top1_matches_gate_argmax():
     router = EnergyRouter(n_experts=4)
     cond_cont, cond_cat = _cond(16)
-    assert torch.equal(
-        router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1)
-    )
+    assert torch.equal(router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1))
 
 
 def test_energy_router_hardens_as_temperature_shrinks():
@@ -92,7 +106,7 @@ def test_energy_router_balance_loss_is_nonnegative_scalar():
 
 
 def test_build_router_ignores_unrecognized_kwargs():
-    # lambda_balance is a model_config.router key but not an EnergyRouter kwarg
+    # lambda_balance is a router config key but not an EnergyRouter kwarg
     router = build_router("energy", 4, temperature=0.3, lambda_balance=0.5)
     assert isinstance(router, EnergyRouter)
     assert router.temperature == 0.3
@@ -118,12 +132,8 @@ def test_energy_router_centers_init_wrong_length_raises():
 
 
 def test_energy_router_centers_init_respects_learn_centers_flag():
-    learned = EnergyRouter(
-        n_experts=3, centers_init=[-1.0, 0.0, 1.0], learn_centers=True
-    )
-    fixed = EnergyRouter(
-        n_experts=3, centers_init=[-1.0, 0.0, 1.0], learn_centers=False
-    )
+    learned = EnergyRouter(n_experts=3, centers_init=[-1.0, 0.0, 1.0], learn_centers=True)
+    fixed = EnergyRouter(n_experts=3, centers_init=[-1.0, 0.0, 1.0], learn_centers=False)
     assert isinstance(learned.centers, torch.nn.Parameter)
     assert not isinstance(fixed.centers, torch.nn.Parameter)
 
@@ -142,6 +152,68 @@ def test_build_router_unknown_type_raises():
     raise AssertionError("expected ValueError for unknown router type")
 
 
+# ── TRUNK_REGISTRY / build_expert_body ──────────────────────────────────────
+
+
+def test_trunk_registry_has_resmlp():
+    assert "resmlp" in TRUNK_REGISTRY
+    assert TRUNK_REGISTRY["resmlp"] is ExpertTrunk
+
+
+def test_build_expert_body_unknown_type_raises():
+    try:
+        build_expert_body("nonexistent", in_dim=4, out_dim=4, hidden_dim=8, n_blocks=1, cond_dim=4)
+    except ValueError:
+        return
+    raise AssertionError("expected ValueError for unknown trunk type")
+
+
+# ── BLOCK_REGISTRY / build_block (gitea #34) ────────────────────────────────
+
+
+def test_block_registry_has_add_film_adaln():
+    assert BLOCK_REGISTRY["add"] is ResBlock
+    assert BLOCK_REGISTRY["film"] is FilmResBlock
+    assert BLOCK_REGISTRY["adaln"] is AdaLNResBlock
+
+
+def test_build_block_unknown_type_raises():
+    try:
+        build_block("nonexistent", dim=8, cond_dim=4)
+    except ValueError:
+        return
+    raise AssertionError("expected ValueError for unknown block conditioning type")
+
+
+@pytest.mark.parametrize("block_type", ["add", "film", "adaln"])
+def test_block_forward_shape(block_type):
+    block = build_block(block_type, dim=8, cond_dim=4)
+    x = torch.randn(5, 8)
+    cond = torch.randn(5, 4)
+    out = block(x, cond)
+    assert out.shape == (5, 8)
+
+
+def test_film_res_block_output_invariant_to_cond_at_init():
+    """Zero-initialized film_proj means gamma=beta=0 at construction, so the
+    output must not depend on which cond is passed in."""
+    block = FilmResBlock(dim=8, cond_dim=4)
+    x = torch.randn(5, 8)
+    cond_a = torch.randn(5, 4)
+    cond_b = torch.randn(5, 4)
+    torch.testing.assert_close(block(x, cond_a), block(x, cond_b))
+
+
+def test_adaln_res_block_is_identity_at_init():
+    """Zero-initialized adaln_proj means scale=shift=gate=0 at construction,
+    so the block must be the exact identity function (the 'Zero' in
+    AdaLN-Zero)."""
+    block = AdaLNResBlock(dim=8, cond_dim=4)
+    x = torch.randn(5, 8)
+    cond = torch.randn(5, 4)
+    torch.testing.assert_close(block(x, cond), x)
+
+
 # ── EnergyRouter learn_width / learn_temperature ────────────────────────────
 
 
@@ -150,9 +222,7 @@ def test_energy_router_learn_width_matches_fixed_temperature_at_init():
     per-expert width must reproduce the fixed-temperature gate exactly."""
     centers_init = [-1.0, 0.0, 0.5, 1.5]
     fixed = EnergyRouter(n_experts=4, temperature=0.3, centers_init=centers_init)
-    learned = EnergyRouter(
-        n_experts=4, temperature=0.3, centers_init=centers_init, learn_width=True
-    )
+    learned = EnergyRouter(n_experts=4, temperature=0.3, centers_init=centers_init, learn_width=True)
     cond_cont, cond_cat = _cond(16)
     torch.testing.assert_close(
         learned.gate(cond_cont, cond_cat),
@@ -185,21 +255,15 @@ def test_energy_router_learn_width_and_temperature_mutually_exclusive_raises():
         EnergyRouter(n_experts=4, learn_width=True, learn_temperature=True)
     except ValueError:
         return
-    raise AssertionError(
-        "expected ValueError for learn_width and learn_temperature both set"
-    )
+    raise AssertionError("expected ValueError for learn_width and learn_temperature both set")
 
 
 def test_energy_router_width_ratio_bounds_must_bracket_one_raises():
     try:
-        EnergyRouter(
-            n_experts=4, learn_width=True, width_min_ratio=1.0, width_max_ratio=2.0
-        )
+        EnergyRouter(n_experts=4, learn_width=True, width_min_ratio=1.0, width_max_ratio=2.0)
     except ValueError:
         return
-    raise AssertionError(
-        "expected ValueError for width_min_ratio/width_max_ratio not bracketing 1.0"
-    )
+    raise AssertionError("expected ValueError for width_min_ratio/width_max_ratio not bracketing 1.0")
 
 
 def test_energy_router_effective_width_stays_within_bounds():
@@ -236,9 +300,7 @@ def test_energy_router_learn_width_hardens_when_pushed_to_floor():
     """Pushing every expert's width toward the (tiny) floor should harden the
     gate to a one-hot at the nearest center, generalizing the fixed-
     temperature->0 hardening test to the per-expert path."""
-    router = EnergyRouter(
-        n_experts=4, learn_width=True, width_min_ratio=1e-4, width_max_ratio=10.0
-    )
+    router = EnergyRouter(n_experts=4, learn_width=True, width_min_ratio=1e-4, width_max_ratio=10.0)
     with torch.no_grad():
         router.raw_width.fill_(-1e6)
     cond_cont, cond_cat = _cond(16)
@@ -254,9 +316,7 @@ def test_energy_router_own_width_controls_own_coverage_independent_of_others():
     expert's own gate share, without needing to touch any other expert's
     width — the "each expert learns its own coverage independently" property
     this feature is meant to add."""
-    router = EnergyRouter(
-        n_experts=2, temperature=1.0, learn_width=True, centers_init=[0.0, 10.0]
-    )
+    router = EnergyRouter(n_experts=2, temperature=1.0, learn_width=True, centers_init=[0.0, 10.0])
     cond_cont, cond_cat = _cond(4)
     cond_cont[:, 3] = 3.0  # fixed energy, unequal distance to each center
 
@@ -270,9 +330,7 @@ def test_energy_router_own_width_controls_own_coverage_independent_of_others():
 
 
 def test_build_router_threads_learn_width_kwargs_through():
-    router = build_router(
-        "energy", 4, learn_width=True, width_min_ratio=0.2, width_max_ratio=8.0
-    )
+    router = build_router("energy", 4, learn_width=True, width_min_ratio=0.2, width_max_ratio=8.0)
     assert isinstance(router, EnergyRouter)
     assert router.learn_width is True
     assert isinstance(router.raw_width, torch.nn.Parameter)
@@ -375,18 +433,18 @@ def test_build_router_from_cfg_sets_gumbel_for_composed_router():
     assert router.gumbel is True
 
 
-def test_routed_denoising_mlp_forward_runs_with_gumbel_enabled():
+def test_routed_stage1_forward_runs_with_gumbel_enabled():
     """End-to-end forward through _route_forward's train branch with
     straight-through Gumbel-softmax combine weights enabled."""
     B = 8
     model = _routed_stage1(n_experts=3)
-    model.router.gumbel = True
-    model.router.gumbel_tau = 0.5
+    model.trunk.router.gumbel = True
+    model.trunk.router.gumbel_tau = 0.5
     model.train()
     x_t = torch.randn(B, X_DIM)
     t = torch.rand(B)
     cond_cont, cond_cat = _cond(B)
-    out = model(x_t, t, cond_cont, cond_cat)
+    out = model(x_t, cond_cont, cond_cat, t=t)
     assert out.shape == (B, X_DIM)
     assert torch.isfinite(out).all()
 
@@ -409,9 +467,7 @@ def test_pdg_router_gate_partition_of_unity():
 def test_pdg_router_top1_matches_gate_argmax():
     router = PdgRouter(n_experts=4, pdg_vocab=3)
     cond_cont, cond_cat = _cond(16)
-    assert torch.equal(
-        router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1)
-    )
+    assert torch.equal(router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1))
 
 
 def test_pdg_router_hardens_as_temperature_shrinks():
@@ -463,59 +519,89 @@ def test_build_router_pdg_type_uses_pdg_vocab():
     assert router.pdg_emb.num_embeddings == 5
 
 
+def _nested_cfg(
+    pdg_vocab,
+    mat_vocab,
+    stage1_router=None,
+    stage2_router=None,
+    particle_type="physical",
+    material_type="physical",
+    **overrides,
+):
+    """Minimal new-shape (v0.3.0) model_config for build_models, with
+    optional router sub-blocks. `overrides` deep-patches stage1_model."""
+    stage1_model = {
+        "active": True,
+        "generator": "flow",
+        "hidden_dim": 16,
+        "n_res_blocks": 2,
+        "dropout": 0.0,
+        "flow": {"time_dim": 16},
+        "router": stage1_router or {"enabled": False},
+    }
+    stage1_model.update(overrides)
+    return {
+        "pdg_vocab": pdg_vocab,
+        "mat_vocab": mat_vocab,
+        "conditioning": {
+            "out_dim": 32,
+            "particle": {"type": particle_type, "emb_dim": 8, "n_layers": 1},
+            "material": {"type": material_type, "emb_dim": 8, "n_layers": 1},
+        },
+        "stage1_model": stage1_model,
+        "stage2_model": {
+            "active": True,
+            "decoder": "one_shot",
+            "generator": "flow",
+            "hidden_dim": 16,
+            "n_res_blocks": 2,
+            "dropout": 0.0,
+            "k_max": K_MAX,
+            "context_dim": 16,
+            "n_sec": {"mode": "head"},
+            "flow": {"time_dim": 16},
+            "router": stage2_router or {"enabled": False, "tie_to_stage1": False},
+        },
+    }
+
+
 def test_build_models_routed_with_pdg_router():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        router={
-            "enabled": True,
-            "type": "pdg",
-            "n_experts": 3,
-        },
+        particle_type="embedding",
+        material_type="embedding",
+        stage1_router={"enabled": True, "type": "pdg", "n_experts": 3},
     )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, RoutedDenoisingMLP)
-    assert isinstance(stage1.router, PdgRouter)
-    assert len(stage1.experts) == 3
-    assert stage1.router.pdg_emb.num_embeddings == 4
+    models = build_models(cfg)
+    stage1 = models["stage1"]
+    assert isinstance(stage1, Stage1Model)
+    assert isinstance(stage1.trunk, RoutedTrunk)
+    assert isinstance(stage1.trunk.router, PdgRouter)
+    assert len(stage1.trunk.experts) == 3
+    assert stage1.trunk.router.pdg_emb.num_embeddings == 4
 
 
 def test_build_models_rejects_pdg_router_with_physical_conditioning():
-    """conditioning="physical" is meant to generalize beyond the training PDG
-    vocab; PdgRouter always uses a training-vocab nn.Embedding regardless of
-    conditioning, so the combination must raise rather than silently building
-    a model that can't actually generalize the way it claims to."""
-    model_config = dict(
+    """conditioning.particle.type="physical" is meant to generalize beyond the
+    training PDG vocab; PdgRouter always uses a training-vocab nn.Embedding
+    regardless of conditioning, so the combination must raise rather than
+    silently building a model that can't actually generalize the way it
+    claims to."""
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        conditioning="physical",
-        router={"enabled": True, "type": "pdg", "n_experts": 3},
+        stage1_router={"enabled": True, "type": "pdg", "n_experts": 3},
     )
     with pytest.raises(ValueError, match="physical"):
-        build_models(model_config)
+        build_models(cfg)
 
 
 def test_build_models_rejects_composed_router_with_pdg_axis_and_physical_conditioning():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        conditioning="physical",
-        router={
+        stage1_router={
             "enabled": True,
             "type": "composed",
             "axis0_type": "energy",
@@ -525,7 +611,7 @@ def test_build_models_rejects_composed_router_with_pdg_axis_and_physical_conditi
         },
     )
     with pytest.raises(ValueError, match="physical"):
-        build_models(model_config)
+        build_models(cfg)
 
