chore: update changelog for v0.3.6 [skip ci]

Reviewed-on: #74

.bumpversion.toml

@@ -0,0 +1,17 @@
+[tool.bumpversion]
+current_version = "0.3.6"
+parse = "(?P<major>\\d+)\\.(?P<minor>\\d+)\\.(?P<patch>\\d+)"
+serialize = ["{major}.{minor}.{patch}"]
+search = "{current_version}"
+replace = "{new_version}"
+regex = false
+allow_dirty = false
+commit = true
+tag = false
+message = "chore: bump version {current_version} -> {new_version} [skip ci]"
+pre_commit_hooks = ["uv lock", "git add uv.lock"]
+
+[[tool.bumpversion.files]]
+filename = "pyproject.toml"
+search = "version = \"{current_version}\""
+replace = "version = \"{new_version}\""

.gitea/workflows/ci.yml

@@ -88,6 +88,92 @@ jobs:
           name: coverage-report
           path: coverage.xml
 
+  bump-version:
+    name: Bump version, tag, and update changelog on merge to master
+    needs: [ruff-check, ruff-format, type-check, test]
+    if: github.ref == 'refs/heads/master' && github.event_name == 'push'
+    runs-on: ubuntu-latest
+    container:
+      image: docker.gitea.com/runner-images:ubuntu-latest
+      volumes:
+        - /srv/act-runner-cache/uv:/uv-cache
+    steps:
+      # CI_TOKEN needs write:repository scope (not just read) — this job
+      # pushes commits and tags to master, unlike ruff-check/ruff-format/
+      # type-check/test above, which only need to check out the repo.
+      - uses: actions/checkout@v4
+        with:
+          token: ${{ secrets.CI_TOKEN }}
+          fetch-depth: 0
+      - name: Check whether this push is a merge commit
+        id: merge_check
+        run: |
+          PARENTS=$(git rev-parse HEAD^@ | wc -l)
+          echo "HEAD has $PARENTS parent(s)"
+          if [ "$PARENTS" -ge 2 ]; then
+            echo "is_merge=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "is_merge=false" >> "$GITHUB_OUTPUT"
+          fi
+      - uses: astral-sh/setup-uv@v5
+        if: steps.merge_check.outputs.is_merge == 'true'
+        with:
+          enable-cache: false
+      - run: |
+          echo "UV_CACHE_DIR=/uv-cache" >> "$GITHUB_ENV"
+          echo "UV_LINK_MODE=copy" >> "$GITHUB_ENV"
+        if: steps.merge_check.outputs.is_merge == 'true'
+      - run: uv sync --extra cpu --extra dev
+        if: steps.merge_check.outputs.is_merge == 'true'
+      - name: Configure git identity
+        if: steps.merge_check.outputs.is_merge == 'true'
+        run: |
+          git config user.name "gitea-actions"
+          git config user.email "actions@git.larsbogner.de"
+      - name: Bump patch version if this merge didn't already bump it
+        if: steps.merge_check.outputs.is_merge == 'true'
+        run: |
+          OLD_VERSION=$(git show "${{ github.event.before }}:pyproject.toml" 2>/dev/null | grep -m1 '^version = ' | sed -E 's/version = "(.*)"/\1/')
+          CURRENT_VERSION=$(uv version --short)
+          if [ -z "$OLD_VERSION" ]; then
+            echo "Could not read pyproject.toml at github.event.before; falling back to HEAD^1"
+            OLD_VERSION=$(git show "HEAD^1:pyproject.toml" | grep -m1 '^version = ' | sed -E 's/version = "(.*)"/\1/')
+          fi
+          if [ "$OLD_VERSION" = "$CURRENT_VERSION" ]; then
+            echo "Version unchanged by this merge ($CURRENT_VERSION); bumping patch"
+            uv run bump-my-version bump patch --current-version "$CURRENT_VERSION"
+          else
+            echo "Branch already bumped the version ($OLD_VERSION -> $CURRENT_VERSION); skipping auto-bump"
+          fi
+      - name: Update changelog for the current version if not already tagged
+        if: steps.merge_check.outputs.is_merge == 'true'
+        run: |
+          VERSION=$(uv version --short)
+          TAG="v$VERSION"
+          if git rev-parse "$TAG" >/dev/null 2>&1; then
+            echo "Tag $TAG already exists; skipping changelog update"
+          else
+            uv run git-cliff --tag "$TAG" --unreleased --prepend CHANGELOG.md
+            git add CHANGELOG.md
+            if ! git diff --cached --quiet -- CHANGELOG.md; then
+              git commit -m "chore: update changelog for $TAG [skip ci]"
+            else
+              git restore --staged CHANGELOG.md
+            fi
+          fi
+      - name: Push commits and tag the current version
+        if: steps.merge_check.outputs.is_merge == 'true'
+        run: |
+          git push origin HEAD:master
+          VERSION=$(uv version --short)
+          TAG="v$VERSION"
+          if git rev-parse "$TAG" >/dev/null 2>&1; then
+            echo "Tag $TAG already exists"
+          else
+            git tag -a "$TAG" -m "$TAG"
+            git push origin "refs/tags/$TAG"
+          fi
+
   sync-version-on-tag:
     name: Sync project version with tag
     if: startsWith(github.ref, 'refs/tags/')

CHANGELOG.md

@@ -0,0 +1,40 @@
+# Changelog
+
+## [0.3.6] - 2026-08-24
+
+### Changed
+
+- Give CriticModel a registry-built trunk and StageModel base [gitea #57](https://git.larsbogner.de/lars/giant/issues/57)
+
+## [0.3.5] - 2026-08-24
+
+### Added
+
+- Add "none" variants for router, history, and trunk [gitea #45](https://git.larsbogner.de/lars/giant/issues/45)
+
+## [0.3.4] - 2026-08-23
+
+### Added
+
+- Add giant model summary command [gitea #46](https://git.larsbogner.de/lars/giant/issues/46)
+
+- Add per-stage init_from/freeze [gitea #42](https://git.larsbogner.de/lars/giant/issues/42)
+
+- Add bf16 autocast to the training loop [gitea #47](https://git.larsbogner.de/lars/giant/issues/47)
+
+- Add class-balanced secondary particle-type loss [gitea #44](https://git.larsbogner.de/lars/giant/issues/44)
+
+
+### Changed
+
+- Implement stage2_model.stage1_context = "sampled" [gitea #41](https://git.larsbogner.de/lars/giant/issues/41)
+
+- Bump patch version to 0.3.3
+
+- Offset event_id across multi-shard reference reads in giant analyze [gitea #22](https://git.larsbogner.de/lars/giant/issues/22)
+
+- Auto-bump patch version, tag, and update changelog on merge to master [gitea #50](https://git.larsbogner.de/lars/giant/issues/50)
+
+- Document CI_TOKEN's write:repository scope requirement [gitea #50](https://git.larsbogner.de/lars/giant/issues/50)
+
+# Changelog

README.md

@@ -139,6 +139,7 @@ Useful flags on `giant train`:
 - `--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
+- `--stage1-init-from`/`--stage2-init-from` (checkpoint `.pt`) + `--stage1-freeze`/`--stage2-freeze` — load a stage's weights from another checkpoint and never update them, so the other stage can be retrained alone against a fixed, known-good one while still producing a complete, rollout-capable checkpoint
 
 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).
 

cliff.toml

@@ -0,0 +1,52 @@
+# git-cliff configuration — see https://git-cliff.org/docs/configuration
+#
+# Commit messages in this repo aren't Conventional Commits; they're plain
+# imperative summaries like "Add class-balanced secondary particle-type loss
+# (gitea #44)". Parsing here is tuned to that convention rather than to
+# feat:/fix:-style prefixes.
+
+[changelog]
+header = "# Changelog\n\n"
+body = """
+{% if version %}\
+## [{{ version | trim_start_matches(pat="v") }}] - {{ timestamp | date(format="%Y-%m-%d") }}
+{% else %}\
+## [Unreleased]
+{% endif %}\
+{% for group, commits in commits | group_by(attribute="group") %}
+### {{ group | striptags | trim | upper_first }}
+{% for commit in commits %}
+- {{ commit.message | upper_first }}
+{% endfor %}
+{% endfor %}
+"""
+trim = true
+render_always = true
+postprocessors = []
+
+[git]
+conventional_commits = false
+filter_unconventional = false
+require_conventional = false
+split_commits = false
+# Keep only the commit subject (first line), then linkify "(gitea #N)".
+commit_preprocessors = [
+    { pattern = "(?s)\n.*", replace = "" },
+    { pattern = "\\(gitea #(\\d+)\\)", replace = "[gitea #${1}](https://git.larsbogner.de/lars/giant/issues/${1})" },
+]
+protect_breaking_commits = false
+commit_parsers = [
+    { message = "^Merge ", skip = true },
+    { message = "\\[skip ci\\]", skip = true },
+    { message = "^Add", group = "<!-- 0 -->Added" },
+    { message = "^(Fix|Clamp|Clip)", group = "<!-- 1 -->Fixed" },
+    { message = "^(Remove|Drop|Deprecate)", group = "<!-- 2 -->Removed" },
+    { message = ".*", group = "<!-- 3 -->Changed" },
+]
+filter_commits = false
+link_parsers = []
+use_branch_tags = false
+topo_order = false
+topo_order_commits = true
+sort_commits = "oldest"
+recurse_submodules = false

giant/analysis/sources.py

@@ -40,6 +40,11 @@ from giant.constants import (
     TERM_MAX_STEPS,
     TERM_UNKNOWN_PDG,
 )
+from giant.data.loader import event_id_offset, find_parquet_files
+
+# Helper column name for the per-shard offset join in open_side; dropped before
+# the LazyFrame is returned, so it never leaks into a caller's schema.
+_SOURCE_PATH_COL = "__source_path"
 
 # The world-frame physical columns both sides share under identical names.
 PHYS_COLS: tuple[str, ...] = (
@@ -108,6 +113,22 @@ def open_side(source: str | Path | pl.LazyFrame, side: Side) -> pl.LazyFrame:
     reference file's upstream ROOT→parquet conversion don't agree on integer
     width, and an uncast mismatch only surfaces later as a ``pl.concat``
     ``SchemaError`` (e.g. in ``build_context``'s pdg-count merge).
+
+    The reference (a rollout's seed ``dataset``) may be a directory of parquet
+    shards, or a ``.manifest`` naming a subset, rather than a single file — each
+    such shard is a separate Geant4 job whose own ``event_id`` numbering
+    restarts from 0, so a multi-shard load offsets every shard's ids by
+    ``giant.data.loader.event_id_offset(file_index)`` to keep them globally
+    unique, exactly as the training/rollout data pipeline already does
+    (``giant/data/loader.py``). ``file_index`` comes from
+    ``find_parquet_files``'s deterministic ordering — the same list and
+    ordering ``giant rollout`` used (via ``_seed_from_data``) to offset the
+    rollout side's own ``event_id``s, so both sides agree on what an
+    ``event_id`` means. There is no overflow guard here (unlike
+    ``loader._offset_event_id``): checking it would cost an eager
+    ``event_id``-column read per shard in every condor compute job, and
+    ``giant rollout`` already ran that check over this exact file list when it
+    produced the seed.
     """
     if isinstance(source, pl.LazyFrame):
         return source.with_columns(pl.col("pdg").cast(pl.Int64))
@@ -116,9 +137,18 @@ def open_side(source: str | Path | pl.LazyFrame, side: Side) -> pl.LazyFrame:
         _check_rollout_metadata(path)
         lf = pl.scan_parquet(path)
     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)
+        files = find_parquet_files(path)
+        if len(files) == 1:
+            lf = pl.scan_parquet(files[0])
+        else:
+            offsets = {str(p): event_id_offset(i) for i, p in enumerate(files)}
+            lf = (
+                pl.scan_parquet(files, include_file_paths=_SOURCE_PATH_COL)
+                .with_columns(
+                    pl.col("event_id") + pl.col(_SOURCE_PATH_COL).replace_strict(offsets, return_dtype=pl.Int64)
+                )
+                .drop(_SOURCE_PATH_COL)
+            )
     return lf.with_columns(pl.col("pdg").cast(pl.Int64))
 
 

giant/cli.py

@@ -41,6 +41,7 @@ from giant.data.transforms import (
 )
 from giant.checkpoint_io import CheckpointCompatibilityError, load_for_inference
 from giant.geometry import GeometryOracle
+from giant.materials import MATERIAL_PROPERTIES
 from giant.pipeline import run_train_job
 from giant.rollout import (
     L1DistCollector,
@@ -487,6 +488,36 @@ def train(
             help="WGAN-GP (--mode wgan only): critic depth for stage 2 (default: same as generator's n_res_blocks)",
         ),
     ] = None,
+    stage1_init_from: Annotated[
+        Optional[Path],
+        typer.Option(
+            "--stage1-init-from",
+            help="Checkpoint .pt to load stage 1's weights from before training starts "
+            "(gitea #42) — combine with --stage1-freeze to retrain stage 2 alone "
+            "against a fixed, known-good stage 1",
+        ),
+    ] = None,
+    stage1_freeze: Annotated[
+        Optional[bool],
+        typer.Option(
+            "--stage1-freeze/--no-stage1-freeze",
+            help="Never update stage 1's weights (requires --stage1-init-from, or --resume)",
+        ),
+    ] = None,
+    stage2_init_from: Annotated[
+        Optional[Path],
+        typer.Option(
+            "--stage2-init-from",
+            help="Checkpoint .pt to load stage 2's weights from before training starts (gitea #42)",
+        ),
+    ] = None,
+    stage2_freeze: Annotated[
+        Optional[bool],
+        typer.Option(
+            "--stage2-freeze/--no-stage2-freeze",
+            help="Never update stage 2's weights (requires --stage2-init-from, or --resume)",
+        ),
+    ] = None,
     val_fraction: Annotated[Optional[float], typer.Option("--val-fraction", "-f")] = None,
     seed: Annotated[
         Optional[int],
@@ -581,6 +612,14 @@ def train(
             "steps (default: 50); per-epoch metrics always log in full",
         ),
     ] = None,
+    precision: Annotated[
+        Optional[str],
+        typer.Option(
+            "--precision",
+            help="Training-step autocast precision: 'fp32' (default) or "
+            "'bf16'. No 'fp16' — see giant.training.amp.resolve_autocast",
+        ),
+    ] = None,
 ) -> None:
     """Train the GIANT surrogate model."""
     batch_size_auto = False
@@ -616,6 +655,7 @@ def train(
         "wandb_project": wandb_project,
         "wandb_run_name": wandb_run_name,
         "wandb_log_every": wandb_log_every,
+        "precision": precision,
         "hidden_dim": hidden_dim,
         "n_blocks": n_blocks,
         "dropout": dropout,
@@ -651,11 +691,15 @@ def train(
         "stage1_critic_n_res_blocks": stage1_critic_n_res_blocks,
         "stage2_critic_hidden_dim": stage2_critic_hidden_dim,
         "stage2_critic_n_res_blocks": stage2_critic_n_res_blocks,
+        "stage1_init_from": str(stage1_init_from) if stage1_init_from is not None else None,
+        "stage1_freeze": stage1_freeze,
+        "stage2_init_from": str(stage2_init_from) if stage2_init_from is not None else None,
+        "stage2_freeze": stage2_freeze,
     }
     overrides = gconfig.overrides_from_flags(flag_values)
 
     cfg = gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, config, overrides)
-    gconfig.validate_config(cfg)
+    gconfig.validate_config(cfg, resume=resume is not None)
     t = cfg["train"]
 
     _device = torch.device(device) if device else gconfig.auto_device()
@@ -690,6 +734,7 @@ def train(
 
     typer.echo(f"device:  {_device}")
     typer.echo(f"out_dir: {out_dir}")
+    typer.echo(f"precision: {t['precision']}")
 
     run_train_job(
         data=data,
@@ -735,6 +780,10 @@ def new_run(
     stage2_k_max: Annotated[Optional[int], typer.Option("--stage2-k-max")] = None,
     stage2_context_dim: Annotated[Optional[int], typer.Option("--stage2-context-dim")] = None,
     stage2_stage1_context: Annotated[Optional[Stage1Context], typer.Option("--stage2-stage1-context")] = None,
+    stage1_init_from: Annotated[Optional[Path], typer.Option("--stage1-init-from")] = None,
+    stage1_freeze: Annotated[Optional[bool], typer.Option("--stage1-freeze/--no-stage1-freeze")] = None,
+    stage2_init_from: Annotated[Optional[Path], typer.Option("--stage2-init-from")] = None,
+    stage2_freeze: Annotated[Optional[bool], typer.Option("--stage2-freeze/--no-stage2-freeze")] = None,
     conditioning: Annotated[Optional[Conditioning], typer.Option("--conditioning")] = None,
     router: Annotated[Optional[bool], typer.Option("--router/--no-router")] = None,
     router_type: Annotated[Optional[str], typer.Option("--router-type")] = None,
@@ -795,6 +844,10 @@ def new_run(
         "stage2_k_max": stage2_k_max,
         "stage2_context_dim": stage2_context_dim,
         "stage2_stage1_context": stage2_stage1_context.value if stage2_stage1_context is not None else None,
+        "stage1_init_from": str(stage1_init_from) if stage1_init_from is not None else None,
+        "stage1_freeze": stage1_freeze,
+        "stage2_init_from": str(stage2_init_from) if stage2_init_from is not None else None,
+        "stage2_freeze": stage2_freeze,
         "mode": mode.value if mode is not None else None,
         "stage1_generator": stage1_generator.value if stage1_generator is not None else None,
         "stage2_generator": stage2_generator.value if stage2_generator is not None else None,
@@ -854,6 +907,52 @@ def new_run(
     typer.echo(f"  giant train {data_arg} --config {config_path} --out {run_dir}")
 
 
+model_app = typer.Typer(
+    no_args_is_help=True,
+    help="Inspect a resolved model architecture without training.",
+)
+app.add_typer(model_app, name="model")
+
+
+@model_app.command("summary")
+def model_summary(
+    config: Annotated[
+        Optional[Path],
+        typer.Option("--config", "-c", help="TOML config file (default: built-in defaults)"),
+    ] = None,
+    pdg_vocab: Annotated[
+        int,
+        typer.Option(
+            "--pdg-vocab",
+            help="Placeholder PDG vocab size for conditioning.particle.type='embedding' "
+            "or a pdg/process router (no dataset attached to derive the real training vocab)",
+        ),
+    ] = 300,
+    mat_vocab: Annotated[
+        int,
+        typer.Option(
+            "--mat-vocab",
+            help="Placeholder material vocab size for conditioning.material.type='embedding' "
+            "or a process router (default: the number of known materials in giant.materials)",
+        ),
+    ] = len(MATERIAL_PROPERTIES),
+) -> None:
+    """Build the resolved model graph from a config with no dataset attached, and
+    print per-module parameter counts, trunk widths, which heads exist, and which
+    conditioning/stage1_model/stage2_model config keys actually shaped it."""
+    from giant.model.summary import render_summary, summarize_model
+
+    cfg = gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, config, {})
+    try:
+        gconfig.validate_config(cfg)
+    except ValueError as exc:
+        typer.echo(f"error: {exc}", err=True)
+        raise typer.Exit(1)
+
+    summary = summarize_model(cfg, pdg_vocab=pdg_vocab, mat_vocab=mat_vocab)
+    typer.echo(render_summary(summary))
+
+
 @app.command()
 def predict(
     data: Annotated[Path, typer.Argument(help="Parquet file or directory of parquet files")],

giant/config.py

@@ -468,6 +468,18 @@ class ParticleTypeConfig:
     # silently the same number). 0 = inherit conditioning.particle.emb_dim,
     # preserving pre-#29 behavior.
     n_classes: int = 0
+    # Class-balances the target = "onehot" cross-entropy loss against the
+    # secondary-species long tail (gitea #44: the failure mode motivating the
+    # v0.3.0 pivot was specifically a species collapse — zero photon
+    # secondaries, hallucinated antineutrinos). "none": plain CE (pre-#44
+    # behavior). "inverse_freq": CE weighted by 1/count per class,
+    # normalized to mean 1 across classes so lambda_weight doesn't need
+    # retuning when this is switched on. validate_config requires target =
+    # "onehot" and stage2_model.generator != "wgan" whenever this isn't
+    # "none" — "embedding"/"physical" have no class CE to weight, and the
+    # WGAN stage-2 path feeds its type slice to the critic via a
+    # straight-through Gumbel relaxation instead of a CE loss.
+    class_weighting: str = "none"
 
     @classmethod
     def from_dict(cls, d: dict | None) -> "ParticleTypeConfig":
@@ -477,6 +489,7 @@ class ParticleTypeConfig:
             lambda_weight=d.get("lambda", 1.0),
             other_policy=d.get("other_policy", "sample"),
             n_classes=d.get("n_classes", 0),
+            class_weighting=d.get("class_weighting", "none"),
         )
 
     def to_dict(self) -> dict:
@@ -485,6 +498,7 @@ class ParticleTypeConfig:
             "lambda": self.lambda_weight,
             "other_policy": self.other_policy,
             "n_classes": self.n_classes,
+            "class_weighting": self.class_weighting,
         }
 