 
 # ── ProcessRouter ────────────────────────────────────────────────────────────
@@ -546,9 +632,7 @@ def test_process_router_gate_partition_of_unity():
 def test_process_router_top1_matches_gate_argmax():
     router = ProcessRouter(n_experts=4, pdg_vocab=3, mat_vocab=2)
     cond_cont, cond_cat = _cond(16)
-    assert torch.equal(
-        router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1)
-    )
+    assert torch.equal(router.top1(cond_cont, cond_cat), router.gate(cond_cont, cond_cat).argmax(-1))
 
 
 def test_process_router_balance_loss_is_nonnegative_scalar():
@@ -597,43 +681,38 @@ def test_build_router_process_type_uses_pdg_mat_vocab():
 
 
 def test_build_models_routed_with_process_router():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        router={
+        particle_type="embedding",
+        material_type="embedding",
+        stage1_router={
             "enabled": True,
             "type": "process",
             "n_experts": 3,
             "lambda_proc": 1.0,
         },
     )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, RoutedDenoisingMLP)
-    assert isinstance(stage1.router, ProcessRouter)
-    assert len(stage1.experts) == 3
-    assert stage1.router.pdg_emb.num_embeddings == 4
-    assert stage1.router.mat_emb.num_embeddings == 2
+    models = build_models(cfg)
+    stage1 = models["stage1"]
+    assert isinstance(stage1, Stage1Model)
+    assert isinstance(stage1.trunk, RoutedTrunk)
+    assert isinstance(stage1.trunk.router, ProcessRouter)
+    assert len(stage1.trunk.experts) == 3
+    assert stage1.trunk.router.pdg_emb.num_embeddings == 4
+    assert stage1.trunk.router.mat_emb.num_embeddings == 2
 
 
 # ── ComposedRouter ───────────────────────────────────────────────────────────
 
 
 def test_composed_router_n_experts_is_product():
-    router = ComposedRouter(
-        [EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)]
-    )
+    router = ComposedRouter([EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)])
     assert router.n_experts == 12
 
 
 def test_composed_router_gate_partition_of_unity():
-    router = ComposedRouter(
-        [EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)]
-    )
+    router = ComposedRouter([EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)])
     cond_cont, cond_cat = _cond(16, pdg=5)
     g = router.gate(cond_cont, cond_cat)
     assert g.shape == (16, 12)
@@ -670,9 +749,7 @@ def test_composed_router_top1_factors_into_per_axis_argmax():
 
 
 def test_composed_router_supports_different_expert_counts_per_axis():
-    router = ComposedRouter(
-        [EnergyRouter(n_experts=5), PdgRouter(n_experts=2, pdg_vocab=5)]
-    )
+    router = ComposedRouter([EnergyRouter(n_experts=5), PdgRouter(n_experts=2, pdg_vocab=5)])
     assert router.n_experts == 10
     cond_cont, cond_cat = _cond(8, pdg=5)
     assert router.gate(cond_cont, cond_cat).shape == (8, 10)
@@ -680,9 +757,7 @@ def test_composed_router_supports_different_expert_counts_per_axis():
 
 def test_composed_router_classify_loss_sums_sub_router_losses():
     """energy/pdg both default to zero, so the composed loss should too."""
-    router = ComposedRouter(
-        [EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)]
-    )
+    router = ComposedRouter([EnergyRouter(n_experts=4), PdgRouter(n_experts=3, pdg_vocab=5)])
     cond_cont, cond_cat = _cond(16, pdg=5)
     labels = torch.randint(0, 4, (16,))
     loss = router.classify_loss(cond_cont, cond_cat, labels)
@@ -777,15 +852,12 @@ def test_build_composed_router_resolves_per_axis_specs():
 
 
 def test_build_models_routed_with_composed_router():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=5,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        router={
+        particle_type="embedding",
+        material_type="embedding",
+        stage1_router={
             "enabled": True,
             "type": "composed",
             "axis0_type": "energy",
@@ -793,29 +865,58 @@ def test_build_models_routed_with_composed_router():
             "axis1_type": "pdg",
             "axis1_n_experts": 3,
         },
+        stage2_router={
+            "enabled": True,
+            "tie_to_stage1": False,
+            "type": "composed",
+            "axis0_type": "energy",
+            "axis0_n_experts": 4,
+            "axis1_type": "pdg",
+            "axis1_n_experts": 3,
+        },
     )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, RoutedDenoisingMLP)
-    assert isinstance(stage1.router, ComposedRouter)
-    assert len(stage1.experts) == 12
-    assert len(sec_decoder.experts) == 12
-    # stage1 and sec_decoder must not share router weights (same convention
-    # as the single-axis routers built by build_models).
-    assert stage1.router is not sec_decoder.router
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
+    assert isinstance(stage1.trunk, RoutedTrunk)
+    assert isinstance(stage1.trunk.router, ComposedRouter)
+    assert len(stage1.trunk.experts) == 12
+    assert len(stage2.trunk.experts) == 12
+    # stage1 and stage2 must not share router weights when tie_to_stage1 is
+    # false (same convention as v0.2's two-independent-routers behaviour).
+    assert stage1.trunk.router is not stage2.trunk.router
+
+
+def test_build_models_routed_stage2_ties_to_stage1_router():
+    cfg = _nested_cfg(
+        pdg_vocab=4,
+        mat_vocab=2,
+        stage1_router={"enabled": True, "type": "energy", "n_experts": 3},
+        stage2_router={
+            "enabled": True,
+            "tie_to_stage1": True,
+            "type": "energy",
+            "n_experts": 3,
+        },
+    )
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
+    assert stage1.trunk.router is stage2.trunk.router
 
 
 def test_build_models_routed_pair_composed_router_is_drop_in_for_sample_flow():
     from giant.sample import sample_flow, sample_secondaries
 
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=3,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=8,
-        expert_n_blocks=1,
-        router={
+        # PdgRouter (axis1) always builds its own training-vocab embedding,
+        # incompatible with conditioning.particle.type="physical" (the
+        # _nested_cfg default) — see _check_router_conditioning_compat.
+        particle_type="embedding",
+        material_type="embedding",
+        stage1_router={
             "enabled": True,
             "type": "composed",
             "axis0_type": "energy",
@@ -824,24 +925,31 @@ def test_build_models_routed_pair_composed_router_is_drop_in_for_sample_flow():
             "axis1_n_experts": 2,
         },
     )
-    stage1, sec_decoder = build_models(model_config)
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
     B = 5
     cond_cont, cond_cat = _cond(B, pdg=3, mat=2)
     stage1_norm, n_sec_pred = sample_flow(stage1, cond_cont, cond_cat, steps=2)
     assert stage1_norm.shape == (B, X_DIM)
+    # A fresh v0.3.0 Stage1Model has no n_sec_head (it moves to stage 2) —
+    # sample_flow returns n_sec_pred=None here, and n_sec must be
+    # asked of stage2 instead, using the just-sampled stage1_norm as context.
+    assert n_sec_pred is None
+    n_sec_pred = stage2.predict_n_sec(cond_cont, cond_cat, stage1_norm).argmax(dim=-1)
     assert n_sec_pred.shape == (B,)
 
     sec_cont, sec_type_emb, sec_valid = sample_secondaries(
-        sec_decoder, cond_cont, cond_cat, stage1_norm, n_sec_pred, steps=2
+        stage2, cond_cont, cond_cat, stage1_norm, n_sec_pred, steps=2
     )
     assert sec_cont.shape == (B, K_MAX, 4)
     assert sec_valid.shape == (B, K_MAX)
 
 
-# ── RoutedDenoisingMLP ───────────────────────────────────────────────────────
+# ── Stage1Model with a routed trunk ─────────────────────────────────────────
 
 
-def test_routed_denoising_mlp_output_shape_train_and_eval():
+def test_routed_stage1_output_shape_train_and_eval():
     B = 8
     model = _routed_stage1()
     x_t = torch.randn(B, X_DIM)
@@ -849,16 +957,16 @@ def test_routed_denoising_mlp_output_shape_train_and_eval():
     cond_cont, cond_cat = _cond(B)
 
     model.train()
-    out_train = model(x_t, t, cond_cont, cond_cat)
+    out_train = model(x_t, cond_cont, cond_cat, t=t)
     assert out_train.shape == (B, X_DIM)
 
     model.eval()
     with torch.no_grad():
-        out_eval = model(x_t, t, cond_cont, cond_cat)
+        out_eval = model(x_t, cond_cont, cond_cat, t=t)
     assert out_eval.shape == (B, X_DIM)
 
 
-def test_routed_denoising_mlp_gradients_flow_in_train_mode():
+def test_routed_stage1_gradients_flow_in_train_mode():
     """Soft mixture in train mode should touch every expert's parameters."""
     B = 8
     model = _routed_stage1(n_experts=3)
@@ -866,14 +974,14 @@ def test_routed_denoising_mlp_gradients_flow_in_train_mode():
     t = torch.rand(B)
     cond_cont, cond_cat = _cond(B)
     model.train()
-    flow_loss = model(x_t, t, cond_cont, cond_cat).sum()
+    flow_loss = model(x_t, cond_cont, cond_cat, t=t).sum()
     nsec_loss = model.predict_n_sec(cond_cont, cond_cat).sum()
     (flow_loss + nsec_loss).backward()
     for name, p in model.named_parameters():
         assert p.grad is not None, f"no grad for {name}"
 
 
-def test_routed_denoising_mlp_eval_dispatch_matches_manual_grouping():
+def test_routed_stage1_eval_dispatch_matches_manual_grouping():
     """Eval-mode grouped top-1 dispatch must equal running each row through
     its assigned expert individually (batch order shouldn't matter)."""
     B = 12
@@ -884,20 +992,20 @@ def test_routed_denoising_mlp_eval_dispatch_matches_manual_grouping():
     cond_cont, cond_cat = _cond(B)
 
     with torch.no_grad():
-        batched = model(x_t, t, cond_cont, cond_cat)
+        batched = model(x_t, cond_cont, cond_cat, t=t)
 
         t_emb = model.time_emb(t)
         c_emb = model.cond_enc(cond_cont, cond_cat)
         cond = torch.cat([t_emb, c_emb], dim=-1)
-        idx = model.router.top1(cond_cont, cond_cat)
+        idx = model.trunk.router.top1(cond_cont, cond_cat)
         manual = torch.zeros_like(x_t)
         for i in range(B):
-            manual[i] = model.experts[int(idx[i])](x_t[i : i + 1], cond[i : i + 1])[0]
+            manual[i] = model.trunk.experts[int(idx[i])](x_t[i : i + 1], cond[i : i + 1])[0]
 
     torch.testing.assert_close(batched, manual, atol=1e-5, rtol=1e-4)
 
 
-def test_routed_denoising_mlp_predict_n_sec_shape():
+def test_routed_stage1_predict_n_sec_shape():
     B = 6
     model = _routed_stage1()
     cond_cont, cond_cat = _cond(B)
@@ -905,15 +1013,15 @@ def test_routed_denoising_mlp_predict_n_sec_shape():
     assert logits.shape == (B, K_MAX + 1)
 
 
-def test_routed_denoising_mlp_has_no_pdg_embedding_weight_method():
+def test_routed_stage1_has_no_pdg_embedding_weight_method():
     model = _routed_stage1(pdg=5, mat=2)
     assert not hasattr(model, "pdg_embedding_weight")
 
 
-# ── RoutedSecondaryDecoder ───────────────────────────────────────────────────
+# ── Stage2OneShot with a routed trunk ────────────────────────────────────────
 
 
-def test_routed_secondary_decoder_output_shape_train_and_eval():
+def test_routed_stage2_output_shape_train_and_eval():
     B = 8
     decoder = _routed_sec_decoder()
     x_t = torch.randn(B, SEC_DIM)
@@ -922,16 +1030,16 @@ def test_routed_secondary_decoder_output_shape_train_and_eval():
     stage1_out = torch.randn(B, X_DIM)
 
     decoder.train()
-    out_train = decoder(x_t, t, cond_cont, cond_cat, stage1_out)
+    out_train = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t)
     assert out_train.shape == (B, SEC_DIM)
 
     decoder.eval()
     with torch.no_grad():
-        out_eval = decoder(x_t, t, cond_cont, cond_cat, stage1_out)
+        out_eval = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t)
     assert out_eval.shape == (B, SEC_DIM)
 
 
-def test_routed_secondary_decoder_gradients_flow():
+def test_routed_stage2_gradients_flow():
     B = 4
     decoder = _routed_sec_decoder(n_experts=3)
     x_t = torch.randn(B, SEC_DIM)
@@ -939,7 +1047,9 @@ def test_routed_secondary_decoder_gradients_flow():
     cond_cont, cond_cat = _cond(B)
     stage1_out = torch.randn(B, X_DIM)
     decoder.train()
-    decoder(x_t, t, cond_cont, cond_cat, stage1_out).sum().backward()
+    flow_loss = decoder(x_t, cond_cont, cond_cat, stage1_out, t=t).sum()
+    nsec_loss = decoder.predict_n_sec(cond_cont, cond_cat, stage1_out).sum()
+    (flow_loss + nsec_loss).backward()
     for name, p in decoder.named_parameters():
         assert p.grad is not None, f"no grad for {name}"
 