 
@@ -594,6 +608,19 @@ class Stage1ModelConfig:
     # false skips building/training stage 1 entirely. The resulting
     # checkpoint holds only stage 2 and cannot be rolled out.
     active: bool = True
+    # Checkpoint .pt to load this stage's weights from before training starts
+    # (its own "model"/"sec_decoder" key, not this run's own resume state) —
+    # "" means start from a fresh init. See `freeze` below for the partial-
+    # retrain use case this exists for (gitea #42).
+    init_from: str = ""
+    # true keeps this stage's weights exactly as loaded from `init_from` —
+    # forward/backward still run every batch (so its loss/grad_norm metrics
+    # stay meaningful, and a WGAN stage's critic still gets a real signal to
+    # report), but its optimizer never steps. Lets a rollout-capable
+    # checkpoint retrain only the *other* stage against a fixed, known-good
+    # one (gitea #42) — `validate_config` requires `init_from` to be set
+    # whenever this is true, unless the run is a `--resume`.
+    freeze: bool = False
     # "flow": conditional flow matching (~10 ODE steps at inference).
     # "ddpm": cosine-schedule diffusion baseline.
     # "wgan": WGAN-GP, single forward pass at inference.
@@ -620,6 +647,8 @@ class Stage1ModelConfig:
         d = d or {}
         return cls(
             active=d.get("active", True),
+            init_from=d.get("init_from", ""),
+            freeze=d.get("freeze", False),
             generator=d.get("generator", "flow"),
             hidden_dim=d.get("hidden_dim", 256),
             n_res_blocks=d.get("n_res_blocks", 6),
@@ -636,6 +665,8 @@ class Stage1ModelConfig:
     def to_dict(self) -> dict:
         return {
             "active": self.active,
+            "init_from": self.init_from,
+            "freeze": self.freeze,
             "generator": self.generator,
             "hidden_dim": self.hidden_dim,
             "n_res_blocks": self.n_res_blocks,
@@ -655,6 +686,9 @@ class Stage2ModelConfig:
     # false trains stage 1 alone. giant rollout must then refuse the
     # checkpoint; giant predict still works.
     active: bool = True
+    # See Stage1ModelConfig.init_from/.freeze — same semantics, this stage.
+    init_from: str = ""
+    freeze: bool = False
     # "one_shot": predict all k_max slots simultaneously with padded slots
     # masked from the loss (v0.2 behaviour).
     # "autoregressive": emit one secondary at a time in descending-energy
@@ -677,6 +711,13 @@ class Stage2ModelConfig:
     # output, closing the train/inference gap at the cost of a sampling pass
     # per batch and a moving target early in training.
     stage1_context: str = "truth"
+    # Ramp for "sampled": P(condition on the ground-truth stage-1 outcome
+    # rather than a fresh sample), linearly interpolated from ctx_p_start
+    # (epoch 0) to ctx_p_end (the final epoch) — the same scheduled-sampling
+    # shape as autoregressive.tf_p_start/tf_p_end, so stage 2 doesn't chase a
+    # wildly moving stage-1 target in early epochs. Unread under "truth".
+    ctx_p_start: float = 1.0
+    ctx_p_end: float = 0.0
     n_sec: NSecConfig = field(default_factory=NSecConfig)
     particle_type: ParticleTypeConfig = field(default_factory=ParticleTypeConfig)
     autoregressive: AutoregressiveConfig = field(default_factory=AutoregressiveConfig)
@@ -692,6 +733,8 @@ class Stage2ModelConfig:
         d = d or {}
         return cls(
             active=d.get("active", True),
+            init_from=d.get("init_from", ""),
+            freeze=d.get("freeze", False),
             decoder=d.get("decoder", "autoregressive"),
             generator=d.get("generator", "wgan"),
             hidden_dim=d.get("hidden_dim", 256),
@@ -701,6 +744,8 @@ class Stage2ModelConfig:
             k_max=d.get("k_max", 15),
             context_dim=d.get("context_dim", 64),
             stage1_context=d.get("stage1_context", "truth"),
+            ctx_p_start=d.get("ctx_p_start", 1.0),
+            ctx_p_end=d.get("ctx_p_end", 0.0),
             n_sec=NSecConfig.from_dict(d.get("n_sec")),
             particle_type=ParticleTypeConfig.from_dict(d.get("particle_type")),
             autoregressive=AutoregressiveConfig.from_dict(d.get("autoregressive")),
@@ -715,6 +760,8 @@ class Stage2ModelConfig:
     def to_dict(self) -> dict:
         return {
             "active": self.active,
+            "init_from": self.init_from,
+            "freeze": self.freeze,
             "decoder": self.decoder,
             "generator": self.generator,
             "hidden_dim": self.hidden_dim,
@@ -724,6 +771,8 @@ class Stage2ModelConfig:
             "k_max": self.k_max,
             "context_dim": self.context_dim,
             "stage1_context": self.stage1_context,
+            "ctx_p_start": self.ctx_p_start,
+            "ctx_p_end": self.ctx_p_end,
             "n_sec": self.n_sec.to_dict(),
             "particle_type": self.particle_type.to_dict(),
             "autoregressive": self.autoregressive.to_dict(),
@@ -766,6 +815,12 @@ class TrainConfig:
     # thousands of steps. Per-epoch metrics (the metrics.csv row) always log
     # in full.
     wandb_log_every: int = 50
+    # Training-step autocast dtype: "fp32" (default, no autocast) or "bf16".
+    # No "fp16" — GradScaler and the double-backward in
+    # giant.model.wgan.gradient_penalty don't mix well, and bf16 alone covers
+    # every training GPU in the fleet (Ampere and newer). See
+    # giant.training.amp.resolve_autocast (gitea #47).
+    precision: str = "fp32"
 
     @classmethod
     def from_dict(cls, d: dict | None) -> "TrainConfig":
@@ -787,6 +842,7 @@ class TrainConfig:
             wandb_project=d.get("wandb_project", "giant"),
             wandb_run_name=d.get("wandb_run_name", ""),
             wandb_log_every=d.get("wandb_log_every", 50),
+            precision=d.get("precision", "fp32"),
         )
 
     def to_dict(self) -> dict:
@@ -807,6 +863,7 @@ class TrainConfig:
             "wandb_project": self.wandb_project,
             "wandb_run_name": self.wandb_run_name,
             "wandb_log_every": self.wandb_log_every,
+            "precision": self.precision,
         }
 
 
@@ -843,6 +900,25 @@ class GiantConfig:
 DEFAULT_CONFIG: dict = GiantConfig().to_dict()
 
 
+def leaf_paths(node: dict, prefix: str = "") -> list[str]:
+    """Every dotted leaf path in a DEFAULT_CONFIG-shaped dict, e.g.
+    "stage1_model.router.n_experts". `[meta]` (run provenance, no schema
+    counterpart) is skipped at the top level, matching `validate_config_keys`.
+    Shared by `tests/test_config_consumed_keys.py` (the static per-identifier
+    audit) and `giant.model.summary` (the runtime per-config audit, gitea
+    #46) so both walk the exact same tree."""
+    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 git_hash() -> str:
     try:
         return subprocess.check_output(["git", "rev-parse", "HEAD"], stderr=subprocess.DEVNULL).decode().strip()
@@ -1060,6 +1136,7 @@ FLAG_SPECS: tuple[FlagSpec, ...] = (
     FlagSpec("wandb_project", ("train.wandb_project",)),
     FlagSpec("wandb_run_name", ("train.wandb_run_name",)),
     FlagSpec("wandb_log_every", ("train.wandb_log_every",)),
+    FlagSpec("precision", ("train.precision",)),
     # --hidden-dim/--n-blocks/--dropout are stage-1-only backward-compat
     # shorthands (they predate stage2_model having its own flags);
     # --stage1-* wins when both are given.
@@ -1112,6 +1189,13 @@ FLAG_SPECS: tuple[FlagSpec, ...] = (
     FlagSpec("stage1_critic_n_res_blocks", ("stage1_model.wgan.critic_n_res_blocks",)),
     FlagSpec("stage2_critic_hidden_dim", ("stage2_model.wgan.critic_hidden_dim",)),
     FlagSpec("stage2_critic_n_res_blocks", ("stage2_model.wgan.critic_n_res_blocks",)),
+    # Partial-retrain (gitea #42): stage-scoped only, no shared alias — a
+    # shared "freeze both stages from the same file" flag has no sensible
+    # meaning (a checkpoint has one set of weights per stage).
+    FlagSpec("stage1_init_from", ("stage1_model.init_from",)),
+    FlagSpec("stage1_freeze", ("stage1_model.freeze",)),
+    FlagSpec("stage2_init_from", ("stage2_model.init_from",)),
+    FlagSpec("stage2_freeze", ("stage2_model.freeze",)),
 )
 
 
@@ -1344,7 +1428,7 @@ def merge_cli_overrides(
     return cfg
 
 
-def validate_config(cfg: dict) -> None:
+def validate_config(cfg: dict, *, resume: bool = False) -> None:
     """Cross-block validation the per-block schema can't express on its own.
 
     Raises ValueError with a clear message on the first violation found. Call
@@ -1352,6 +1436,10 @@ def validate_config(cfg: dict) -> None:
     these checks need to see across blocks, so they don't belong in
     `migrate_config` (which only ever sees one dict's own keys) or in any
     single block's defaults.
+
+    `resume=True` (only `giant train --resume` passes this) relaxes the
+    `stage{1,2}_model.freeze` -> `.init_from` requirement below: a resumed
+    frozen stage's weights come from the resume checkpoint, not `init_from`.
     """
     particle_type = _get_path(cfg, "conditioning.particle.type")
 
@@ -1364,7 +1452,34 @@ def validate_config(cfg: dict) -> None:
             f"{particle_type!r})"
         )
 
+    class_weighting = _get_path(cfg, "stage2_model.particle_type.class_weighting")
+    if class_weighting not in ("none", "inverse_freq"):
+        raise ValueError(
+            f"stage2_model.particle_type.class_weighting = {class_weighting!r} — must be 'none' or 'inverse_freq'"
+        )
+    if class_weighting != "none" and pt_target != "onehot":
+        raise ValueError(
+            "stage2_model.particle_type.class_weighting != 'none' requires "
+            f"stage2_model.particle_type.target = 'onehot' (there is no class "
+            f"cross-entropy to weight under target = {pt_target!r})"
+        )
+    if class_weighting != "none" and _get_path(cfg, "stage2_model.generator") == "wgan":
+        raise ValueError(
+            "stage2_model.particle_type.class_weighting != 'none' is "
+            "incompatible with stage2_model.generator = 'wgan' — that path "
+            "feeds the type slice to the critic via a straight-through "
+            "Gumbel relaxation instead of a class cross-entropy, so there is "
+            "nothing to weight"
+        )
+
     for stage_name in ("stage1_model", "stage2_model"):
+        if _get_path(cfg, f"{stage_name}.freeze") and not _get_path(cfg, f"{stage_name}.init_from") and not resume:
+            raise ValueError(
+                f"{stage_name}.freeze = true requires {stage_name}.init_from "
+                "to be set (or --resume) — freezing a randomly-initialized "
+                "model is almost certainly a mistake"
+            )
+
         router = _get_path(cfg, f"{stage_name}.router") or {}
         if router.get("enabled") and router.get("type") in ("pdg", "process") and particle_type == "physical":
             raise ValueError(
@@ -1401,14 +1516,36 @@ def validate_config(cfg: dict) -> None:
     if stop_sampling not in ("greedy", "sample"):
         raise ValueError(f"stage2_model.n_sec.stop_sampling = {stop_sampling!r} — must be 'greedy' or 'sample'")
 
-    if _get_path(cfg, "stage2_model.stage1_context") == "sampled":
+    precision = _get_path(cfg, "train.precision")
+    if precision not in ("fp32", "bf16"):
         raise ValueError(
-            "stage2_model.stage1_context = 'sampled' is accepted by the schema "
-            "but not implemented — trainers.py always trains stage 2 against "
-            "the ground-truth stage-1 output; use 'truth' (default) instead "
-            "(see issues.md Issue 16 for the planned implementation)"
+            f"train.precision = {precision!r} — must be 'fp32' or 'bf16' "
+            "('fp16' is not supported: see giant.training.amp.resolve_autocast)"
         )
 
+    stage1_context = _get_path(cfg, "stage2_model.stage1_context")
+    if stage1_context not in ("truth", "sampled"):
+        raise ValueError(f"stage2_model.stage1_context = {stage1_context!r} — must be 'truth' or 'sampled'")
+    if stage1_context == "sampled":
+        if not (_get_path(cfg, "stage1_model.active") and _get_path(cfg, "stage2_model.active")):
+            raise ValueError(
+                "stage2_model.stage1_context = 'sampled' requires both "
+                "stage1_model.active and stage2_model.active = true — there is "
+                "no stage-1 model to sample from in a stage-2-only run"
+            )
+        ctx_p_start = _get_path(cfg, "stage2_model.ctx_p_start")
+        ctx_p_end = _get_path(cfg, "stage2_model.ctx_p_end")
+        for name, value in (("ctx_p_start", ctx_p_start), ("ctx_p_end", ctx_p_end)):
+            if not (0.0 <= value <= 1.0):
+                raise ValueError(f"stage2_model.{name} = {value} — must be in [0, 1]")
+        if ctx_p_start == 1.0 and ctx_p_end == 1.0:
+            raise ValueError(
+                "stage2_model.stage1_context = 'sampled' with ctx_p_start = "
+                "ctx_p_end = 1.0 always conditions on the ground truth — "
+                "identical to 'truth' but silently so; use 'truth' instead or "
+                "lower ctx_p_end"
+            )
+
     if (
         _get_path(cfg, "stage2_model.n_sec.mode") == "truth"
         and _get_path(cfg, "stage1_model.active")
@@ -1569,6 +1706,8 @@ _OUT_DIR_NAME_CANDIDATES = [
     ),
     ("particle_conditioning", _conditioning_candidate("particle", "c")),
     ("material_conditioning", _conditioning_candidate("material", "m")),
+    ("stage1_freeze", _path_candidate("stage1_model.freeze", "s1frozen", formatter=lambda _: "")),
+    ("stage2_freeze", _path_candidate("stage2_model.freeze", "s2frozen", formatter=lambda _: "")),
     ("stage1_hidden_dim", _path_candidate("stage1_model.hidden_dim", "h")),
     ("stage2_hidden_dim", _path_candidate("stage2_model.hidden_dim", "s2h")),
     ("stage1_n_res_blocks", _path_candidate("stage1_model.n_res_blocks", "b")),

giant/data/loader.py

@@ -1,4 +1,4 @@
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Iterator
 
@@ -256,24 +256,32 @@ def _rank_by_frequency_from_files(files: list[Path], column: str, cast) -> dict:
     return counts
 
 
-def _topn_plus_other_map(counts: dict, n_classes: int) -> tuple[dict, dict]:
+def _topn_plus_other_map(counts: dict, n_classes: int) -> tuple[dict, 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).
+    Returns `(class_map, other_members, class_counts)` — `other_members` is
+    `{key: count}` for every key bucketed into "other" (the empirical
+    within-bucket distribution, for `other_policy = "sample"` at rollout);
+    `class_counts` is `{index: total_count}` for every resulting class index
+    (0-indexed; the "other" index's count is the sum of `other_members`),
+    the per-class frequencies `stage2_model.particle_type.class_weighting`
+    (gitea #44) needs and that would otherwise be dropped once `counts` is
+    collapsed into `class_map`.
     """
     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)}
+    class_counts = {i: counts[k] 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
+    if other_members:
+        class_counts[other_idx] = sum(other_members.values())
+    return class_map, other_members, class_counts
 
 
 def build_process_map_from_files(files: list[Path], n_experts: int) -> dict[str, int]:
@@ -287,7 +295,7 @@ def build_process_map_from_files(files: list[Path], n_experts: int) -> dict[str,
     fixed-width n_sec_head classifier.
     """
     counts = _rank_by_frequency_from_files(files, "process", str)
-    class_map, _ = _topn_plus_other_map(counts, n_experts)
+    class_map, _, _ = _topn_plus_other_map(counts, n_experts)
     return class_map
 
 
@@ -299,6 +307,11 @@ class TopNMap:
 
     class_map: dict
     other_members: dict
+    # {class_index: total_count} — see _topn_plus_other_map. Empty for a
+    # TopNMap decoded from a checkpoint/sidecar predating gitea #44; only
+    # stage2_model.particle_type.class_weighting reads it, and it raises
+    # loudly if it needs counts that aren't there (giant/training/trainers.py).
+    class_counts: dict = field(default_factory=dict)
 
 
 def build_topn_map_from_files(files: list[Path], column: str, n_classes: int, cast=str) -> TopNMap:
@@ -315,8 +328,8 @@ def build_topn_map_from_files(files: list[Path], column: str, n_classes: int, ca
     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)
+    class_map, other_members, class_counts = _topn_plus_other_map(counts, n_classes)
+    return TopNMap(class_map=class_map, other_members=other_members, class_counts=class_counts)
 
 
 def build_pdg_topn_map_from_files(files: list[Path], n_classes: int) -> TopNMap:
@@ -347,5 +360,5 @@ def build_pdg_topn_map_from_files(files: list[Path], n_classes: int) -> TopNMap:
         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)
+    class_map, other_members, class_counts = _topn_plus_other_map(counts, n_classes)
+    return TopNMap(class_map=class_map, other_members=other_members, class_counts=class_counts)

giant/data/setup_cache.py

@@ -35,7 +35,10 @@ from giant.data.transforms import Normalizer, sorted_membership
 # v3: NormalizerEntry.energy_reservoir_sample (100k raw values) replaced by
 # energy_quantiles (a fixed ENERGY_QUANTILE_LEVELS-point quantile grid) — a
 # v2 sidecar has no such grid to fall back on, so it must be recomputed.
-_CACHE_FORMAT_VERSION = 3
+# v4: TopNMap gained class_counts (gitea #44, stage2_model.particle_type.
+# class_weighting) — a v3 sidecar's cached topn_maps have no counts, so they
+# must be rebuilt rather than silently cached with class_counts={}.
+_CACHE_FORMAT_VERSION = 4
 
 _DIMS = {
     "COND_DIM": COND_DIM,
@@ -131,6 +134,7 @@ 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()},
+        "class_counts": {str(k): v for k, v in m.class_counts.items()},
     }
 
 
@@ -139,6 +143,11 @@ def topnmap_from_json(d: dict, axis: str) -> TopNMap:
     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()},
+        # Missing for a checkpoint's topn maps predating gitea #44 — {} is
+        # the correct decode there (inference never reads class_counts; only
+        # stage2_model.particle_type.class_weighting does, at train time, and
+        # it raises loudly if it needs counts a checkpoint doesn't have).
+        class_counts={int(k): v for k, v in d.get("class_counts", {}).items()},
     )
 
 

giant/model/builders.py

@@ -202,6 +202,8 @@ def build_critics(model_config: dict) -> dict[str, nn.Module | None]:
             cond_out_dim=cond_out_dim,
             dropout=s1_spec.dropout,
             stage="stage1",
+            trunk_type=s1_spec.trunk.type,
+            block_conditioning=s1_spec.trunk.block_conditioning,
         )
 
     if s2_spec.active and build_objective(s2_spec.generator).is_adversarial:
@@ -225,6 +227,8 @@ def build_critics(model_config: dict) -> dict[str, nn.Module | None]:
             dropout=s2_spec.dropout,
             stage="stage2",
             context_dim=s2_spec.context_dim,
+            trunk_type=s2_spec.trunk.type,
+            block_conditioning=s2_spec.trunk.block_conditioning,
         )
 
     return result

giant/model/history.py

@@ -63,6 +63,23 @@ def build_history(name: str, in_dim: int, out_dim: int, **kwargs) -> HistoryEnco
     return cls(in_dim, out_dim, **filtered)
 
 
+@register_history("none")
+class NoHistory(HistoryEncoder):
+    """No history signal at all — ignores feat/has_prev entirely and always
+    returns zeros. Ablates whether the AR decoder's history conditioning is
+    earning its parameters. `init_cache`/`step` use the base class's O(1)
+    defaults unmodified (this encoder's own `forward` is already O(1) per
+    call regardless of prefix length)."""
+
+    def __init__(self, in_dim: int, out_dim: int) -> None:
+        super().__init__()
+        self.out_dim = out_dim
+
+    def forward(self, feat: torch.Tensor, has_prev: torch.Tensor) -> torch.Tensor:
+        B, K, _ = feat.shape
+        return torch.zeros(B, K, self.out_dim, device=feat.device, dtype=feat.dtype)
+
+
 @register_history("markov")
 class MarkovHistory(HistoryEncoder):
     """Summarizes the previous secondary's own `(energy_fraction, direction,

giant/model/models.py

@@ -8,7 +8,7 @@ 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.layers import ContextAdapter, SinusoidalEmbedding, build_mlp_head
 from giant.model.objectives import build_objective
 from giant.model.routers import Router
 from giant.model.trunks import build_trunk
@@ -188,6 +188,26 @@ class StageModel(nn.Module):
             hidden = max(1, round(hidden_dim * head_cfg.hidden_ratio))
             self.stop_head = build_mlp_head(cond_out_dim, 1, hidden, head_cfg.depth)
 
+    def _build_context_fusion(self, x_dim: int, context_dim: int, cond_out_dim: int) -> None:
+        """Builds `self.context_adapter`/`self.fuse` — the stage-2-style
+        context-fusion pattern (project the previous stage's outcome down to
+        `context_dim` via `ContextAdapter`, concat onto the base conditioning,
+        project back to `cond_out_dim`) shared by `Stage2OneShot` and a
+        `stage="stage2"` `CriticModel` (gitea #57). Call from a subclass's
+        `__init__` before using `_cond_embed`."""
+        self.context_adapter = ContextAdapter(x_dim, context_dim)
+        self.fuse = nn.Sequential(
+            nn.Linear(cond_out_dim + context_dim, cond_out_dim),
+            nn.SiLU(),
+        )
+
+    def _cond_embed(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor, stage1_out: torch.Tensor) -> torch.Tensor:
+        """Fuses base conditioning with the previous stage's outcome — pairs
+        with `_build_context_fusion`."""
+        base = self.cond_enc(cond_cont, cond_cat)
+        ctx = self.context_adapter(stage1_out)
+        return self.fuse(torch.cat([base, ctx], dim=-1))
+
     def _require_n_sec_head(self) -> None:
         if self.n_sec_head is None:
             raise RuntimeError(
@@ -363,11 +383,7 @@ class Stage2OneShot(StageModel):
             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(),
-        )
+        self._build_context_fusion(x_dim, context_dim, cond_out_dim)
         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(
@@ -386,11 +402,6 @@ class Stage2OneShot(StageModel):
             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,
@@ -702,11 +713,25 @@ class Stage2Autoregressive(StageModel):
         return self.stop_head(c_emb.reshape(B * K, -1)).view(B, K)
 