@@ -948,46 +1058,37 @@ def test_routed_secondary_decoder_gradients_flow():
 
 
 def test_build_models_monolith_when_router_absent():
-    model_config = dict(
-        pdg_vocab=4,
-        mat_vocab=2,
-        hidden_dim=32,
-        n_blocks=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-    )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, DenoisingMLP)
-    assert isinstance(sec_decoder, SecondaryDecoder)
+    cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2)
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert isinstance(stage1, Stage1Model)
+    assert isinstance(stage2, Stage2OneShot)
+    assert not isinstance(stage1.trunk, RoutedTrunk)
+    assert not isinstance(stage2.trunk, RoutedTrunk)
+    assert isinstance(stage1.trunk, ExpertTrunk)
+    assert isinstance(stage2.trunk, ExpertTrunk)
 
 
 def test_build_models_monolith_when_router_disabled():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        hidden_dim=32,
-        n_blocks=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        router={"enabled": False, "type": "energy", "n_experts": 4},
+        stage1_router={"enabled": False, "type": "energy", "n_experts": 4},
     )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, DenoisingMLP)
-    assert isinstance(sec_decoder, SecondaryDecoder)
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
+    assert not isinstance(stage1.trunk, RoutedTrunk)
+    assert not isinstance(stage2.trunk, RoutedTrunk)
+    assert isinstance(stage1.trunk, ExpertTrunk)
+    assert isinstance(stage2.trunk, ExpertTrunk)
 
 
 def test_build_models_routed_when_enabled():
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=4,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=16,
-        expert_n_blocks=2,
-        router={
+        stage1_router={
             "enabled": True,
             "type": "energy",
             "n_experts": 4,
@@ -995,37 +1096,102 @@ def test_build_models_routed_when_enabled():
             "learn_centers": True,
             "lambda_balance": 0.0,
         },
+        stage2_router={
+            "enabled": True,
+            "tie_to_stage1": False,
+            "type": "energy",
+            "n_experts": 4,
+            "temperature": 0.5,
+            "learn_centers": True,
+        },
     )
-    stage1, sec_decoder = build_models(model_config)
-    assert isinstance(stage1, RoutedDenoisingMLP)
-    assert isinstance(sec_decoder, RoutedSecondaryDecoder)
-    assert len(stage1.experts) == 4
-    assert len(sec_decoder.experts) == 4
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
+    assert isinstance(stage1.trunk, RoutedTrunk)
+    assert isinstance(stage2.trunk, RoutedTrunk)
+    assert len(stage1.trunk.experts) == 4
+    assert len(stage2.trunk.experts) == 4
+
+
+def test_build_models_explicit_resmlp_trunk_type_matches_default():
+    """stage1_model.trunk.type = 'resmlp' is the default's spelled-out
+    equivalent, not a behaviour change — gitea #33."""
+    default_cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2)
+    explicit_cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2, trunk={"type": "resmlp"})
+    default_stage1 = build_models(default_cfg)["stage1"]
+    explicit_stage1 = build_models(explicit_cfg)["stage1"]
+    assert default_stage1 is not None and explicit_stage1 is not None
+    assert type(default_stage1.trunk) is type(explicit_stage1.trunk) is ExpertTrunk
+    assert default_stage1.trunk.input_proj.weight.shape == explicit_stage1.trunk.input_proj.weight.shape
+    default_params = sum(p.numel() for p in default_stage1.parameters())
+    explicit_params = sum(p.numel() for p in explicit_stage1.parameters())
+    assert default_params == explicit_params
+
+
+def test_build_models_explicit_add_block_conditioning_matches_default():
+    """stage1_model.trunk.block_conditioning = 'add' is the default's
+    spelled-out equivalent, not a behaviour change — gitea #34."""
+    default_cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2)
+    explicit_cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2, trunk={"type": "resmlp", "block_conditioning": "add"})
+    default_stage1 = build_models(default_cfg)["stage1"]
+    explicit_stage1 = build_models(explicit_cfg)["stage1"]
+    assert default_stage1 is not None and explicit_stage1 is not None
+    assert type(default_stage1.trunk.blocks[0]) is type(explicit_stage1.trunk.blocks[0]) is ResBlock
+    default_params = sum(p.numel() for p in default_stage1.parameters())
+    explicit_params = sum(p.numel() for p in explicit_stage1.parameters())
+    assert default_params == explicit_params
+
+
+@pytest.mark.parametrize("block_type,cls", [("film", FilmResBlock), ("adaln", AdaLNResBlock)])
+def test_build_models_selects_block_conditioning(block_type, cls):
+    cfg = _nested_cfg(pdg_vocab=4, mat_vocab=2, trunk={"type": "resmlp", "block_conditioning": block_type})
+    stage1 = build_models(cfg)["stage1"]
+    assert stage1 is not None
+    assert isinstance(stage1.trunk, ExpertTrunk)
+    assert all(isinstance(b, cls) for b in stage1.trunk.blocks)
+
+
+def test_build_models_routed_trunk_uses_block_conditioning_for_every_expert():
+    cfg = _nested_cfg(
+        pdg_vocab=4,
+        mat_vocab=2,
+        trunk={"type": "resmlp", "block_conditioning": "film"},
+        stage1_router={"enabled": True, "type": "energy", "n_experts": 3},
+    )
+    stage1 = build_models(cfg)["stage1"]
+    assert stage1 is not None
+    assert isinstance(stage1.trunk, RoutedTrunk)
+    assert len(stage1.trunk.experts) == 3
+    for expert in stage1.trunk.experts:
+        assert all(isinstance(b, FilmResBlock) for b in expert.blocks)
 
 
 def test_build_models_routed_pair_is_drop_in_for_sample_flow():
     """Exercise the exact calling convention giant/sample.py uses."""
     from giant.sample import sample_flow, sample_secondaries
 
-    model_config = dict(
+    cfg = _nested_cfg(
         pdg_vocab=3,
         mat_vocab=2,
-        emb_dim=16,
-        dropout=0.1,
-        k_max=K_MAX,
-        expert_hidden_dim=8,
-        expert_n_blocks=1,
-        router={"enabled": True, "type": "energy", "n_experts": 2},
+        stage1_router={"enabled": True, "type": "energy", "n_experts": 2},
     )
-    stage1, sec_decoder = build_models(model_config)
+    models = build_models(cfg)
+    stage1, stage2 = models["stage1"], models["stage2"]
+    assert stage1 is not None and stage2 is not None
     B = 5
     cond_cont, cond_cat = _cond(B, pdg=3, mat=2)
     stage1_norm, n_sec_pred = sample_flow(stage1, cond_cont, cond_cat, steps=2)
     assert stage1_norm.shape == (B, X_DIM)
+    # A fresh v0.3.0 Stage1Model has no n_sec_head (it moves to stage 2) —
+    # sample_flow returns n_sec_pred=None here, and n_sec must be
+    # asked of stage2 instead, using the just-sampled stage1_norm as context.
+    assert n_sec_pred is None
+    n_sec_pred = stage2.predict_n_sec(cond_cont, cond_cat, stage1_norm).argmax(dim=-1)
     assert n_sec_pred.shape == (B,)
 
     sec_cont, sec_type_emb, sec_valid = sample_secondaries(
-        sec_decoder, cond_cont, cond_cat, stage1_norm, n_sec_pred, steps=2
+        stage2, cond_cont, cond_cat, stage1_norm, n_sec_pred, steps=2
     )
     assert sec_cont.shape == (B, K_MAX, 4)
     assert sec_valid.shape == (B, K_MAX)

tests/test_router_gating.py

@@ -36,7 +36,8 @@ def _model_cfg() -> dict:
 
 def _write_checkpoint(tmp_path) -> str:
     cfg = _model_cfg()
-    stage1, _ = build_models(cfg)
+    stage1 = build_models(cfg)["stage1"]
+    assert stage1 is not None
     norm = Normalizer()
     norm.mean = np.zeros(15, dtype=np.float32)
     norm.std = np.ones(15, dtype=np.float32)

tests/test_sample.py

@@ -0,0 +1,225 @@
+"""Tests for giant/sample.py's v0.3.0 stage-model sampling — the AR loop
+(`sample_secondaries_ar`) and non-"physical" `particle_type.target` coverage
+for the one-shot samplers."""
+
+import pytest
+import torch
+
+from giant.config import ConditioningAxisConfig, ParticleTypeConfig
+from giant.constants import COND_DIM, CONT_SLOT_DIM, K_MAX, PARTICLE_PHYS_DIM, X_DIM
+from giant.model.network import (
+    Stage1Model,
+    Stage2Autoregressive,
+    Stage2OneShot,
+    stage2_trunk_sec_dim,
+)
+from giant.sample import (
+    sample_flow,
+    sample_secondaries,
+    sample_secondaries_ar,
+    sample_secondaries_wgan,
+    sample_wgan,
+)
+
+_PHYS_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+
+
+def _particle_material_cfg(conditioning: str, emb_dim: int) -> tuple[ConditioningAxisConfig, ConditioningAxisConfig]:
+    cfg = ConditioningAxisConfig(type=conditioning, emb_dim=emb_dim, n_layers=1)
+    return cfg, cfg
+
+
+def _cond(B: int, pdg: int = 3, mat: int = 2) -> tuple[torch.Tensor, torch.Tensor]:
+    cond_cont = torch.randn(B, COND_DIM)
+    cond_cat = torch.stack([torch.randint(0, pdg, (B,)), torch.randint(0, mat, (B,))], dim=1)
+    return cond_cont, cond_cat
+
+
+def _conditioning_for(target: str) -> str:
+    # target="embedding" regresses against the conditioning's own embedding
+    # table — only meaningful when the
+    # conditioning axis is itself "embedding".
+    return "embedding" if target == "embedding" else "physical"
+
+
+def _stage2_oneshot(target: str, generator: str, emb_dim: int = 6, pdg: int = 3, mat: int = 2) -> Stage2OneShot:
+    particle_cfg, material_cfg = _particle_material_cfg(_conditioning_for(target), emb_dim)
+    particle_type_cfg = ParticleTypeConfig(target=target)
+    # build_models (giant/model/network.py) computes sec_dim this same way
+    # before constructing Stage2OneShot — its own default (SEC_DIM, the
+    # "physical" width) is only correct for target="physical".
+    sec_dim = stage2_trunk_sec_dim(particle_type_cfg, generator, K_MAX, emb_dim)
+    return Stage2OneShot(
+        pdg_vocab=pdg,
+        mat_vocab=mat,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        generator=generator,
+        time_dim=16,
+        noise_dim=8,
+        sec_dim=sec_dim,
+        particle_type_cfg=particle_type_cfg,
+    ).eval()
+
+
+def _stage2_ar(
+    target: str,
+    generator: str,
+    emb_dim: int = 6,
+    pdg: int = 3,
+    mat: int = 2,
+    k_max: int = 5,
+    history: str = "markov",
+) -> Stage2Autoregressive:
+    particle_cfg, material_cfg = _particle_material_cfg(_conditioning_for(target), emb_dim)
+    return Stage2Autoregressive(
+        pdg_vocab=pdg,
+        mat_vocab=mat,
+        particle_cfg=particle_cfg,
+        material_cfg=material_cfg,
+        hidden_dim=32,
+        n_res_blocks=2,
+        generator=generator,
+        time_dim=16,
+        noise_dim=8,
+        k_max=k_max,
+        particle_type_cfg=ParticleTypeConfig(target=target),
+        history=history,
+        attn_n_heads=2,
+        attn_n_layers=1,
+    ).eval()
+
+
+def _expected_type_dim(target: str, emb_dim: int) -> int:
+    return PARTICLE_PHYS_DIM if target == "physical" else emb_dim
+
+
+# ── Stage-1 n_sec ownership ──────────────────────────────────────────────────
+
+
+def test_sample_flow_returns_none_n_sec_when_stage1_owns_no_head():
+    model = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+    )
+    cond_cont, cond_cat = _cond(4)
+    sample, n_sec = sample_flow(model, cond_cont, cond_cat, steps=2)
+    assert sample.shape == (4, X_DIM)
+    assert n_sec is None
+
+
+def test_sample_wgan_returns_none_n_sec_when_stage1_owns_no_head():
+    model = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="wgan",
+        noise_dim=8,
+    )
+    cond_cont, cond_cat = _cond(4)
+    sample, n_sec = sample_wgan(model, cond_cont, cond_cat)
+    assert sample.shape == (4, X_DIM)
+    assert n_sec is None
+
+
+def test_sample_flow_returns_n_sec_for_legacy_stage1():
+    model = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PHYS_CFG,
+        material_cfg=_PHYS_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        n_sec_head_k_max=K_MAX,
+    )
+    cond_cont, cond_cat = _cond(5)
+    _, n_sec = sample_flow(model, cond_cont, cond_cat, steps=2)
+    assert n_sec is not None and n_sec.shape == (5,)
+
+
+# ── Stage2OneShot: non-"physical" particle_type.target ──────────────────────
+
+
+@pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
+def test_sample_secondaries_flow_shapes_by_target(target):
+    B, emb_dim = 5, 6
+    decoder = _stage2_oneshot(target, "flow", emb_dim=emb_dim)
+    cond_cont, cond_cat = _cond(B)
+    stage1_out = torch.randn(B, X_DIM)
+    n_sec_pred = torch.randint(0, K_MAX + 1, (B,))
+    sec_cont, sec_type, sec_valid = sample_secondaries(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2)
+    assert sec_cont.shape == (B, K_MAX, CONT_SLOT_DIM)
+    assert sec_type.shape == (B, K_MAX, _expected_type_dim(target, emb_dim))
+    assert sec_valid.shape == (B, K_MAX)
+    assert torch.isfinite(sec_cont).all()
+    assert torch.isfinite(sec_type).all()
+
+
+@pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
+def test_sample_secondaries_wgan_shapes_by_target(target):
+    B, emb_dim = 5, 6
+    decoder = _stage2_oneshot(target, "wgan", emb_dim=emb_dim)
+    cond_cont, cond_cat = _cond(B)
+    stage1_out = torch.randn(B, X_DIM)
+    n_sec_pred = torch.randint(0, K_MAX + 1, (B,))
+    sec_cont, sec_type, sec_valid = sample_secondaries_wgan(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred)
+    assert sec_cont.shape == (B, K_MAX, CONT_SLOT_DIM)
+    assert sec_type.shape == (B, K_MAX, _expected_type_dim(target, emb_dim))
+    assert sec_valid.shape == (B, K_MAX)
+
+
+# ── Stage2Autoregressive ─────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("history", ["markov", "attention"])
+@pytest.mark.parametrize("generator", ["flow", "wgan"])
+@pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
+def test_sample_secondaries_ar_shapes(target, generator, history):
+    B, k_max, emb_dim = 4, 5, 6
+    decoder = _stage2_ar(target, generator, emb_dim=emb_dim, k_max=k_max, history=history)
+    cond_cont, cond_cat = _cond(B)
+    stage1_out = torch.randn(B, X_DIM)
+    n_sec_pred = torch.randint(0, k_max + 1, (B,))
+    sec_cont, sec_type, sec_valid = sample_secondaries_ar(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2)
+    assert sec_cont.shape == (B, k_max, CONT_SLOT_DIM)
+    assert sec_type.shape == (B, k_max, _expected_type_dim(target, emb_dim))
+    assert sec_valid.shape == (B, k_max)
+    assert torch.isfinite(sec_cont).all()
+    assert torch.isfinite(sec_type).all()
+
+
+@pytest.mark.parametrize("generator", ["flow", "wgan"])
+@pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
+def test_sample_secondaries_ar_valid_mask_matches_n_sec(target, generator):
+    B, k_max, emb_dim = 3, 5, 6
+    decoder = _stage2_ar(target, generator, emb_dim=emb_dim, k_max=k_max)
+    cond_cont, cond_cat = _cond(B)
+    stage1_out = torch.randn(B, X_DIM)
+    n_sec_pred = torch.tensor([0, 2, k_max])
+    _, _, sec_valid = sample_secondaries_ar(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2)
+    for i, n in enumerate(n_sec_pred.tolist()):
+        assert sec_valid[i, :n].all()
+        assert not sec_valid[i, n:].any()
+
+
+def test_sample_secondaries_ar_first_slot_has_no_history():
+    """Slot 0 always has has_prev=False internally — nothing to assert on
+    the public API directly, but a k_max=1 run should not crash on the
+    "previous token" path at all (has_prev never true)."""
+    B, emb_dim = 3, 6
+    decoder = _stage2_ar("physical", "flow", emb_dim=emb_dim, k_max=1)
+    cond_cont, cond_cat = _cond(B)
+    stage1_out = torch.randn(B, X_DIM)
+    n_sec_pred = torch.tensor([0, 1, 1])
+    sec_cont, sec_type, sec_valid = sample_secondaries_ar(decoder, cond_cont, cond_cat, stage1_out, n_sec_pred, steps=2)
+    assert sec_cont.shape == (B, 1, CONT_SLOT_DIM)
+    assert sec_valid.tolist() == [[False], [True], [True]]

tests/test_setup_cache.py

@@ -7,15 +7,14 @@ import pandas as pd
 import pytest
 
 from giant.data import setup_cache
+from giant.data.loader import TopNMap
 from giant.data.setup_cache import NormalizerEntry, SetupCache
 from giant.data.transforms import Normalizer
 
 
 def _touch_parquet(path, n=1):
     path.parent.mkdir(parents=True, exist_ok=True)
-    pd.DataFrame(
-        {"pdg": [11] * n, "material": ["G4_AIR"] * n, "process": ["eIoni"] * n}
-    ).to_parquet(path)
+    pd.DataFrame({"pdg": [11] * n, "material": ["G4_AIR"] * n, "process": ["eIoni"] * n}).to_parquet(path)
     return path
 
 
@@ -28,9 +27,7 @@ def _normalizer(width=3):
 
 def _entry(n_train_steps=100, sample=None):
     sample = np.array([1.0, 2.0, 3.0], dtype=np.float32) if sample is None else sample
-    return NormalizerEntry(
-        _normalizer(), _normalizer(), _normalizer(2), n_train_steps, sample
-    )
+    return NormalizerEntry(_normalizer(), _normalizer(), _normalizer(2), n_train_steps, sample)
 
 
 # ── sidecar_path ─────────────────────────────────────────────────────────
@@ -38,9 +35,7 @@ def _entry(n_train_steps=100, sample=None):
 
 def test_sidecar_path_single_file(tmp_path):
     f = tmp_path / "shard.parquet"
-    assert (
-        setup_cache.sidecar_path(f) == tmp_path / "shard.parquet.giant_train_cache.json"
-    )
+    assert setup_cache.sidecar_path(f) == tmp_path / "shard.parquet.giant_train_cache.json"
 
 
 def test_sidecar_path_directory(tmp_path):
@@ -50,10 +45,7 @@ def test_sidecar_path_directory(tmp_path):
 
 def test_sidecar_path_manifest(tmp_path):
     m = tmp_path / "pools" / "full.manifest"
-    assert (
-        setup_cache.sidecar_path(m)
-        == tmp_path / "pools" / "full.manifest.giant_train_cache.json"
-    )
+    assert setup_cache.sidecar_path(m) == tmp_path / "pools" / "full.manifest.giant_train_cache.json"
 
 
 # ── fingerprint_files ────────────────────────────────────────────────────
@@ -101,6 +93,37 @@ def test_save_load_round_trip(tmp_path):
     np.testing.assert_allclose(entry.energy_quantiles, [1.0, 2.0, 3.0])
 
 
+def test_save_load_round_trip_topn_maps(tmp_path):
+    data = _touch_parquet(tmp_path / "shard.parquet")
+    files = [data]
+
+    cache = SetupCache.empty(files)
+    cache.topn_maps[setup_cache.topn_key("pdg", 3)] = TopNMap(
+        class_map={22: 0, 11: 1, 2212: 2}, other_members={2212: 5}
+    )
+    cache.topn_maps[setup_cache.topn_key("material", 2)] = TopNMap(
+        class_map={"G4_AIR": 0, "PbWO4": 1}, other_members={}
+    )
+
+    setup_cache.save(data, files, cache)
+    loaded = setup_cache.load(data, files)
+
+    assert loaded is not None
+    pdg_m = loaded.topn_maps[setup_cache.topn_key("pdg", 3)]
+    assert pdg_m.class_map == {22: 0, 11: 1, 2212: 2}
+    assert pdg_m.other_members == {2212: 5}
+    # key type is int (matches pdg_map's own key type), not str
+    assert all(isinstance(k, int) for k in pdg_m.class_map)
+
+    mat_m = loaded.topn_maps[setup_cache.topn_key("material", 2)]
+    assert mat_m.class_map == {"G4_AIR": 0, "PbWO4": 1}
+
+
+def test_topn_key_unknown_axis_raises():
+    with pytest.raises(ValueError, match="unknown top-N map axis"):
+        setup_cache.topn_key("process", 4)
+
+
 def test_load_missing_sidecar_returns_none(tmp_path):
     data = _touch_parquet(tmp_path / "shard.parquet")
     assert setup_cache.load(data, [data]) is None
@@ -149,6 +172,25 @@ def test_load_invalidates_on_file_content_change(tmp_path):
     assert setup_cache.load(data, files) is None
 
 
+def test_load_returns_none_on_malformed_cache_body(tmp_path):
+    """format_version/dims/fingerprint all check out, but the cache body
+    itself doesn't match SetupCache.from_json's expected shape (e.g. hand-
+    edited or written by a version that changed a nested key) — a clean
+    miss, not a crash."""
+    data = _touch_parquet(tmp_path / "shard.parquet")
+    files = [data]
+    setup_cache.save(data, files, SetupCache.empty(files))
+
+    path = setup_cache.sidecar_path(data)
+    raw = json.loads(path.read_text())
+    raw["vocab"] = {"pdg_map": {"11": 0}}  # missing required "mat_map" key
+    path.write_text(json.dumps(raw))
+
+    echoed = []
+    assert setup_cache.load(data, files, echo=echoed.append) is None
+    assert any("malformed" in m for m in echoed)
+
+
 def test_load_soft_warns_on_git_hash_mismatch_but_still_hits(tmp_path, capsys):
     data = _touch_parquet(tmp_path / "shard.parquet")
     files = [data]
@@ -325,3 +367,11 @@ def test_compute_event_index_from_files_single_file_unaffected(tmp_path):
 
     np.testing.assert_array_equal(unique_ids, [5, 7])
     np.testing.assert_array_equal(counts, [2, 1])
+
+
+def test_compute_event_index_from_files_empty_file_list():
+    unique_ids, counts = setup_cache.compute_event_index_from_files([])
+    assert unique_ids.size == 0
+    assert counts.size == 0
+    assert unique_ids.dtype == np.int64
+    assert counts.dtype == np.int64

tests/test_steps_to_parquet.py

@@ -1,6 +1,6 @@
 import polars as pl
 
-from scripts import steps_to_parquet
+from giant.tools import steps_to_parquet
 
 
 def _frame() -> pl.DataFrame:
@@ -24,18 +24,14 @@ def _frame() -> pl.DataFrame:
 def test_e_sec_sums_child_first_step_energy():
     out, n_orphaned = steps_to_parquet._add_secondary_attributes(_frame())
     assert n_orphaned == 0
-    e_sec = dict(
-        zip(zip(out["track_id"], out["step_no"], out["event_id"]), out["e_sec"])
-    )
+    e_sec = dict(zip(zip(out["track_id"], out["step_no"], out["event_id"]), out["e_sec"]))
     assert e_sec[(1, 0, 0)] == 15.0  # one child, first-step pre_E 15
     assert e_sec[(1, 0, 1)] == 50.0  # two children, 20 + 30
 
 
 def test_e_sec_zero_when_no_children():
     out, _ = steps_to_parquet._add_secondary_attributes(_frame())
-    childless = out.filter(
-        (pl.col("event_id") == 0) & (pl.col("track_id") == 1) & (pl.col("step_no") == 1)
-    )
+    childless = out.filter((pl.col("event_id") == 0) & (pl.col("track_id") == 1) & (pl.col("step_no") == 1))
     assert childless["e_sec"].item() == 0.0
 
 
@@ -69,9 +65,7 @@ def test_orphaned_child_track_is_dropped_not_nulled():
         }
     )
     out, n_orphaned = steps_to_parquet._add_secondary_attributes(df)
-    row = out.filter(
-        (pl.col("event_id") == 0) & (pl.col("track_id") == 1) & (pl.col("step_no") == 0)
-    )
+    row = out.filter((pl.col("event_id") == 0) & (pl.col("track_id") == 1) & (pl.col("step_no") == 0))
 
     assert n_orphaned == 1
     assert row["child_track_ids"].to_list() == [[2]]

tests/test_steps_to_parquet_parallel.py

@@ -2,7 +2,7 @@ import json
 import sys
 from pathlib import Path
 
-from scripts import steps_to_parquet_parallel
+from giant.tools import steps_to_parquet_parallel
 
 run_parallel = steps_to_parquet_parallel.run_parallel
 resolve_destination = steps_to_parquet_parallel.resolve_destination
@@ -124,31 +124,13 @@ def _make_dataset(tmp_path: Path, schemas: list[str] | None = None) -> Path:
 def test_resolve_destination_uses_latest_schema(tmp_path):
     root_file = _make_dataset(tmp_path, schemas=["schema1", "schema3", "schema2"])
     dest = resolve_destination(root_file, tmp_path, schema_override=None)
-    assert (
-        dest
-        == tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema3"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    assert dest == tmp_path / "processed" / "steps" / "gen1" / "schema3" / "pbwo4" / "shard-000.parquet"
 
 
 def test_resolve_destination_schema_override_wins(tmp_path):
     root_file = _make_dataset(tmp_path, schemas=["schema1", "schema3"])
     dest = resolve_destination(root_file, tmp_path, schema_override="schema9")
-    assert (
-        dest
-        == tmp_path
-        / "processed"
-        / "steps"
-        / "gen1"
-        / "schema9"
-        / "pbwo4"
-        / "shard-000.parquet"
-    )
+    assert dest == tmp_path / "processed" / "steps" / "gen1" / "schema9" / "pbwo4" / "shard-000.parquet"
 
 
 def test_resolve_destination_errors_without_any_schema(tmp_path):

tests/test_train.py

@@ -1,6 +1,50 @@
-"""Tests for giant/train.py helpers."""
+"""Tests for giant/training/."""
 