 
-class CriticModel(nn.Module):
+class CriticModel(StageModel):
     """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"`."""
+    `Stage2OneShot`, via `StageModel._build_context_fusion`/`_cond_embed`).
+    Used only when that stage's `generator == "wgan"`.
+
+    Subclasses `StageModel` for the `cond_enc` construction and (stage 2)
+    context-fusion scaffolding only — its trunk is built directly via
+    `build_trunk` (output width 1) rather than through
+    `_build_trunk_and_heads`, since that helper is shaped around a
+    generator's `Objective`/time-embedding/flow-matching concerns
+    (`forward`'s `(x_t, cond) -> vector` shape) that don't apply to a critic's
+    `(x, cond) -> scalar` (gitea #57). `generator="wgan"` is passed to the
+    base purely because that's factually when a critic exists; nothing here
+    ever calls `_build_trunk_and_heads`, so no head/time-embedding machinery
+    is built from it. Never routed (MoE) — that's a separate, unrequested
+    axis of scope; see gitea #57's proposal, which covers only the trunk/
+    block registries."""
 
     def __init__(
         self,
@@ -722,22 +747,26 @@ class CriticModel(nn.Module):
         stage: str = "stage1",
         context_dim: int = 64,
         context_in_dim: int = X_DIM,
+        trunk_type: str = "resmlp",
+        block_conditioning: str = "add",
     ) -> None:
-        super().__init__()
+        super().__init__(
+            pdg_vocab,
+            mat_vocab,
+            particle_cfg,
+            material_cfg,
+            cond_out_dim=cond_out_dim,
+            generator="wgan",
+            noise_dim=0,
+        )
         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)
+            self._build_context_fusion(context_in_dim, context_dim, cond_out_dim)
+        self.trunk = build_trunk(
+            None, trunk_type, in_dim, 1, hidden_dim, n_res_blocks, cond_out_dim, dropout, block_conditioning
+        )
 
     def forward(
         self,
@@ -746,13 +775,9 @@ class CriticModel(nn.Module):
         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))
+            assert stage1_out is not None, "stage='stage2' CriticModel requires stage1_out"
+            cond = self._cond_embed(cond_cont, cond_cat, stage1_out)
         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)
+            cond = self.cond_enc(cond_cont, cond_cat)
+        return self.trunk(x, cond, cond_cont, cond_cat).squeeze(-1)

giant/model/network.py

@@ -15,6 +15,7 @@ from giant.model.history import (
     AttentionHistory,
     HistoryEncoder,
     MarkovHistory,
+    NoHistory,
     _CausalAttnBlock,
     build_history,
     register_history,
@@ -54,6 +55,7 @@ from giant.model.routers import (
     ROUTER_REGISTRY,
     ComposedRouter,
     EnergyRouter,
+    NoneRouter,
     PdgRouter,
     ProcessRouter,
     Router,
@@ -67,6 +69,7 @@ from giant.model.routers import (
 from giant.model.trunks import (
     TRUNK_REGISTRY,
     ExpertTrunk,
+    LinearTrunk,
     RoutedTrunk,
     Trunk,
     _route_forward,
@@ -90,7 +93,10 @@ __all__ = [
     "FlowObjective",
     "HISTORY_REGISTRY",
     "HistoryEncoder",
+    "LinearTrunk",
     "MarkovHistory",
+    "NoHistory",
+    "NoneRouter",
     "OBJECTIVE_REGISTRY",
     "Objective",
     "PdgRouter",

giant/model/routers.py

@@ -45,21 +45,36 @@ class Router(nn.Module):
         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.
+
+        Forced fp32 (`torch.autocast(..., enabled=False)`) regardless of the
+        caller's ambient `train.precision` autocast region: `clamp_min(1e-8)`
+        below sits under bf16's precision but *above* fp16's ~6e-8 subnormal
+        floor, so `log_probs` degrading here is exactly the kind of quiet
+        drift that cost a whole rollout benchmark before (see the MoE section
+        of CLAUDE.md's Roadmap) — cheap to rule out (gitea #47).
         """
-        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)
+        with torch.autocast(cond_cont.device.type, enabled=False):
+            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)
+        with torch.autocast(cond_cont.device.type, enabled=False):
+            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
+        """Importance CV^2 load-balancing auxiliary loss (Shazeer et al. 2017).
+
+        Forced fp32 — `importance` sums `gate()` over the whole batch (a
+        large-magnitude accumulation in reduced precision), then takes a
+        `std/mean` ratio: a classic catastrophic-cancellation shape (gitea
+        #47)."""
+        with torch.autocast(cond_cont.device.type, enabled=False):
+            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.
@@ -76,12 +91,17 @@ class Router(nn.Module):
 
     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
+        the full explanation, unchanged in v0.3.0.
+
+        Forced fp32, same rationale as `balance_loss`/`combine_weights`: the
+        `+ 1e-8` epsilon here is `entropy_loss`'s training-loss path too, not
+        just a diagnostic (gitea #47)."""
+        with torch.autocast(cond_cont.device.type, enabled=False):
+            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]] = {}
@@ -124,6 +144,24 @@ def _inverse_bounded_interp(value: float, lo: float, hi: float) -> float:
     return math.log(p / (1 - p))
 
 
+@register_router("none")
+class NoneRouter(Router):
+    """Uniform 1/n_experts gate — no learned routing signal at all.
+
+    Still builds n_experts expert trunks via RoutedTrunk (same parameter
+    budget as a real router), but every row gets an identical weight
+    regardless of conditioning. Ablates whether the *learned routing
+    signal* — as opposed to simply having multiple experts — is earning
+    its parameters. `top1()` (the base class default) always dispatches to
+    expert 0 (argmax of a uniform vector), which still exercises
+    RoutedTrunk's real per-expert grouped-dispatch code path at eval time.
+    """
+
+    def gate(self, cond_cont: torch.Tensor, cond_cat: torch.Tensor) -> torch.Tensor:
+        B = cond_cont.shape[0]
+        return torch.full((B, self.n_experts), 1.0 / self.n_experts, device=cond_cont.device)
+
+
 @register_router("energy")
 class EnergyRouter(Router):
     """Soft turn-on gate over normalized pre-step log-energy.