-from giant.train import _gumbel_tau, _wandb_run_config
+import copy
+import csv
+import math
+import tempfile
+from pathlib import Path
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+
+from giant.config import ParticleTypeConfig
+from giant.constants import (
+    COND_DIM,
+    CONT_SLOT_DIM,
+    K_MAX,
+    PARTICLE_PHYS_DIM,
+    SEC_SLOT_DIM,
+    X_DIM,
+)
+from giant.data.dataset import StepBatch
+from giant.model.network import build_critics, build_models
+from giant.training import (
+    FlowDDPMStageTrainer,
+    StageSpec,
+    WGANStageTrainer,
+    build_stage_trainers,
+    train,
+)
+from giant.training.metrics import _wandb_run_config
+from giant.training.stage2_inputs import (
+    _ar_has_prev,
+    _assemble_stage2_ar_inputs,
+    _assemble_stage2_ar_target,
+    _assemble_stage2_real,
+    _gumbel_tau,
+    _relax_onehot_type_slice,
+    _remaining_energy_fraction,
+    _shift_prev,
+    _stage2_tf_prob,
+    _stick_fraction,
+    _type_repr,
+)
+
+PDG_VOCAB = 6
+MAT_VOCAB = 3
 
 
 def test_gumbel_tau_at_step_zero_is_start():
@@ -26,72 +70,681 @@ def test_gumbel_tau_handles_zero_total_steps():
     assert abs(_gumbel_tau(1, 0, 1.0, 0.1) - 0.1) < 1e-9
 
 
-def _base_wandb_kwargs(**overrides):
-    kwargs = dict(
-        mode="flow",
-        epochs=30,
-        lr=3e-4,
-        warmup_epochs=3,
-        weight_decay=0.01,
-        ema_decay=0.9999,
-        lambda_nsec=0.1,
-        lambda_s2=1.0,
-        lambda_balance=0.035,
-        lambda_proc=0.0,
-        lambda_entropy=0.0,
-        gumbel_tau_start=1.0,
-        gumbel_tau_end=0.1,
-        n_critic=5,
-        gp_weight=10.0,
-        model_config={"router": {"enabled": False}},
-        stage1_params=100,
-        sec_decoder_params=50,
-        critic_params=0,
-        sec_critic_params=0,
-        total_params=150,
-    )
-    kwargs.update(overrides)
-    return kwargs
-
-
-def test_wandb_run_config_omits_router_knobs_when_router_disabled():
-    cfg = _wandb_run_config(**_base_wandb_kwargs())
-    for key in (
-        "lambda_balance",
-        "lambda_proc",
-        "lambda_entropy",
-        "gumbel_tau_start",
-        "gumbel_tau_end",
-    ):
-        assert key not in cfg
-    # still present, nested, regardless of router state
-    assert cfg["model"] == {"router": {"enabled": False}}
-
-
-def test_wandb_run_config_includes_router_knobs_when_router_enabled():
-    cfg = _wandb_run_config(
-        **_base_wandb_kwargs(model_config={"router": {"enabled": True}})
-    )
-    assert cfg["lambda_balance"] == 0.035
-    assert cfg["lambda_proc"] == 0.0
-    assert cfg["lambda_entropy"] == 0.0
-    assert cfg["gumbel_tau_start"] == 1.0
-    assert cfg["gumbel_tau_end"] == 0.1
-
-
-def test_wandb_run_config_omits_wgan_knobs_when_mode_is_not_wgan():
-    cfg = _wandb_run_config(**_base_wandb_kwargs(mode="flow"))
-    assert "n_critic" not in cfg
-    assert "gp_weight" not in cfg
-
-
-def test_wandb_run_config_includes_wgan_knobs_when_mode_is_wgan():
-    cfg = _wandb_run_config(**_base_wandb_kwargs(mode="wgan"))
-    assert cfg["n_critic"] == 5
-    assert cfg["gp_weight"] == 10.0
+def test_wandb_run_config_includes_full_cfg_and_param_counts():
+    cfg = {
+        "train": {"lr": 3e-4},
+        "conditioning": {"out_dim": 128},
+        "stage1_model": {"generator": "flow"},
+        "stage2_model": {"generator": "wgan"},
+    }
+    wcfg = _wandb_run_config(cfg, model_config={"pdg_vocab": 3}, param_counts={"stage1": 100})
+    assert wcfg["train"] == {"lr": 3e-4}
+    assert wcfg["stage1_model"] == {"generator": "flow"}
+    assert wcfg["stage2_model"] == {"generator": "wgan"}
+    assert wcfg["model_config"] == {"pdg_vocab": 3}
+    assert wcfg["param_counts"] == {"stage1": 100}
 
 
 def test_wandb_run_config_handles_missing_model_config():
-    cfg = _wandb_run_config(**_base_wandb_kwargs(model_config=None))
-    assert cfg["model"] == {}
-    assert "lambda_balance" not in cfg
+    cfg = {"train": {}, "conditioning": {}, "stage1_model": {}, "stage2_model": {}}
+    wcfg = _wandb_run_config(cfg, model_config=None, param_counts={})
+    assert wcfg["model_config"] == {}
+
+
+# --- AR helper functions (v0.3.0 step 5) ---------
+
+
+def test_stick_fraction_matches_sigmoid_of_logit():
+    sec_cont = torch.zeros(2, 3, SEC_SLOT_DIM)
+    sec_cont[..., 0] = torch.tensor([[0.0, 2.0, -2.0], [1.0, -1.0, 0.0]])
+    frac = _stick_fraction(sec_cont)
+    assert torch.allclose(frac, torch.sigmoid(sec_cont[..., 0]))
+
+
+def test_remaining_energy_fraction_hand_computed():
+    fraction = torch.tensor([[0.5, 0.5, 1.0]])
+    remaining = _remaining_energy_fraction(fraction)
+    assert torch.allclose(remaining, torch.tensor([[1.0, 0.5, 0.25]]))
+
+
+def test_shift_prev_shifts_and_zero_pads_slot0():
+    x = torch.arange(2 * 4 * 3).reshape(2, 4, 3).float()
+    shifted = _shift_prev(x)
+    assert torch.all(shifted[:, 0] == 0)
+    assert torch.equal(shifted[:, 1:], x[:, :-1])
+
+
+def test_ar_has_prev_false_only_at_slot_zero():
+    has_prev = _ar_has_prev(5, torch.device("cpu"))
+    assert has_prev.shape == (1, 5)
+    assert has_prev.tolist() == [[False, True, True, True, True]]
+
+
+# --- _stage2_tf_prob (v0.3.0 step 7) -----------
+
+
+def test_stage2_tf_prob_always_is_constant_one():
+    assert _stage2_tf_prob("always", 1.0, 0.0, 0, 10) == 1.0
+    assert _stage2_tf_prob("always", 1.0, 0.0, 9, 10) == 1.0
+
+
+def test_stage2_tf_prob_never_is_constant_zero():
+    assert _stage2_tf_prob("never", 1.0, 1.0, 0, 10) == 0.0
+    assert _stage2_tf_prob("never", 1.0, 1.0, 9, 10) == 0.0
+
+
+def test_stage2_tf_prob_scheduled_interpolates_linearly():
+    assert _stage2_tf_prob("scheduled", 1.0, 0.0, 0, 11) == 1.0
+    assert abs(_stage2_tf_prob("scheduled", 1.0, 0.0, 5, 11) - 0.5) < 1e-9
+    assert _stage2_tf_prob("scheduled", 1.0, 0.0, 10, 11) == 0.0
+
+
+def test_stage2_tf_prob_scheduled_clamps_beyond_total_epochs():
+    end = _stage2_tf_prob("scheduled", 1.0, 0.0, 10, 11)
+    beyond = _stage2_tf_prob("scheduled", 1.0, 0.0, 50, 11)
+    assert beyond == end
+
+
+def test_stage2_tf_prob_scheduled_handles_single_epoch():
+    # total_epochs=1 is guarded to a denominator of 1 internally (like
+    # _gumbel_tau's total_steps=0 guard) — epoch=0 gives zero progress.
+    assert _stage2_tf_prob("scheduled", 1.0, 0.0, 0, 1) == 1.0
+
+
+@pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
+def test_type_repr_shapes_and_values(target):
+    B, K, emb_dim = 3, 4, 6
+    sec_cont = torch.randn(B, K, SEC_SLOT_DIM)
+    sec_type_idx = torch.randint(0, emb_dim, (B, K))
+    cond_enc = torch.nn.Module()
+    if target == "embedding":
+        cond_enc.pdg_emb = torch.nn.Embedding(emb_dim, emb_dim)
+    repr_ = _type_repr(sec_type_idx, sec_cont, ParticleTypeConfig(target=target), cond_enc, emb_dim)
+    expected_width = PARTICLE_PHYS_DIM if target == "physical" else emb_dim
+    assert repr_.shape == (B, K, expected_width)
+    if target == "physical":
+        assert torch.equal(repr_, sec_cont[..., CONT_SLOT_DIM : CONT_SLOT_DIM + PARTICLE_PHYS_DIM])
+    if target == "onehot":
+        assert torch.all(repr_.sum(-1) == 1.0)
+
+
+@pytest.mark.parametrize(
+    "target,generator",
+    [
+        ("physical", "flow"),
+        ("physical", "wgan"),
+        ("onehot", "flow"),
+        ("onehot", "wgan"),
+        ("embedding", "flow"),
+        ("embedding", "wgan"),
+    ],
+)
+def test_assemble_stage2_ar_target_matches_assemble_stage2_real_flattened(target, generator):
+    """Regression test tying the refactor together: _assemble_stage2_real is
+    now defined as _assemble_stage2_ar_target(...).flatten(1)."""
+    B, emb_dim = 4, 6
+    sec_cont = torch.randn(B, K_MAX, SEC_SLOT_DIM)
+    sec_type_idx = torch.randint(0, emb_dim, (B, K_MAX))
+    cond_enc = torch.nn.Module()
+    if target == "embedding":
+        cond_enc.pdg_emb = torch.nn.Embedding(emb_dim, emb_dim)
+    particle_type_cfg = ParticleTypeConfig(target=target)
+    flat = _assemble_stage2_real(sec_cont, sec_type_idx, particle_type_cfg, generator, cond_enc, emb_dim)
+    unflat = _assemble_stage2_ar_target(sec_cont, sec_type_idx, particle_type_cfg, generator, cond_enc, emb_dim)
+    assert torch.equal(unflat.flatten(1), flat)
+
+
+def test_assemble_stage2_ar_inputs_shapes_and_history_feat_width():
+    B, emb_dim = 3, 6
+    sec_cont = torch.randn(B, K_MAX, SEC_SLOT_DIM)
+    sec_type_idx = torch.randint(0, emb_dim, (B, K_MAX))
+    cond_enc = torch.nn.Module()
+    out = _assemble_stage2_ar_inputs(sec_cont, sec_type_idx, ParticleTypeConfig(target="physical"), cond_enc, emb_dim)
+    assert out["history_feat"].shape == (B, K_MAX, CONT_SLOT_DIM + PARTICLE_PHYS_DIM)
+    assert out["has_prev"].shape == (B, K_MAX)
+    assert out["remaining_frac"].shape == (B, K_MAX)
+    assert out["slot_idx"].shape == (B, K_MAX)
+    assert torch.all(out["slot_idx"][:, 0] == 0.0)
+    assert torch.all(out["slot_idx"][:, -1] == 1.0)
+
+
+def test_relax_onehot_type_slice_grad_probe_populates_both_norms():
+    B, k_max, cont_dim, type_dim = 4, K_MAX, CONT_SLOT_DIM, 6
+    x_flat = torch.randn(B, k_max * (cont_dim + type_dim), requires_grad=True)
+    grad_probe: dict[str, float] = {}
+    out = _relax_onehot_type_slice(x_flat, k_max, cont_dim, type_dim, tau=0.5, grad_probe=grad_probe)
+    out.sum().backward()
+    assert grad_probe["cont"] >= 0.0
+    assert grad_probe["type"] >= 0.0
+
+
+def test_relax_onehot_type_slice_grad_probe_none_is_backward_compatible():
+    B, k_max, cont_dim, type_dim = 4, K_MAX, CONT_SLOT_DIM, 6
+    x_flat = torch.randn(B, k_max * (cont_dim + type_dim), requires_grad=True)
+    out = _relax_onehot_type_slice(x_flat, k_max, cont_dim, type_dim, tau=0.5)
+    out.sum().backward()
+    assert x_flat.grad is not None
+
+
+# --- end-to-end train() integration tests -----------------------------------
+
+PARTICLE_CFG = {"type": "physical", "emb_dim": 8, "n_layers": 1}
+MATERIAL_CFG = {"type": "physical", "emb_dim": 8, "n_layers": 1}
+
+
+def _base_cfg():
+    return {
+        "conditioning": {
+            "out_dim": 32,
+            "share_stages": False,
+            "particle": dict(PARTICLE_CFG),
+            "material": dict(MATERIAL_CFG),
+        },
+        "stage1_model": {
+            "active": True,
+            "generator": "flow",
+            "hidden_dim": 24,
+            "n_res_blocks": 2,
+            "dropout": 0.0,
+            "lambda": 1.0,
+            "flow": {"time_dim": 16},
+            "ddpm": {"time_dim": 16, "n_steps": 50},
+            "wgan": {
+                "noise_dim": 16,
+                "n_critic": 2,
+                "gp_weight": 10.0,
+                "critic_lr": 0.0,
+            },
+            "router": {"enabled": False},
+        },
+        "stage2_model": {
+            "active": True,
+            "decoder": "one_shot",
+            "generator": "wgan",
+            "hidden_dim": 24,
+            "n_res_blocks": 2,
+            "dropout": 0.0,
+            "lambda": 1.0,
+            "k_max": K_MAX,
+            "context_dim": 16,
+            "n_sec": {"mode": "head", "lambda": 0.1},
+            # Explicit, not relying on the fallback default (which is
+            # "onehot", matching DEFAULT_CONFIG — see issues.md Issue 1):
+            # the "physical"-labelled cases below (and this fixture's own
+            # comment history) intend this as the base "physical" case,
+            # with "*_onehot"/"*_embedding" cases opting in explicitly.
+            "particle_type": {"target": "physical", "lambda": 1.0},
+            "flow": {"time_dim": 16},
+            "ddpm": {"time_dim": 16, "n_steps": 50},
+            "wgan": {
+                "noise_dim": 16,
+                "n_critic": 2,
+                "gp_weight": 10.0,
+                "critic_lr": 0.0,
+            },
+            "router": {"enabled": False, "tie_to_stage1": False},
+        },
+        "train": {
+            "epochs": 2,
+            "batch_size": 8,
+            "lr": 3e-4,
+            "weight_decay": 0.01,
+            "ema_decay": 0.999,
+            "warmup_epochs": 0,
+            "val_fraction": 0.1,
+            "max_val_batches": 0,
+            "num_workers": 0,
+            "seed": 0,
+            "validate_every": 0,
+            "validate_steps": 2,
+            "wandb": False,
+        },
+    }
+
+
+def _fake_batches(n_batches, batch_size, seed=0):
+    g = torch.Generator().manual_seed(seed)
+    batches = []
+    for _ in range(n_batches):
+        cond_cont = torch.randn(batch_size, COND_DIM, generator=g)
+        cond_cat = torch.stack(
+            [
+                torch.randint(0, PDG_VOCAB, (batch_size,), generator=g),
+                torch.randint(0, MAT_VOCAB, (batch_size,), generator=g),
+            ],
+            dim=1,
+        )
+        x1 = torch.randn(batch_size, X_DIM, generator=g)
+        n_sec = torch.randint(0, K_MAX, (batch_size,), generator=g)
+        sec_cont = torch.randn(batch_size, K_MAX, SEC_SLOT_DIM, generator=g)
+        proc_idx = torch.zeros(batch_size, dtype=torch.long)
+        sec_type_idx = torch.zeros(batch_size, K_MAX, dtype=torch.long)
+        batches.append(StepBatch(cond_cont, cond_cat, x1, n_sec, sec_cont, proc_idx, sec_type_idx))
+    return batches
+
+
+def _model_config(cfg):
+    return {
+        "pdg_vocab": PDG_VOCAB,
+        "mat_vocab": MAT_VOCAB,
+        "conditioning": cfg["conditioning"],
+        "stage1_model": cfg["stage1_model"],
+        "stage2_model": cfg["stage2_model"],
+    }
+
+
+def _run_train(cfg, out_dir, resume_path=None):
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    train_loader = _fake_batches(4, cfg["train"]["batch_size"])
+    val_loader = _fake_batches(2, cfg["train"]["batch_size"], seed=1)
+    train(
+        cfg=cfg,
+        models=models,
+        critics=critics,
+        train_loader=train_loader,
+        val_loader=val_loader,
+        device=torch.device("cpu"),
+        out_dir=out_dir,
+        normalizer_dict={"cond": {}, "target": {}, "sec_phys": {}},
+        pdg_map={"22": 0},
+        mat_map={"G4_AIR": 0},
+        proc_map=None,
+        model_config=model_config,
+        total_train_batches=4,
+        resume_path=resume_path,
+    )
+
+
+@pytest.mark.parametrize(
+    "label,mutate",
+    [
+        ("both_flow", lambda cfg: None),
+        ("both_wgan", lambda cfg: cfg["stage1_model"].__setitem__("generator", "wgan")),
+        (
+            "mixed_stage1_flow_stage2_wgan",
+            lambda cfg: None,  # already the default
+        ),
+        (
+            "mixed_stage1_wgan_stage2_flow",
+            lambda cfg: (
+                cfg["stage1_model"].__setitem__("generator", "wgan"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+            ),
+        ),
+        ("stage1_only", lambda cfg: cfg["stage2_model"].__setitem__("active", False)),
+        ("stage2_only", lambda cfg: cfg["stage1_model"].__setitem__("active", False)),
+        (
+            "both_ddpm_stage1_flow_stage2",
+            lambda cfg: (
+                cfg["stage1_model"].__setitem__("generator", "ddpm"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+            ),
+        ),
+        (
+            "routed_stage1_energy_gumbel",
+            lambda cfg: cfg["stage1_model"].__setitem__(
+                "router",
+                {
+                    "enabled": True,
+                    "type": "energy",
+                    "n_experts": 3,
+                    "temperature": 0.5,
+                    "learn_centers": True,
+                    "lambda_balance": 0.1,
+                    "lambda_entropy": 0.01,
+                    "gumbel": True,
+                    "gumbel_tau_start": 1.0,
+                    "gumbel_tau_end": 0.1,
+                },
+            ),
+        ),
+        (
+            "stage2_onehot_target_wgan",
+            lambda cfg: cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0}),
+        ),
+        (
+            "stage2_onehot_target_flow",
+            lambda cfg: (
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "stage2_embedding_target_wgan",
+            lambda cfg: (
+                cfg["conditioning"]["particle"].__setitem__("type", "embedding"),
+                cfg["conditioning"]["material"].__setitem__("type", "embedding"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "embedding", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "stage2_embedding_target_flow",
+            lambda cfg: (
+                cfg["conditioning"]["particle"].__setitem__("type", "embedding"),
+                cfg["conditioning"]["material"].__setitem__("type", "embedding"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "embedding", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "ar_wgan_onehot",
+            lambda cfg: (
+                cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "ar_wgan_physical",
+            lambda cfg: cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+        ),
+        (
+            "ar_flow_onehot",
+            lambda cfg: (
+                cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "ar_flow_embedding",
+            lambda cfg: (
+                cfg["conditioning"]["particle"].__setitem__("type", "embedding"),
+                cfg["conditioning"]["material"].__setitem__("type", "embedding"),
+                cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "embedding", "lambda": 1.0}),
+            ),
+        ),
+        (
+            "ar_stage2_only",
+            lambda cfg: (
+                cfg["stage1_model"].__setitem__("active", False),
+                cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+            ),
+        ),
+        (
+            "ar_mixed_stage1_wgan_stage2_flow_onehot",
+            lambda cfg: (
+                cfg["stage1_model"].__setitem__("generator", "wgan"),
+                cfg["stage2_model"].__setitem__("generator", "flow"),
+                cfg["stage2_model"].__setitem__("decoder", "autoregressive"),
+                cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0}),
+            ),
+        ),
+    ],
+)
+def test_train_end_to_end(label, mutate):
+    cfg = _base_cfg()
+    mutate(cfg)
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        assert (out_dir / "last.pt").exists()
+        assert (out_dir / "metrics.csv").exists()
+        ckpt = torch.load(out_dir / "last.pt", weights_only=False)
+        if cfg["stage1_model"]["active"]:
+            assert "model" in ckpt
+        else:
+            assert "model" not in ckpt
+        if cfg["stage2_model"]["active"]:
+            assert "sec_decoder" in ckpt
+        else:
+            assert "sec_decoder" not in ckpt
+
+
+def test_train_resume_continues_from_checkpoint():
+    cfg = _base_cfg()
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        ckpt_before = torch.load(out_dir / "last.pt", weights_only=False)
+        assert ckpt_before["epoch"] == 2
+
+        cfg2 = copy.deepcopy(cfg)
+        cfg2["train"]["epochs"] = 3
+        _run_train(cfg2, out_dir, resume_path=out_dir / "last.pt")
+        ckpt_after = torch.load(out_dir / "last.pt", weights_only=False)
+        assert ckpt_after["epoch"] == 3
+        assert ckpt_after["global_step"] > ckpt_before["global_step"]
+
+
+def test_train_raises_when_no_active_stage():
+    cfg = _base_cfg()
+    cfg["stage1_model"]["active"] = False
+    cfg["stage2_model"]["active"] = False
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    with tempfile.TemporaryDirectory() as tmp:
+        with pytest.raises(ValueError, match="no active stage"):
+            train(
+                cfg=cfg,
+                models=models,
+                critics=critics,
+                train_loader=_fake_batches(1, 8),
+                val_loader=_fake_batches(1, 8),
+                device=torch.device("cpu"),
+                out_dir=Path(tmp) / "run",
+                model_config=model_config,
+                total_train_batches=1,
+            )
+
+
+def test_metrics_csv_columns_are_stage_prefixed():
+    cfg = _base_cfg()
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        header = (out_dir / "metrics.csv").read_text().splitlines()[0].split(",")
+        assert "stage1/train/loss" in header
+        assert "stage2/train/d_loss" in header
+        assert "val/loss" in header
+        assert "epoch" in header
+
+
+def test_wgan_stage_trainer_skips_generator_step_when_no_grad_this_batch():
+    """Regression test: on a non-generator-step batch, if this stage's model
+    has no n_sec_head (n_sec defaults to stage 2), g_loss is a
+    graph-less zero — .backward() must not be called on it."""
+    cfg = _base_cfg()
+    cfg["stage1_model"]["generator"] = "wgan"
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    assert models["stage1"] is not None and critics["stage1"] is not None
+    spec = StageSpec(
+        name="stage1",
+        is_stage2=False,
+        generator="wgan",
+        n_critic=1000,  # never a generator step in this test
+        ema_decay=0.0,
+        steps_per_epoch=4,
+    )
+    trainer = WGANStageTrainer(spec, models["stage1"], critics["stage1"], torch.device("cpu"))
+    assert trainer.model.n_sec_head is None
+    batch = _fake_batches(1, 8)[0]
+    stats = trainer.step(batch, torch.device("cpu"), global_step=1)
+    assert stats["did_g_step"] is False
+
+
+def test_flow_stage_trainer_ddpm_not_implemented_for_stage2():
+    spec = StageSpec(name="stage2", is_stage2=True, generator="ddpm", ddpm_n_steps=50, ema_decay=0.0)
+    with pytest.raises(NotImplementedError):
+        FlowDDPMStageTrainer(spec, torch.nn.Linear(1, 1), torch.device("cpu"))
+
+
+def test_stage_spec_from_config_omitted_decoder_and_particle_type_match_default_config():
+    """Regression for issues.md Issue 1: StageSpec.from_config's own fallback
+    defaults for stage2_model.decoder/particle_type must equal
+    DEFAULT_CONFIG's ("autoregressive" / "onehot"), not the old, now-wrong
+    ("one_shot" / "physical") literals a .get(key, default) call used to
+    supply when a hand-built cfg omitted these keys."""
+    cfg = _base_cfg()
+    del cfg["stage2_model"]["decoder"]
+    del cfg["stage2_model"]["particle_type"]
+    spec = StageSpec.from_config(cfg, "stage2", is_stage2=True, steps_per_epoch=1)
+    assert spec.decoder == "autoregressive"
+    assert spec.particle_type.target == "onehot"
+
+
+def _routed_stage1_trainer(lambda_balance, lambda_proc, lambda_entropy):
+    cfg = _base_cfg()
+    cfg["stage1_model"]["router"] = {
+        "enabled": True,
+        "type": "energy",
+        "n_experts": 3,
+        "temperature": 0.5,
+        "learn_centers": True,
+        "lambda_balance": lambda_balance,
+        "lambda_proc": lambda_proc,
+        "lambda_entropy": lambda_entropy,
+    }
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    trainers = build_stage_trainers(cfg, models, critics, torch.device("cpu"), total_train_batches=4)
+    return trainers["stage1"]
+
+
+def test_router_aux_losses_skipped_when_lambda_zero_but_run_when_positive():
+    """Gitea #31: FlowDDPMStageTrainer._compute must not call
+    router.balance_loss/classify_loss/entropy_loss when the corresponding
+    lambda is 0 (the default) -- those calls do their own router.gate(...)
+    forward pass that is wasted once the term is masked out of the total
+    loss anyway. Checked both ways: zero lambdas must skip all three calls,
+    positive lambdas must still make them (the guard must not accidentally
+    suppress the real path)."""
+    batch = _fake_batches(1, 4)[0]
+    device = torch.device("cpu")
+
+    trainer_zero = _routed_stage1_trainer(0.0, 0.0, 0.0)
+    router_zero = trainer_zero.router
+    router_zero.balance_loss = MagicMock(wraps=router_zero.balance_loss)
+    router_zero.classify_loss = MagicMock(wraps=router_zero.classify_loss)
+    router_zero.entropy_loss = MagicMock(wraps=router_zero.entropy_loss)
+    stats_zero = trainer_zero.step(batch, device, global_step=1)
+    assert router_zero.balance_loss.call_count == 0
+    assert router_zero.classify_loss.call_count == 0
+    assert router_zero.entropy_loss.call_count == 0
+    assert stats_zero["loss_balance"] == 0.0
+    assert stats_zero["loss_proc"] == 0.0
+    assert stats_zero["loss_entropy"] == 0.0
+
+    trainer_pos = _routed_stage1_trainer(0.1, 0.1, 0.01)
+    router_pos = trainer_pos.router
+    router_pos.balance_loss = MagicMock(wraps=router_pos.balance_loss)
+    router_pos.classify_loss = MagicMock(wraps=router_pos.classify_loss)
+    router_pos.entropy_loss = MagicMock(wraps=router_pos.entropy_loss)
+    trainer_pos.step(batch, device, global_step=1)
+    assert router_pos.balance_loss.call_count == 1
+    assert router_pos.classify_loss.call_count == 1
+    assert router_pos.entropy_loss.call_count == 1
+
+
+# --- AR trainer wiring (v0.3.0 step 5) --------------------------------------
+
+
+@pytest.mark.parametrize("teacher_forcing", ["always", "scheduled", "never"])
+@pytest.mark.parametrize("history", ["markov", "attention"])
+@pytest.mark.parametrize("stage2_generator", ["wgan", "flow"])
+def test_build_stage_trainers_ar_scheduled_and_attention_step_runs(teacher_forcing, history, stage2_generator):
+    """v0.3.0 step 7: history='attention' and teacher_forcing in
+    {'scheduled', 'never'} must actually train — a stage-2 AR trainer.step()
+    must run and produce a finite loss, for every {history} x
+    {teacher_forcing} x {generator} combination."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["decoder"] = "autoregressive"
+    cfg["stage2_model"]["generator"] = stage2_generator
+    cfg["stage2_model"]["autoregressive"] = {
+        "history": history,
+        "teacher_forcing": teacher_forcing,
+        "tf_p_start": 1.0,
+        "tf_p_end": 0.0,
+        "attn_n_heads": 2,
+        "attn_n_layers": 1,
+    }
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    trainers = build_stage_trainers(cfg, models, critics, torch.device("cpu"), total_train_batches=4)
+    trainer = trainers["stage2"]
+    batch = _fake_batches(1, 4)[0]
+    stats = trainer.step(batch, torch.device("cpu"), global_step=1)
+    loss_key = "g_loss" if stage2_generator == "wgan" else "loss"
+    assert math.isfinite(stats[loss_key])
+
+
+@pytest.mark.parametrize("stage2_generator", ["wgan", "flow"])
+def test_train_end_to_end_ar_attention_history_scheduled_teacher_forcing(
+    stage2_generator,
+):
+    """Full `train()` run (not just one `trainer.step()` call) with
+    history='attention' AND teacher_forcing='scheduled' together — the
+    combination v0.3.0 step 7 exists to land — must complete and write a
+    checkpoint + metrics.csv with finite losses throughout."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["decoder"] = "autoregressive"
+    cfg["stage2_model"]["generator"] = stage2_generator
+    cfg["stage2_model"]["autoregressive"] = {
+        "history": "attention",
+        "teacher_forcing": "scheduled",
+        "tf_p_start": 1.0,
+        "tf_p_end": 0.0,
+        "attn_n_heads": 2,
+        "attn_n_layers": 1,
+    }
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        assert (out_dir / "last.pt").exists()
+        with open(out_dir / "metrics.csv", newline="") as f:
+            rows = list(csv.DictReader(f))
+        assert len(rows) == cfg["train"]["epochs"]
+        loss_col = "stage2/train/g_loss" if stage2_generator == "wgan" else "stage2/train/loss"
+        assert all(math.isfinite(float(r[loss_col])) for r in rows)
+
+
+def test_ar_wgan_onehot_grad_norm_instrumentation_populates_metrics():
+    """Differentiability validation-obligation instrumentation: the
+    trunk-gradient-norm-by-slice columns must appear and actually fire for
+    generator='wgan' + particle_type.target='onehot' under decoder=
+    'autoregressive' (added at v0.3.0 step 5 to accrue evidence during the
+    architecture comparison)."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["decoder"] = "autoregressive"
+    cfg["stage2_model"]["particle_type"] = {"target": "onehot", "lambda": 1.0}
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        with open(out_dir / "metrics.csv", newline="") as f:
+            rows = list(csv.DictReader(f))
+        assert "stage2/train/grad_norm_type_slice" in rows[0]
+        assert "stage2/train/grad_norm_cont_slice" in rows[0]
+        assert any(float(r["stage2/train/grad_norm_type_slice"]) > 0 for r in rows)
+        assert any(float(r["stage2/train/grad_norm_cont_slice"]) > 0 for r in rows)
+
+
+def test_wgan_onehot_one_shot_also_gets_grad_norm_instrumentation():
+    """The instrumentation is decoder-agnostic — one_shot + wgan + onehot
+    must populate the same columns."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["particle_type"] = {"target": "onehot", "lambda": 1.0}
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        with open(out_dir / "metrics.csv", newline="") as f:
+            rows = list(csv.DictReader(f))
+        assert any(float(r["stage2/train/grad_norm_type_slice"]) > 0 for r in rows)
+        assert any(float(r["stage2/train/grad_norm_cont_slice"]) > 0 for r in rows)
+
+
+def test_wgan_physical_omits_grad_norm_slice_columns():
+    cfg = _base_cfg()  # _base_cfg's stage2_model.particle_type.target is "physical"
+    with tempfile.TemporaryDirectory() as tmp:
+        out_dir = Path(tmp) / "run"
+        _run_train(cfg, out_dir)
+        header = (out_dir / "metrics.csv").read_text().splitlines()[0].split(",")
+        assert "stage2/train/grad_norm_type_slice" not in header
+        assert "stage2/train/grad_norm_cont_slice" not in header