giant/model/summary.py

@@ -0,0 +1,318 @@
+"""Build-only model introspection (gitea #46): construct the resolved
+Stage1/Stage2/critic graph from a config with no dataset attached, and report
+per-module parameter counts, trunk widths, which heads exist, and — via
+differential probing — which `conditioning`/`stage1_model`/`stage2_model`
+config keys actually shape the built model. This is the runtime counterpart
+to `tests/test_config_consumed_keys.py`'s static per-identifier audit: that
+test asks "does any code reference this key's name at all", this module asks
+"given *this* resolved config, does the key change what `build_models`/
+`build_critics` (`giant/model/builders.py`) actually produces".
+
+Differential probing, not identifier matching: build the model once from the
+resolved config and take a structural fingerprint (`_fingerprint` — which
+submodules exist, every parameter's/buffer's shape+dtype, every plain scalar
+attribute stored on any module). Then, for each in-scope leaf key, perturb
+just that one value (`_perturb`), rebuild, and re-fingerprint. A changed
+fingerprint — or a rebuild that raises — means the key was consumed; an
+identical fingerprint means construction never looked at it under this
+particular config. A key can be genuinely inert under one config and live
+under another (e.g. any `stage1_model.router.*` key when `router.enabled =
+false`) — that config-dependence is exactly the "silently degenerate
+combination" issue #46 is after, so it is reported per-run rather than
+baked into a static table.
+
+Keys legitimately owned by the trainer/sampler/rollout rather than by
+`build_models`/`build_critics` (loss weights, WGAN-GP training
+hyperparameters, teacher-forcing and stage1-context schedules, ...) are
+cataloged in `_NOT_BUILD_TIME` below so the report doesn't flag them as
+suspicious. One leaf is inert under every config today —
+`stage2_model.autoregressive.order` — matching
+`tests/test_config_consumed_keys.py`'s own `_KNOWN_UNUSED` entry; it is
+deliberately *not* in `_NOT_BUILD_TIME`, since "always inert" is itself the
+finding those two tests independently converge on.
+"""
+
+import copy
+from dataclasses import dataclass, field
+
+import torch.nn as nn
+
+from giant.config import _get_path, _set_path, leaf_paths
+from giant.model.builders import build_critics, build_models
+from giant.model.trunks import RoutedTrunk
+
+_IN_SCOPE_ROOTS = ("conditioning", "stage1_model", "stage2_model")
+
+_PROBE_STR = "__giant_model_summary_probe__"
+
+# A handful of string leaves branch on equality against one specific literal
+# (e.g. `builders.py`: `stop_token = s2_spec.n_sec.mode == "stop_token"`),
+# where every value other than that literal behaves identically. A single
+# generic sentinel probe would then falsely read as inert whenever the
+# config's *current* value is already one of those identically-behaving
+# "other" values (e.g. mode="head") — it never crosses the one boundary that
+# actually matters. Named here so probing tries the real alternative(s) too;
+# every other string leaf is registry-validated (raises on garbage, still
+# correctly detected as consumed) or genuinely value-independent, so doesn't
+# need an entry.
+_STRING_ALTERNATIVES: dict[str, tuple[str, ...]] = {
+    "stage2_model.n_sec.owner": ("stage1", "stage2"),
+    "stage2_model.n_sec.mode": ("stop_token", "head", "truth"),
+    "stage2_model.particle_type.target": ("physical", "onehot", "embedding"),
+}
+
+# Verified by reading giant/training/trainers.py, giant/training/stage2_inputs.py
+# and giant/rollout.py while implementing gitea #46 — not auto-derived, so a
+# future reader touching these fields should re-check this table still holds.
+_NOT_BUILD_TIME: dict[str, str] = {
+    "stage1_model.init_from": "training/checkpoint.py's init_stages_from_checkpoints, run before build_stage_trainers (gitea #42)",
+    "stage1_model.freeze": "trainers.py: StageSpec.freeze, gates StageTrainer._step_optimizer (gitea #42)",
+    "stage2_model.init_from": "training/checkpoint.py's init_stages_from_checkpoints, run before build_stage_trainers (gitea #42)",
+    "stage2_model.freeze": "trainers.py: StageSpec.freeze, gates StageTrainer._step_optimizer (gitea #42)",
+    "stage1_model.lambda": "trainers.py: StageSpec.lambda_weight, the total-loss mix weight",
+    "stage2_model.lambda": "trainers.py: StageSpec.lambda_weight, the total-loss mix weight",
+    "stage2_model.n_sec.lambda": "trainers.py: StageSpec.n_sec_lambda, the n_sec-head loss weight",
+    "stage2_model.particle_type.lambda": "trainers.py: Stage2Trainer.particle_type_lambda, the type-head loss weight",
+    "stage2_model.particle_type.other_policy": "giant/rollout.py: resolves an 'other'-bucket secondary's PDG code at inference",
+    "stage2_model.particle_type.class_weighting": "trainers.py: FlowDDPMStageTrainer.type_class_weights, shapes the type-head loss, not the built graph (gitea #44)",
+    "stage2_model.autoregressive.teacher_forcing": "giant/training/stage2_inputs.py's training-time input assembly",
+    "stage2_model.autoregressive.tf_p_start": "trainers.py's teacher-forcing schedule",
+    "stage2_model.autoregressive.tf_p_end": "trainers.py's teacher-forcing schedule",
+    "stage2_model.stage1_context": "trainers.py's stage1/stage2 boundary — StageTrainer._stage1_context",
+    "stage2_model.ctx_p_start": "trainers.py's stage1-context sampling schedule",
+    "stage2_model.ctx_p_end": "trainers.py's stage1-context sampling schedule",
+    "stage1_model.router.lambda_balance": "trainers.py's load-balancing auxiliary loss weight",
+    "stage1_model.router.lambda_entropy": "trainers.py's entropy-regularization auxiliary loss weight",
+    "stage1_model.router.lambda_proc": "trainers.py's supervised process-classification auxiliary loss weight",
+    "stage1_model.router.gumbel_tau_start": "trainers.py's expert-combination Gumbel-softmax temperature anneal",
+    "stage1_model.router.gumbel_tau_end": "trainers.py's expert-combination Gumbel-softmax temperature anneal",
+    "stage2_model.router.lambda_balance": "trainers.py's load-balancing auxiliary loss weight",
+    "stage2_model.router.lambda_entropy": "trainers.py's entropy-regularization auxiliary loss weight",
+    "stage2_model.router.lambda_proc": "trainers.py's supervised process-classification auxiliary loss weight",
+    "stage2_model.router.gumbel_tau_start": "trainers.py's expert-combination Gumbel-softmax temperature anneal",
+    "stage2_model.router.gumbel_tau_end": "trainers.py's expert-combination Gumbel-softmax temperature anneal",
+    "stage1_model.wgan.n_critic": "trainers.py's WGAN-GP critic-update cadence",
+    "stage1_model.wgan.gp_weight": "trainers.py's WGAN-GP gradient-penalty coefficient",
+    "stage1_model.wgan.critic_lr": "trainers.py's critic optimizer learning rate",
+    "stage2_model.wgan.n_critic": "trainers.py's WGAN-GP critic-update cadence",
+    "stage2_model.wgan.gp_weight": "trainers.py's WGAN-GP gradient-penalty coefficient",
+    "stage2_model.wgan.critic_lr": "trainers.py's critic optimizer learning rate",
+    "stage2_model.wgan.gumbel_tau_start": "trainers.py's type-slice Gumbel-softmax temperature anneal (type_gumbel_tau_start)",
+    "stage2_model.wgan.gumbel_tau_end": "trainers.py's type-slice Gumbel-softmax temperature anneal (type_gumbel_tau_end)",
+}
+
+
+@dataclass
+class ModelSummary:
+    modules: dict[str, nn.Module]
+    consumed: list[str]
+    inert: list[str]
+    elsewhere: list[str]
+    pdg_vocab: int
+    mat_vocab: int
+    vocab_caveats: list[str] = field(default_factory=list)
+
+
+def _build_model_config(cfg: dict, pdg_vocab: int, mat_vocab: int) -> dict:
+    return {
+        "pdg_vocab": pdg_vocab,
+        "mat_vocab": mat_vocab,
+        "conditioning": cfg["conditioning"],
+        "stage1_model": cfg["stage1_model"],
+        "stage2_model": cfg["stage2_model"],
+    }
+
+
+def _built_modules(cfg: dict, pdg_vocab: int, mat_vocab: int) -> dict[str, nn.Module]:
+    model_config = _build_model_config(cfg, pdg_vocab, mat_vocab)
+    modules: dict[str, nn.Module] = {}
+    for name, m in build_models(model_config).items():
+        if m is not None:
+            modules[name] = m
+    for name, m in build_critics(model_config).items():
+        if m is not None:
+            modules[f"{name}_critic"] = m
+    return modules
+
+
+def _fingerprint(modules: dict[str, nn.Module]) -> list:
+    """A config-shape fingerprint of the built graph: which submodules
+    exist, every parameter's/buffer's shape+dtype (never values — those are
+    randomly initialized and irrelevant to *structure*), and every plain
+    scalar attribute any module stores on itself (e.g. `Stage2Autoregressive
+    .stop_sampling`, `EnergyRouter.temperature`) — this is what makes a
+    non-parametric key's effect on construction observable."""
+    sig = []
+    for stage_name, module in modules.items():
+        for mod_name, m in module.named_modules():
+            full = f"{stage_name}.{mod_name}" if mod_name else stage_name
+            for k, v in vars(m).items():
+                if k.startswith("_"):
+                    continue
+                if v is None or isinstance(v, (bool, int, float, str)):
+                    sig.append((full, k, v))
+        for pname, p in module.named_parameters():
+            sig.append((stage_name, "param", pname, tuple(p.shape), str(p.dtype)))
+        for bname, b in module.named_buffers():
+            sig.append((stage_name, "buffer", bname, tuple(b.shape), str(b.dtype)))
+    return sorted(sig, key=repr)
+
+
+def _perturb_candidates(path: str, value) -> list:
+    """Values to try perturbing `path`'s current `value` to, in order —
+    probing stops at the first one that changes the fingerprint or raises.
+    Almost always a single candidate; see `_STRING_ALTERNATIVES`."""
+    if isinstance(value, bool):
+        return [not value]
+    if isinstance(value, int):
+        return [value + 1]
+    if isinstance(value, float):
+        return [value + 1.0]
+    if isinstance(value, str):
+        alternatives = [v for v in _STRING_ALTERNATIVES.get(path, ()) if v != value]
+        return [*alternatives, _PROBE_STR]
+    raise TypeError(f"gitea #46 probing: unsupported leaf value type {type(value)!r} ({value!r})")
+
+
+def _vocab_caveats(cfg: dict) -> list[str]:
+    caveats = []
+    if _get_path(cfg, "conditioning.particle.type") == "embedding":
+        caveats.append(
+            "conditioning.particle.type = 'embedding' -- pdg_vocab below is a "
+            "placeholder (no dataset attached to derive the real training vocab size)"
+        )
+    if _get_path(cfg, "conditioning.material.type") == "embedding":
+        caveats.append(
+            "conditioning.material.type = 'embedding' -- mat_vocab below is a "
+            "placeholder (no dataset attached to derive the real training vocab size)"
+        )
+    for stage in ("stage1_model", "stage2_model"):
+        router_type = _get_path(cfg, f"{stage}.router.type")
+        if _get_path(cfg, f"{stage}.router.enabled") and router_type in ("pdg", "process"):
+            caveats.append(
+                f"{stage}.router.type = {router_type!r} builds its own pdg_vocab-sized "
+                "embedding -- the count above is a placeholder"
+            )
+    return caveats
+
+
+def summarize_model(cfg: dict, pdg_vocab: int, mat_vocab: int) -> ModelSummary:
+    """Build `cfg`'s model with no dataset attached and report its resolved
+    graph, plus which `conditioning`/`stage1_model`/`stage2_model` config
+    keys actually shaped it (differential probing — see module docstring).
+    `cfg` must already be a fully-merged v0.3 config (`merge_cli_overrides`
+    output) — this does not migrate or validate it."""
+    modules = _built_modules(cfg, pdg_vocab, mat_vocab)
+    baseline_fp = _fingerprint(modules)
+
+    in_scope = [p for p in leaf_paths(cfg) if p.split(".", 1)[0] in _IN_SCOPE_ROOTS]
+    consumed: list[str] = []
+    inert: list[str] = []
+    elsewhere: list[str] = []
+    for path in in_scope:
+        original = _get_path(cfg, path)
+        changed = False
+        for candidate in _perturb_candidates(path, original):
+            probe_cfg = copy.deepcopy(
+                {
+                    "conditioning": cfg["conditioning"],
+                    "stage1_model": cfg["stage1_model"],
+                    "stage2_model": cfg["stage2_model"],
+                }
+            )
+            _set_path(probe_cfg, path, candidate)
+            try:
+                changed = _fingerprint(_built_modules(probe_cfg, pdg_vocab, mat_vocab)) != baseline_fp
+            except Exception:
+                changed = True
+            if changed:
+                break
+        if changed:
+            consumed.append(path)
+        elif path in _NOT_BUILD_TIME:
+            elsewhere.append(path)
+        else:
+            inert.append(path)
+
+    return ModelSummary(
+        modules=modules,
+        consumed=sorted(consumed),
+        inert=sorted(inert),
+        elsewhere=sorted(elsewhere),
+        pdg_vocab=pdg_vocab,
+        mat_vocab=mat_vocab,
+        vocab_caveats=_vocab_caveats(cfg),
+    )
+
+
+def _tree_lines(module: nn.Module, name: str, indent: int = 0) -> list[str]:
+    total = sum(p.numel() for p in module.parameters())
+    in_dim = getattr(module, "in_dim", None)
+    out_dim = getattr(module, "out_dim", None)
+    widths = f"  [in={in_dim}, out={out_dim}]" if in_dim is not None and out_dim is not None else ""
+    lines = [f"{'  ' * indent}{name} ({type(module).__name__}): {total:,}{widths}"]
+    for child_name, child in module.named_children():
+        lines.extend(_tree_lines(child, child_name, indent + 1))
+    return lines
+
+
+_HEAD_NAMES = ("n_sec_head", "type_head", "stop_head")
+
+
+def _stage_header(name: str, module: nn.Module) -> list[str]:
+    total = sum(p.numel() for p in module.parameters())
+    lines = [f"{name}: {type(module).__name__} -- {total:,} parameters"]
+    generator = getattr(module, "generator_kind", None)
+    if generator is not None:
+        lines.append(f"  generator: {generator}")
+    trunk = getattr(module, "trunk", None)
+    if trunk is not None:
+        in_dim = getattr(trunk, "in_dim", "?")
+        out_dim = getattr(trunk, "out_dim", "?")
+        if isinstance(trunk, RoutedTrunk):
+            detail = f"routed, n_experts={trunk.router.n_experts}, expert type={type(trunk.experts[0]).__name__}"
+        else:
+            detail = f"unrouted, {type(trunk).__name__}"
+        lines.append(f"  trunk: {detail}, in={in_dim}, out={out_dim}")
+    history_kind = getattr(module, "history_kind", None)
+    if history_kind is not None:
+        lines.append(f"  autoregressive history: {history_kind}")
+    present = [h for h in _HEAD_NAMES if getattr(module, h, None) is not None]
+    absent = [h for h in _HEAD_NAMES if hasattr(module, h) and getattr(module, h) is None]
+    if present or absent:
+        lines.append(f"  heads present: {', '.join(present) if present else 'none'}")
+        if absent:
+            lines.append(f"  heads absent: {', '.join(absent)}")
+    return lines
+
+
+def render_summary(summary: ModelSummary) -> str:
+    lines: list[str] = []
+    for name, module in summary.modules.items():
+        lines.extend(_stage_header(name, module))
+        lines.extend(_tree_lines(module, name, indent=1))
+        lines.append("")
+
+    lines.append(
+        f"config keys read during construction: {len(summary.consumed)} / "
+        f"read elsewhere (trainer/sampler/rollout): {len(summary.elsewhere)} / "
+        f"inert under this config: {len(summary.inert)}"
+    )
+    if summary.elsewhere:
+        lines.append("read elsewhere, not by construction:")
+        for path in summary.elsewhere:
+            lines.append(f"  {path}  ({_NOT_BUILD_TIME[path]})")
+    lines.append("inert under this config (declared, parsed, but doing nothing here):")
+    if summary.inert:
+        for path in summary.inert:
+            lines.append(f"  {path}")
+    else:
+        lines.append("  (none)")
+
+    if summary.vocab_caveats:
+        lines.append("")
+        lines.append("vocab placeholder caveats:")
+        for caveat in summary.vocab_caveats:
+            lines.append(f"  {caveat}")
+
+    return "\n".join(lines)

giant/model/trunks.py

@@ -73,6 +73,7 @@ class ExpertTrunk(nn.Module):
         block_conditioning: str = "add",
     ) -> None:
         super().__init__()
+        self.in_dim = in_dim
         self.out_dim = out_dim
         self.input_proj = nn.Linear(in_dim, hidden_dim)
         self.blocks = nn.ModuleList(
@@ -93,6 +94,49 @@ class ExpertTrunk(nn.Module):
         return self.out_proj(x)
 
 
+@register_trunk("linear")
+class LinearTrunk(nn.Module):
+    """`nn.Linear(in_dim + cond_dim, out_dim)` over `concat([x, cond])` —
+    the trivial trunk body: no hidden layer, no ResBlock stack, no
+    nonlinearity. Ablates whether trunk depth/nonlinearity is earning its
+    parameters, holding everything else (heads, ConditionEncoder,
+    generator, ...) fixed. Composes for free with `router.enabled = true`
+    (gitea #33): a RoutedTrunk of n_experts linear bodies is "mixture of
+    trivial linear experts". `hidden_dim`/`n_blocks`/`dropout`/
+    `block_conditioning` are accepted and ignored, matching
+    `build_expert_body`'s shared factory signature.
+
+    `x` — the trunk's own input (e.g. the noised primary vector for flow
+    matching) — does not already carry conditioning; that's fused in
+    per-body via `cond`. So this concatenates `x` and `cond` itself to
+    remain a valid, conditioning-dependent model.
+    """
+
+    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.in_dim = in_dim
+        self.out_dim = out_dim
+        self.linear = nn.Linear(in_dim + cond_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:
+        return self.linear(torch.cat([x, cond], dim=-1))
+
+
 def _route_forward(
     experts: nn.ModuleList,
     router: Router,
@@ -108,21 +152,50 @@ def _route_forward(
     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.
+
+    The accumulator's dtype is deferred to the first expert call rather than
+    fixed at fp32: under autocast (`train.precision = "bf16"`, gitea #47) an
+    expert's `ResBlock` stack returns bf16, and an fp32-fixed accumulator
+    would silently upcast every mixture term (train mode) or downcast every
+    dispatched row via `index_put_` (eval mode) — making a `RoutedTrunk`
+    return a different dtype than the unrouted `ExpertTrunk` it's a drop-in
+    replacement for, purely because `router.enabled` was set.
+
+    `router.combine_weights` is deliberately fp32 internally (it forces its
+    own autocast-disabled region — see `Router.combine_weights`'s docstring),
+    so `weights` itself is always fp32 regardless of the ambient precision.
+    Left as-is, `weights[:, i:i+1] * expert(x, cond)` would type-promote the
+    whole mixture back to fp32 by ordinary PyTorch promotion rules — the same
+    dtype-mismatch bug this function exists to avoid, just moved one line
+    over. `weights` is cast down to each expert's own output dtype right
+    before combining: the softmax stays numerically stable at fp32, but its
+    *result* (values in [0, 1], not precision-sensitive to represent) loses
+    nothing meaningful by then being used at bf16.
     """
     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)
+        weights = router.combine_weights(cond_cont, cond_cat)  # (B, n_experts), fp32
+        out = None
         for i, expert in enumerate(experts):
-            out = out + weights[:, i : i + 1] * expert(x, cond)
+            expert_out = expert(x, cond)
+            term = weights[:, i : i + 1].to(expert_out.dtype) * expert_out
+            out = term if out is None else out + term
+        assert out is not None, "RoutedTrunk built with zero experts"
         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)
+    out = None
     for i, expert in enumerate(experts):
         mask = idx == i
         if mask.any():
-            out[mask] = expert(x[mask], cond[mask])
+            expert_out = expert(x[mask], cond[mask])
+            if out is None:
+                out = torch.zeros(x.shape[0], expert_out.shape[-1], device=x.device, dtype=expert_out.dtype)
+            out[mask] = expert_out
+    if out is None:
+        # No row was ever dispatched (only reachable with an empty batch,
+        # x.shape[0] == 0) — nothing to infer a dtype from, so fall back to
+        # x's own, matching this function's pre-autocast behavior.
+        out = torch.zeros(x.shape[0], experts[0].out_dim, device=x.device, dtype=x.dtype)
     return out
 
 
@@ -130,7 +203,10 @@ 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."""
+    stage. Implementations are expected to expose `in_dim`/`out_dim`
+    attributes (as `ExpertTrunk`/`RoutedTrunk` do) — `giant.model.summary`
+    (gitea #46) reads them to report trunk widths without needing to know the
+    body architecture."""
 
     def forward(
         self,
@@ -157,6 +233,8 @@ class RoutedTrunk(Trunk):
     ) -> None:
         super().__init__()
         self.router = router
+        self.in_dim = in_dim
+        self.out_dim = out_dim
         self.experts = nn.ModuleList(
             [
                 build_expert_body(

giant/model/wgan.py

@@ -22,21 +22,31 @@ def gradient_penalty(
     norm to 1 — `x_hat`/`grad` are forced to all-zero for such a row, which
     would otherwise contribute a constant `(||0|| - 1)^2 == 1` bias to the
     mean regardless of critic behavior — so they're excluded from the mean.
+
+    Deliberately kept fp32 (`torch.autocast(..., enabled=False)`) regardless
+    of the caller's ambient `train.precision` autocast region: this is a
+    `create_graph=True` double-backward, and `grad.norm(2, dim=1)` sums
+    squares over the critic's full input width (hundreds of dims for stage
+    2), which overflows bf16's range at gradient magnitudes well within
+    normal early-WGAN-GP territory. Disclosed cost: the critic forward
+    inside this function always runs fp32, even when the rest of the WGAN
+    stage's step is bf16 (gitea #47).
     """
-    eps = torch.rand(real.size(0), 1, device=real.device)
-    x_hat = eps * real + (1 - eps) * fake
-    if mask is not None:
-        x_hat = x_hat * mask
-    x_hat = x_hat.requires_grad_(True)
-    scores = critic_fn(x_hat)
-    grad = torch.autograd.grad(outputs=scores.sum(), inputs=x_hat, create_graph=True)[0]
-    if mask is not None:
-        grad = grad * mask
-    penalty = (grad.norm(2, dim=1) - 1) ** 2
-    if mask is not None:
-        valid = (mask.sum(dim=1) > 0).float()
-        return (penalty * valid).sum() / valid.sum().clamp_min(1.0)
-    return penalty.mean()
+    with torch.autocast(real.device.type, enabled=False):
+        eps = torch.rand(real.size(0), 1, device=real.device)
+        x_hat = eps * real.float() + (1 - eps) * fake.float()
+        if mask is not None:
+            x_hat = x_hat * mask
+        x_hat = x_hat.requires_grad_(True)
+        scores = critic_fn(x_hat)
+        grad = torch.autograd.grad(outputs=scores.sum(), inputs=x_hat, create_graph=True)[0]
+        if mask is not None:
+            grad = grad * mask
+        penalty = (grad.norm(2, dim=1) - 1) ** 2
+        if mask is not None:
+            valid = (mask.sum(dim=1) > 0).float()
+            return (penalty * valid).sum() / valid.sum().clamp_min(1.0)
+        return penalty.mean()
 
 
 def critic_loss(

giant/pipeline.py

@@ -359,7 +359,7 @@ def run_train_job(
             "section)"
         )
 
-    config.validate_config(cfg)
+    config.validate_config(cfg, resume=resume is not None)
     particle_conditioning = cfg["conditioning"]["particle"]["type"]
     material_conditioning = cfg["conditioning"]["material"]["type"]
     k_max = cfg["stage2_model"]["k_max"]

giant/training/__init__.py

@@ -5,7 +5,7 @@ Split out of the former single-module `giant/train.py`. The public surface is
 that tests and tooling construct directly.
 """
 
-from giant.training.checkpoint import build_checkpoint, load_checkpoint
+from giant.training.checkpoint import build_checkpoint, init_stages_from_checkpoints, load_checkpoint
 from giant.training.metrics import MetricsCollector, MetricSpec
 from giant.training.loop import train
 from giant.training.trainers import (
@@ -25,6 +25,7 @@ __all__ = [
     "WGANStageTrainer",
     "build_checkpoint",
     "build_stage_trainers",
+    "init_stages_from_checkpoints",
     "load_checkpoint",
     "train",
 ]

giant/training/amp.py

@@ -0,0 +1,47 @@
+"""Mixed-precision training support (`train.precision`, gitea #47).
+
+Only `"fp32"` (no autocast) and `"bf16"` are supported — no `"fp16"`/
+`GradScaler`. bf16 needs no gradient scaler and covers every training GPU in
+the fleet (Ampere and newer: A100, L40S, H200, RTX 4070); fp16 would need a
+scaler *and* fixes to two fragile spots that stay correct under bf16 but break
+under fp16's narrower range — `giant.model.routers`' `1e-8` epsilons (below
+fp16's ~6e-8 subnormal floor) and `giant.model.wgan.gradient_penalty`'s
+sum-of-squares gradient norm (overflows fp16 above ~65504). Revisit if a
+pre-Ampere (V100) training target ever shows up.
+"""
+
+import torch
+
+_SUPPORTED_DEVICE_TYPES = ("cuda", "cpu")
+
+
+def resolve_autocast(precision: str, device: torch.device) -> tuple[str, torch.dtype, bool]:
+    """Resolves `train.precision` + a target device into the
+    `(device_type, dtype, enabled)` triple `torch.autocast` takes as kwargs —
+    computed once per `StageTrainer` rather than re-derived every step.
+
+    Raises `ValueError` rather than silently falling back to fp32: a training
+    run that's quietly not using the mixed precision it was configured for is
+    a wasted GPU-week, not a warning.
+    """
+    if precision == "fp32":
+        return device.type, torch.float32, False
+    if precision != "bf16":
+        raise ValueError(f"unknown precision {precision!r}; must be 'fp32' or 'bf16'")
+
+    if device.type == "cuda":
+        if not torch.cuda.is_bf16_supported():
+            cap = torch.cuda.get_device_capability(device)
+            raise ValueError(
+                f"train.precision = 'bf16' but {torch.cuda.get_device_name(device)} "
+                f"(compute capability {cap[0]}.{cap[1]}) has no native bf16 support "
+                "(needs Ampere/sm_80 or newer) — use train.precision = 'fp32' instead"
+            )
+        return "cuda", torch.bfloat16, True
+    if device.type == "cpu":
+        # torch 2.3's CPU autocast supports bf16 unconditionally — this is
+        # also what lets the bf16 training path be tested without a GPU.
+        return "cpu", torch.bfloat16, True
+    raise ValueError(
+        f"train.precision = 'bf16' is not supported on device type {device.type!r} (only {_SUPPORTED_DEVICE_TYPES} are)"
+    )

giant/training/checkpoint.py

@@ -8,6 +8,8 @@ and per-stage `optimizer_<stage>` / `optimizer_d_<stage>` / `lr_sched_<stage>`
 entries.
 """
 