tests/test_transforms.py

@@ -2,6 +2,7 @@ import warnings
 
 import numpy as np
 import pytest
+from giant.cond_layout import AXIS_TYPES, CondLayout
 from giant.constants import COND_DIM, COND_DIM_BASE, K_MAX
 from giant.data.transforms import (
     build_cond_features,
@@ -162,9 +163,7 @@ def test_local_frame_rotation_normalizes_non_unit_pre_dir():
     post_dir = rng.standard_normal((N, 3)).astype(np.float32)
     post_dir /= np.linalg.norm(post_dir, axis=1, keepdims=True)
 
-    pre_dir_scaled = pre_dir_unit * rng.uniform(0.9, 1.1, size=(N, 1)).astype(
-        np.float32
-    )
+    pre_dir_scaled = pre_dir_unit * rng.uniform(0.9, 1.1, size=(N, 1)).astype(np.float32)
     expected = local_frame_rotation(pre_dir_unit, post_dir)
     result = local_frame_rotation(pre_dir_scaled, post_dir)
     np.testing.assert_allclose(result, expected, atol=1e-4)
@@ -200,14 +199,10 @@ def test_reconstruct_post_pos_straight_line():
     step_length = rng.uniform(0.1, 5.0, size=N).astype(np.float32)
     post_pos = pre_pos + step_length[:, None] * pre_dir
 
-    travel_dir_local = local_frame_rotation(
-        pre_dir, travel_direction(pre_pos, post_pos)
-    )
+    travel_dir_local = local_frame_rotation(pre_dir, travel_direction(pre_pos, post_pos))
     np.testing.assert_allclose(travel_dir_local, np.tile([0, 0, 1], (N, 1)), atol=1e-4)
 
-    reconstructed = reconstruct_post_pos(
-        pre_pos, pre_dir, step_length, travel_dir_local
-    )
+    reconstructed = reconstruct_post_pos(pre_pos, pre_dir, step_length, travel_dir_local)
     np.testing.assert_allclose(reconstructed, post_pos, atol=1e-4)
 
 
@@ -221,12 +216,8 @@ def test_reconstruct_post_pos_general_roundtrip():
     post_pos = pre_pos + rng.standard_normal((N, 3)).astype(np.float32)
     step_length = np.linalg.norm(post_pos - pre_pos, axis=1).astype(np.float32)
 
-    travel_dir_local = local_frame_rotation(
-        pre_dir, travel_direction(pre_pos, post_pos)
-    )
-    reconstructed = reconstruct_post_pos(
-        pre_pos, pre_dir, step_length, travel_dir_local
-    )
+    travel_dir_local = local_frame_rotation(pre_dir, travel_direction(pre_pos, post_pos))
+    reconstructed = reconstruct_post_pos(pre_pos, pre_dir, step_length, travel_dir_local)
     np.testing.assert_allclose(reconstructed, post_pos, atol=1e-4)
 
 
@@ -319,7 +310,7 @@ def test_build_features_clamps_n_sec_label_to_k_max():
     pdg_map = {11: 0}
     mat_map = {"PbWO4": 0}
 
-    _, _, _, n_sec, _, _, _, _ = build_features(data, pdg_map, mat_map)
+    _, _, _, n_sec, _, _, _, _, _ = build_features(data, pdg_map, mat_map)
 
     assert n_sec.max() <= K_MAX
     np.testing.assert_array_equal(n_sec, [0, 5, K_MAX])
@@ -356,24 +347,20 @@ def _step_data_no_sec_lists(n_sec: np.ndarray) -> dict:
 
 
 def test_build_features_proc_idx_zero_without_proc_map():
-    data = _minimal_step_data(
-        3, process=np.array(["compt", "phot", "eIoni"], dtype=object)
-    )
+    data = _minimal_step_data(3, process=np.array(["compt", "phot", "eIoni"], dtype=object))
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
 
-    *_, proc_idx, _, _ = build_features(data, pdg_map, mat_map)
+    *_, proc_idx, _, _, _ = build_features(data, pdg_map, mat_map)
 
     np.testing.assert_array_equal(proc_idx, [0, 0, 0])
 
 
 def test_build_features_proc_idx_looks_up_proc_map():
-    data = _minimal_step_data(
-        3, process=np.array(["compt", "phot", "eIoni"], dtype=object)
-    )
+    data = _minimal_step_data(3, process=np.array(["compt", "phot", "eIoni"], dtype=object))
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
     proc_map = {"compt": 0, "phot": 1, "eIoni": 2}
 
-    *_, proc_idx, _, _ = build_features(data, pdg_map, mat_map, proc_map=proc_map)
+    *_, proc_idx, _, _, _ = build_features(data, pdg_map, mat_map, proc_map=proc_map)
 
     np.testing.assert_array_equal(proc_idx, [0, 1, 2])
 
@@ -396,9 +383,7 @@ def test_build_features_require_secondaries_ok_when_no_secondaries():
     data = _step_data_no_sec_lists(np.zeros(3, dtype=np.int32))
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
 
-    _, _, _, _, sec_cont, *_ = build_features(
-        data, pdg_map, mat_map, require_secondaries=True
-    )
+    _, _, _, _, sec_cont, *_ = build_features(data, pdg_map, mat_map, require_secondaries=True)
 
     assert not sec_cont.any()
 
@@ -413,11 +398,7 @@ def fake_material_props(monkeypatch):
     in by the user (see giant.materials.MaterialPropertiesNotFilledError)."""
     import giant.materials as gm
 
-    fake = {
-        "PbWO4": gm.MaterialProperties(
-            z_eff=75.6, a_eff=205.3, density=8.28, x0=0.89, lambda_int=20.7
-        )
-    }
+    fake = {"PbWO4": gm.MaterialProperties(z_eff=75.6, a_eff=205.3, density=8.28, x0=0.89, lambda_int=20.7)}
     monkeypatch.setattr(gm, "MATERIAL_PROPERTIES", fake)
     return fake
 
@@ -426,7 +407,13 @@ def test_build_features_embedding_mode_zero_fills_physical_columns():
     data = _minimal_step_data(3)
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
 
-    cond_cont, *_ = build_features(data, pdg_map, mat_map, conditioning="embedding")
+    cond_cont, *_ = build_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning="embedding",
+        material_conditioning="embedding",
+    )
 
     assert cond_cont.shape[1] == COND_DIM
     np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE:], 0.0)
@@ -438,14 +425,18 @@ def test_build_features_physical_mode_shape_and_values(fake_material_props):
     data = _minimal_step_data(3)
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
 
-    cond_cont, *_ = build_features(data, pdg_map, mat_map, conditioning="physical")
+    cond_cont, *_ = build_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning="physical",
+        material_conditioning="physical",
+    )
 
     assert cond_cont.shape[1] == COND_DIM
     mass, charge = particle_mass_charge(11)
     expected_log_mass = log_transform(np.array([mass]))[0]
-    np.testing.assert_allclose(
-        cond_cont[:, COND_DIM_BASE], expected_log_mass, atol=1e-5
-    )
+    np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE], expected_log_mass, atol=1e-5)
     np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE + 1], charge)
     np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE + 2], 75.6)  # z_eff
     np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE + 3], 205.3)  # a_eff
@@ -461,7 +452,13 @@ def test_build_features_physical_mode_unfilled_material_raises():
     pdg_map, mat_map = {11: 0}, {"G4_LYSO": 0}
 
     with pytest.raises(MaterialPropertiesNotFilledError):
-        build_features(data, pdg_map, mat_map, conditioning="physical")
+        build_features(
+            data,
+            pdg_map,
+            mat_map,
+            particle_conditioning="physical",
+            material_conditioning="physical",
+        )
 
 
 def test_build_cond_features_mass_charge_override(fake_material_props):
@@ -474,11 +471,15 @@ def test_build_cond_features_mass_charge_override(fake_material_props):
     data["charge"] = np.array([2.0, -2.0], dtype=np.float32)
     pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
 
-    cond_cont, _ = build_cond_features(data, pdg_map, mat_map, conditioning="physical")
-
-    np.testing.assert_allclose(
-        cond_cont[:, COND_DIM_BASE], log_transform(np.array([123.0, 456.0]))
+    cond_cont, _ = build_cond_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning="physical",
+        material_conditioning="physical",
     )
+
+    np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE], log_transform(np.array([123.0, 456.0])))
     np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE + 1], [2.0, -2.0])
 
 
@@ -494,7 +495,12 @@ def test_build_cond_features_pads_legacy_normalizer_in_embedding_mode():
     legacy_norm.std = np.ones(COND_DIM_BASE, dtype=np.float32)
 
     cond_cont, _ = build_cond_features(
-        data, pdg_map, mat_map, cond_normalizer=legacy_norm, conditioning="embedding"
+        data,
+        pdg_map,
+        mat_map,
+        cond_normalizer=legacy_norm,
+        particle_conditioning="embedding",
+        material_conditioning="embedding",
     )
 
     assert cond_cont.shape[-1] == COND_DIM
@@ -521,10 +527,134 @@ def test_build_cond_features_rejects_legacy_normalizer_in_physical_mode(
             pdg_map,
             mat_map,
             cond_normalizer=legacy_norm,
-            conditioning="physical",
+            particle_conditioning="physical",
+            material_conditioning="physical",
         )
 
 
+# ── build_cond_features / build_features share one column layout (gitea #37) ──
+
+
+@pytest.mark.parametrize("particle_type", AXIS_TYPES)
+@pytest.mark.parametrize("material_type", AXIS_TYPES)
+def test_both_builders_agree_column_for_column(particle_type, material_type, fake_material_props):
+    """The two builders used to lay out cond_cont/cond_cat independently and
+    drift apart silently. They now share `_build_cond_arrays`, so for every
+    mode pair they must produce identical arrays."""
+    data = _minimal_step_data(3)
+    pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
+    pdg_topn = {11: 0} if particle_type == "onehot" else None
+    mat_topn = {"PbWO4": 0} if material_type == "onehot" else None
+
+    cond_cont, cond_cat = build_cond_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning=particle_type,
+        material_conditioning=material_type,
+        pdg_topn_map=pdg_topn,
+        mat_topn_map=mat_topn,
+    )
+    feats = build_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning=particle_type,
+        material_conditioning=material_type,
+        pdg_topn_map=pdg_topn,
+        mat_topn_map=mat_topn,
+    )
+
+    layout = CondLayout.from_types(particle_type, material_type)
+    assert cond_cat.shape[1] == layout.cat_dim
+    np.testing.assert_array_equal(feats.cond_cont, cond_cont)
+    np.testing.assert_array_equal(feats.cond_cat, cond_cat)
+
+
+def test_build_features_physical_mode_tolerates_out_of_vocab_pdg_and_material():
+    """The permissive vocab lookup added for "physical"/"onehot" mode (see
+    build_cond_features) applies to build_features too — `giant predict` on a
+    file whose pdg/material aren't in the checkpoint's dense vocab must not
+    KeyError when nothing reads those indices."""
+    pdg_map = {11: 0, 22: 1}
+    mat_map = {"G4_AIR": 0}
+    data = _minimal_step_data(2)
+    data["pdg"] = np.full(2, 13, dtype=np.int64)  # not in pdg_map
+    data["material"] = np.full(2, "G4_Pb", dtype=object)  # not in mat_map
+
+    _, cond_cat, *_ = build_features(
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning="physical",
+        material_conditioning="physical",
+    )
+    np.testing.assert_array_equal(cond_cat, [[0, 0], [0, 0]])  # dummy indices, no raise
+
+    with pytest.raises(KeyError):
+        build_features(
+            data,
+            pdg_map,
+            mat_map,
+            particle_conditioning="embedding",
+            material_conditioning="embedding",
+        )
+
+
+def test_build_features_pads_legacy_normalizer_in_embedding_mode():
+    """The legacy-normalizer padding (a pre-physical-conditioning checkpoint's
+    cond normalizer is COND_DIM_BASE wide) applies to build_features too —
+    `giant predict` reaches build_features, not build_cond_features."""
+    data = _minimal_step_data(3)
+    pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
+    legacy_norm = Normalizer()
+    legacy_norm.mean = np.zeros(COND_DIM_BASE, dtype=np.float32)
+    legacy_norm.std = np.ones(COND_DIM_BASE, dtype=np.float32)
+
+    cond_cont, *_ = build_features(
+        data,
+        pdg_map,
+        mat_map,
+        cond_normalizer=legacy_norm,
+        particle_conditioning="embedding",
+        material_conditioning="embedding",
+    )
+
+    assert cond_cont.shape[-1] == COND_DIM
+    np.testing.assert_allclose(cond_cont[:, COND_DIM_BASE:], 0.0)
+
+
+def test_build_features_rejects_legacy_normalizer_in_physical_mode(fake_material_props):
+    data = _minimal_step_data(3)
+    pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
+    legacy_norm = Normalizer()
+    legacy_norm.mean = np.zeros(COND_DIM_BASE, dtype=np.float32)
+    legacy_norm.std = np.ones(COND_DIM_BASE, dtype=np.float32)
+
+    with pytest.raises(ValueError, match="predates physical-property conditioning"):
+        build_features(
+            data,
+            pdg_map,
+            mat_map,
+            cond_normalizer=legacy_norm,
+            particle_conditioning="physical",
+            material_conditioning="physical",
+        )
+
+
+def test_onehot_axis_without_its_topn_map_raises():
+    """`cond_cat`'s width is the layout's call, so a "onehot" axis with no
+    top-N map is a hard error rather than a silently-narrower array that
+    ConditionEncoder would then index out of bounds."""
+    data = _minimal_step_data(2)
+    pdg_map, mat_map = {11: 0}, {"PbWO4": 0}
+
+    with pytest.raises(ValueError, match="needs pdg_topn_map"):
+        build_cond_features(data, pdg_map, mat_map, particle_conditioning="onehot")
+    with pytest.raises(ValueError, match="needs mat_topn_map"):
+        build_cond_features(data, pdg_map, mat_map, material_conditioning="onehot")
+
+
 # ── sorted_membership / _vectorized_map_lookup ──────────────────────────────
 
 
@@ -610,13 +740,23 @@ def test_build_cond_features_physical_mode_tolerates_out_of_vocab_pdg_and_materi
     }
 
     cond_cont, cond_cat = build_cond_features(
-        data, pdg_map, mat_map, conditioning="physical"
+        data,
+        pdg_map,
+        mat_map,
+        particle_conditioning="physical",
+        material_conditioning="physical",
     )
     assert cond_cont.shape[-1] == COND_DIM
     np.testing.assert_array_equal(cond_cat, [[0, 0]])  # dummy indices, no raise
 
     with pytest.raises(KeyError):
-        build_cond_features(data, pdg_map, mat_map, conditioning="embedding")
+        build_cond_features(
+            data,
+            pdg_map,
+            mat_map,
+            particle_conditioning="embedding",
+            material_conditioning="embedding",
+        )
 
 
 # ── _WelfordAccumulator ──────────────────────────────────────────────────────
@@ -674,9 +814,7 @@ def test_welford_accumulator_matches_naive_running_mean_reference():
     naive_M2 = np.zeros(F)
     naive_n = 0
     for chunk in chunks:
-        naive_mean, naive_M2, naive_n = naive_update(
-            naive_mean, naive_M2, naive_n, chunk
-        )
+        naive_mean, naive_M2, naive_n = naive_update(naive_mean, naive_M2, naive_n, chunk)
 
     acc = _WelfordAccumulator(F)
     for chunk in chunks:

tests/test_type_embedding_distance.py

@@ -0,0 +1,43 @@
+"""Tests for the secondary-type embedding-distance diagnostic
+(giant.analysis.type_embedding_distance)."""
+
+from __future__ import annotations
+
+from giant.analysis.type_embedding_distance import compute_type_embedding_l1_distance
+
+
+def _summary(n=100):
+    return {
+        "n": n,
+        "mean": 1.23,
+        "std": 0.45,
+        "min": 0.01,
+        "max": 9.87,
+        "hist_edges": [0.0, 1.0, 2.0, 3.0],
+        "hist_counts": [30, 40, 30],
+    }
+
+
+def test_none_is_unavailable():
+    r = compute_type_embedding_l1_distance(None)
+    assert r.kind == "unavailable"
+    assert r.id == "type_embedding_l1_distance"
+    assert r.payload["note"]
+
+
+def test_summary_produces_single_hist():
+    r = compute_type_embedding_l1_distance(_summary())
+    assert r.kind == "single_hist"
+    assert r.id == "type_embedding_l1_distance"
+    assert r.payload["edges"] == [0.0, 1.0, 2.0, 3.0]
+    assert r.payload["rollout"] == [30, 40, 30]
+    assert r.payload["log_x"] is True
+    assert r.payload["log_y"] is True
+    assert "n=100" in r.payload["note"]
+
+
+def test_single_hist_payload_shape_matches_render_contract():
+    """_render_single (giant.analysis.render) requires len(rollout) ==
+    len(edges) - 1."""
+    r = compute_type_embedding_l1_distance(_summary())
+    assert len(r.payload["rollout"]) == len(r.payload["edges"]) - 1

tests/test_validate.py

@@ -1,30 +1,65 @@
 import numpy as np
 import torch
 
-from giant.constants import COND_DIM, K_MAX, SEC_SLOT_DIM, X_DIM
-from giant.model.network import DenoisingMLP, SecondaryDecoder
+from giant.config import ConditioningAxisConfig, ParticleTypeConfig
+from giant.constants import COND_DIM, SEC_SLOT_DIM, X_DIM
+from giant.data.dataset import StepBatch
+from giant.model.network import Stage1Model, Stage2OneShot, stage2_trunk_sec_dim
 from giant.validate import validate_marginals
 
+_PARTICLE_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+_MATERIAL_CFG = ConditioningAxisConfig(type="physical", emb_dim=8, n_layers=1)
+_K_MAX = 5
 
-def _tiny_models():
-    s1 = DenoisingMLP(pdg_vocab=3, mat_vocab=2, hidden_dim=16, n_blocks=1)
-    s2 = SecondaryDecoder(pdg_vocab=3, mat_vocab=2, hidden_dim=16, n_blocks=1)
+
+def _tiny_models(particle_type_cfg: ParticleTypeConfig | None = None):
+    """A fresh v0.3.0 pair: Stage1Model owns no n_sec_head, so n_sec always
+    comes from Stage2OneShot."""
+    s1 = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PARTICLE_CFG,
+        material_cfg=_MATERIAL_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+    )
+    resolved_type_cfg = particle_type_cfg or ParticleTypeConfig(target="physical")
+    target = resolved_type_cfg.target
+    sec_dim = stage2_trunk_sec_dim(
+        resolved_type_cfg,
+        "flow",
+        _K_MAX,
+        _PARTICLE_CFG.emb_dim,
+    )
+    s2 = Stage2OneShot(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=_PARTICLE_CFG,
+        material_cfg=_MATERIAL_CFG,
+        hidden_dim=16,
+        n_res_blocks=1,
+        generator="flow",
+        time_dim=16,
+        k_max=_K_MAX,
+        sec_dim=sec_dim,
+        particle_type_cfg=particle_type_cfg,
+    )
+    assert s2.particle_type_cfg.target == target
     return s1.eval(), s2.eval()
 
 
-def _zero_secondaries_loader(B=4, n_batches=2):
-    """A val_loader whose every batch has n_sec=0 (real side) — matches the
-    (cond_cont, cond_cat, target_s1, n_sec, sec_cont, proc_idx) tuple shape
-    StreamingStepsDataset yields."""
+def _loader(B: int = 4, n_batches: int = 2, n_sec_value: int = 0, n_classes: int = 8):
+    """A val_loader matching StreamingStepsDataset's StepBatch shape."""
     batches = []
     for _ in range(n_batches):
         cond_cont = torch.randn(B, COND_DIM)
         cond_cat = torch.zeros(B, 2, dtype=torch.long)
         x1 = torch.randn(B, X_DIM)
-        n_sec = torch.zeros(B, dtype=torch.long)
-        sec_cont = torch.zeros(B, K_MAX, SEC_SLOT_DIM)
+        n_sec = torch.full((B,), n_sec_value, dtype=torch.long)
+        sec_cont = torch.randn(B, _K_MAX, SEC_SLOT_DIM)
         proc_idx = torch.zeros(B, dtype=torch.long)
-        batches.append((cond_cont, cond_cat, x1, n_sec, sec_cont, proc_idx))
+        sec_type_idx = torch.randint(0, n_classes, (B, _K_MAX), dtype=torch.long)
+        batches.append(StepBatch(cond_cont, cond_cat, x1, n_sec, sec_cont, proc_idx, sec_type_idx))
     return batches
 
 
@@ -32,22 +67,54 @@ def test_validate_marginals_all_zero_secondaries_returns_nan_phys_kl(monkeypatch
     """If n_sec_pred collapses to 0 across the whole validated set (realistic
     during early/unstable training), phys_kl must degrade to NaN instead of
     crashing on the empty-array .min()/.max() reduction inside
-    _histogram_kl -- a regression the old species/bincount code this
-    replaced explicitly guarded against."""
+    _histogram_kl."""
     s1, s2 = _tiny_models()
-    loader = _zero_secondaries_loader()
+    loader = _loader(n_sec_value=0)
 
-    # Force the Stage-1 n_sec head's prediction to 0 for every sample too, so
-    # the generated side's valid-slot mask is also empty (real side is
-    # already all n_sec=0 by construction of the fake loader above).
-    def _fake_sample_flow(model, cond_cont, cond_cat, **kw):
-        B = cond_cont.size(0)
-        return torch.randn(B, X_DIM), torch.zeros(B, dtype=torch.long)
+    def _fake_resolve_n_sec(stage1_model, sec_decoder, cond_cont, cond_cat, stage1_out, n_sec_pred):
+        return torch.zeros(cond_cont.size(0), dtype=torch.long)
 
-    monkeypatch.setattr("giant.validate.sample_flow", _fake_sample_flow)
+    monkeypatch.setattr("giant.validate.resolve_n_sec", _fake_resolve_n_sec)
 
-    result = validate_marginals(s1, loader, sec_decoder=s2, n_batches=2)
+    result = validate_marginals(s1, loader, sec_decoder=s2, n_batches=2, steps=2)
 
     assert np.asarray(result["phys_real"]).shape == (0, 2)
     assert np.asarray(result["phys_generated"]).shape == (0, 2)
     assert np.isnan(np.asarray(result["phys_kl"])).all()
+
+
+def test_validate_marginals_physical_target_shapes():
+    s1, s2 = _tiny_models()
+    loader = _loader(n_sec_value=2)
+
+    result = validate_marginals(s1, loader, sec_decoder=s2, n_batches=1, steps=2)
+
+    assert np.asarray(result["real"]).shape == (4, X_DIM)
+    assert np.asarray(result["generated"]).shape == (4, X_DIM)
+    assert np.asarray(result["kl_divergence"]).shape == (X_DIM,)
+    assert "phys_real" in result and "phys_generated" in result and "phys_kl" in result
+    assert "type_class_real" not in result
+
+
+def test_validate_marginals_onehot_type_class_marginal():
+    particle_type_cfg = ParticleTypeConfig(target="onehot")
+    s1, s2 = _tiny_models(particle_type_cfg)
+    loader = _loader(n_sec_value=2, n_classes=s2.type_dim)
+
+    result = validate_marginals(s1, loader, sec_decoder=s2, n_batches=1, steps=2)
+
+    assert "phys_real" not in result
+    # Real/generated valid-slot counts need not agree (real: ground-truth
+    # n_sec=2 always; generated: the untrained n_sec_head's own prediction).
+    assert np.asarray(result["type_class_real"]).ndim == 1
+    assert np.asarray(result["type_class_gen"]).ndim == 1
+    assert np.asarray(result["type_class_real"]).shape[0] > 0
+
+
+def test_validate_marginals_without_sec_decoder_returns_stage1_only():
+    s1, _ = _tiny_models()
+    loader = _loader(n_sec_value=0)
+
+    result = validate_marginals(s1, loader, n_batches=1, steps=2)
+
+    assert set(result) == {"real", "generated", "kl_divergence"}