+import torch
+
 from giant.training.trainers import StageTrainer
 
 #: Stage name -> the checkpoint key its weights live under. Historical: stage
@@ -46,6 +48,37 @@ def build_checkpoint(
     return ckpt
 
 
+def init_stages_from_checkpoints(trainers: dict[str, StageTrainer]) -> list[str]:
+    """Load each trainer's `spec.init_from` checkpoint (gitea #42) into its
+    model, before training starts — the partial-retrain counterpart to
+    `load_checkpoint`'s full-run `--resume`. Only weights move: unlike
+    `load_checkpoint`, this never touches optimizer/lr_sched/epoch state, so
+    it composes cleanly with `--resume` (call this first; a resume's own
+    `load_checkpoint` then overwrites whatever this loaded with the resumed
+    run's own weights).
+
+    A stage with no `init_from` set (`""`, the default) is left alone. The
+    EMA companion (`<key>_ema`) is loaded too when both the source checkpoint
+    and this trainer have one, so `--weights ema` at inference still sees the
+    source's EMA shadow rather than a copy of its raw weights. Returns one
+    description string per stage actually initialized, for the caller to
+    echo.
+    """
+    loaded = []
+    for name, trainer in trainers.items():
+        init_from = trainer.spec.init_from
+        if not init_from:
+            continue
+        key = _STAGE_KEY[name]
+        ckpt = torch.load(init_from, map_location="cpu", weights_only=False)
+        trainer.model.load_state_dict(ckpt[key])
+        ema_key = f"{key}_ema"
+        if trainer.ema_model is not None and ema_key in ckpt:
+            trainer.ema_model.load_state_dict(ckpt[ema_key])
+        loaded.append(f"{name}: loaded from {init_from}" + (" (frozen)" if trainer.frozen else ""))
+    return loaded
+
+
 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

giant/training/loop.py

@@ -20,7 +20,7 @@ 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.checkpoint import build_checkpoint, init_stages_from_checkpoints, load_checkpoint
 from giant.training.metrics import MetricsCollector
 from giant.training.trainers import (
     FlowDDPMStageTrainer,
@@ -132,9 +132,12 @@ def train(
     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)
+    sec_type_class_counts = sec_type_topn_map.class_counts if sec_type_topn_map is not None else None
+    trainers = build_stage_trainers(cfg, models, critics, device, total_train_batches, sec_type_class_counts)
     if not trainers:
         raise ValueError("no active stage — stage1_model.active and stage2_model.active are both false")
+    for line in init_stages_from_checkpoints(trainers):
+        print(line)
     has_adversarial = any(not tr.supports_val_loss for tr in trainers.values())
 
     checkpoint_extras = {

giant/training/stage2_inputs.py

@@ -120,9 +120,17 @@ def _remaining_energy_fraction(fraction: torch.Tensor) -> torch.Tensor:
     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)
+    the batch's ground-truth stick logits alone, no `e_sec` required).
+
+    Forced fp32 regardless of the caller's ambient `train.precision` autocast
+    region: a `cumprod` over `K_MAX` slots in bf16 underflows to zero within a
+    handful of slots, killing `remaining_frac` as a conditioning signal — the
+    numpy encoder (`giant.data.transforms.encode_secondaries`'s stick-breaking
+    twin) already promotes to float64 for exactly this reason (gitea #47)."""
+    with torch.autocast(fraction.device.type, enabled=False):
+        fraction = fraction.float()
+        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:
@@ -202,21 +210,37 @@ def _assemble_stage2_ar_inputs(
     return {"history_feat": history_feat, **_ar_meta(K, B, device, fraction)}
 
 
+def _linear_schedule(p_start: float, p_end: float, epoch: int, total_epochs: int) -> float:
+    """Linear interpolation from `p_start` (epoch 0) to `p_end` (the final
+    epoch) — standard scheduled sampling (Bengio et al. 2015), shared by
+    every train-time schedule keyed on epoch."""
+    frac = epoch / max(total_epochs - 1, 1)
+    frac = min(max(frac, 0.0), 1.0)
+    return p_start + (p_end - p_start) * frac
+
+
 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)."""
+    linearly interpolates `p_start` to `p_end` via `_linear_schedule`."""
     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
+    return _linear_schedule(p_start, p_end, epoch, total_epochs)
+
+
+def _ctx_truth_prob(mode: str, p_start: float, p_end: float, epoch: int, total_epochs: int) -> float:
+    """P(condition stage 2 on the TRUE stage-1 outcome rather than a fresh
+    stage-1 sample), for the current epoch (`stage2_model.stage1_context`).
+    `"truth"` is the degenerate constant 1.0; `"sampled"` linearly
+    interpolates `ctx_p_start` to `ctx_p_end` via `_linear_schedule` — the
+    stage-boundary counterpart of `_stage2_tf_prob`."""
+    if mode == "truth":
+        return 1.0
+    return _linear_schedule(p_start, p_end, epoch, total_epochs)
 
 
 def _history_repr_from_ar_sample(

giant/training/trainers.py

@@ -27,10 +27,13 @@ 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.sample import sample_stage1
+from giant.training.amp import resolve_autocast
 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,
+    _ctx_truth_prob,
     _gumbel_tau,
     _relax_onehot_type_slice,
     _stage2_tf_prob,
@@ -73,6 +76,41 @@ def _batch_to_device(batch: StepBatch, device: torch.device) -> StepBatch:
     return type(batch)(*(t.to(device) for t in batch))
 
 
+def _type_class_weight_vector(class_counts: dict[int, int], n_classes: int, scheme: str) -> list[float] | None:
+    """Per-class `F.cross_entropy(weight=...)` vector for the stage-2 type
+    head's `class_weighting` (gitea #44), or `None` under `"none"` (the
+    pre-#44 unweighted-CE behavior — the caller must pass that through as
+    `weight=None`, not a vector of ones, so old runs stay bit-identical).
+
+    `"inverse_freq"`: `1 / count` per class, normalized to mean 1 over
+    `n_classes` so switching this on doesn't rescale the type loss against
+    `particle_type.lambda` / the generator loss it's summed with. A class
+    with zero training examples (fewer distinct species than `n_classes - 1`
+    slots) clamps its count to 1 — its weight is otherwise undefined, and
+    since it never appears in a batch's labels the value is inert anyway.
+
+    Raises if `scheme != "none"` and `class_counts` is empty: that means the
+    `TopNMap` behind this run predates gitea #44 (a stale checkpoint's decode
+    map, or a not-yet-rebuilt setup-cache sidecar) and truly has no
+    frequency information to weight by — silently falling back to uniform
+    weights would look like the feature is active when it isn't.
+    """
+    if scheme == "none":
+        return None
+    if not class_counts:
+        raise ValueError(
+            f"stage2_model.particle_type.class_weighting = {scheme!r} requires "
+            "per-class counts, but this run's sec_type_topn_map has none "
+            "(class_counts={}) — it was built before gitea #44 or loaded "
+            "from a stale setup-cache sidecar/checkpoint; rebuild the setup "
+            "cache (giant train --rebuild-setup-cache) or retrain."
+        )
+    counts = [max(class_counts.get(i, 0), 1) for i in range(n_classes)]
+    inv = [1.0 / c for c in counts]
+    mean_inv = sum(inv) / len(inv)
+    return [w / mean_inv for w in inv]
+
+
 @dataclass(frozen=True)
 class StageSpec:
     """One stage's resolved training configuration.
@@ -87,6 +125,10 @@ class StageSpec:
     generator: str
     decoder: str = "one_shot"
 
+    # partial-retrain (gitea #42)
+    init_from: str = ""
+    freeze: bool = False
+
     # loss weights
     lambda_weight: float = 1.0
     n_sec_lambda: float = 0.1
@@ -95,6 +137,10 @@ class StageSpec:
     # particle-type target (stage 2 only)
     particle_type: ParticleTypeConfig = field(default_factory=ParticleTypeConfig)
     particle_type_n_classes: int = 16
+    # Resolved by from_config from sec_type_class_counts (dataset-derived,
+    # not itself a cfg value — see _type_class_weight_vector) crossed with
+    # particle_type.class_weighting (gitea #44). None under "none".
+    type_class_weights: list[float] | None = None
 
     # optimization
     lr: float = 3e-4
@@ -103,6 +149,7 @@ class StageSpec:
     warmup_epochs: int = 0
     epochs: int = 1
     steps_per_epoch: int = 1
+    precision: str = "fp32"
 
     # routing auxiliaries
     lambda_balance: float = 0.0
@@ -117,6 +164,11 @@ class StageSpec:
     tf_p_end: float = 1.0
     ar_sample_steps: int = 10
 
+    # stage-1/stage-2 boundary (stage 2 only)
+    stage1_context: str = "truth"
+    ctx_p_start: float = 1.0
+    ctx_p_end: float = 0.0
+
     # generator-specific
     ddpm_n_steps: int = 1000
     n_critic: int = 5
@@ -126,7 +178,18 @@ class StageSpec:
     type_gumbel_tau_end: float = 0.1
 
     @classmethod
-    def from_config(cls, cfg: dict, name: str, is_stage2: bool, steps_per_epoch: int) -> "StageSpec":
+    def from_config(
+        cls,
+        cfg: dict,
+        name: str,
+        is_stage2: bool,
+        steps_per_epoch: int,
+        sec_type_class_counts: dict[int, int] | None = None,
+    ) -> "StageSpec":
+        """`sec_type_class_counts` is dataset-derived (`sec_type_topn_map.class_counts`,
+        gitea #44), not a `cfg` value — it's the one input to `StageSpec` that
+        doesn't come from `cfg`, kept separate from the "only place that
+        reads `cfg`" invariant below on purpose."""
         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
@@ -139,17 +202,23 @@ class StageSpec:
         # 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"])
+        particle_type_n_classes = resolve_type_n_classes(
+            s2_spec.particle_type, cfg["conditioning"]["particle"]["emb_dim"]
+        )
         return cls(
             name=name,
             is_stage2=is_stage2,
             generator=stage_spec.generator,
             decoder=s2_spec.decoder if is_stage2 else "one_shot",
+            init_from=stage_spec.init_from,
+            freeze=stage_spec.freeze,
             lambda_weight=stage_spec.lambda_weight,
             n_sec_lambda=s2_spec.n_sec.lambda_weight,
             n_sec_mode=s2_spec.n_sec.mode,
             particle_type=s2_spec.particle_type,
-            particle_type_n_classes=resolve_type_n_classes(
-                s2_spec.particle_type, cfg["conditioning"]["particle"]["emb_dim"]
+            particle_type_n_classes=particle_type_n_classes,
+            type_class_weights=_type_class_weight_vector(
+                sec_type_class_counts or {}, particle_type_n_classes, s2_spec.particle_type.class_weighting
             ),
             # train.* keys are all guaranteed by DEFAULT_CONFIG's deep-merge
             # (giant/config.py), so TrainConfig.from_dict never has to fall
@@ -161,6 +230,7 @@ class StageSpec:
             warmup_epochs=t.warmup_epochs,
             epochs=t.epochs,
             steps_per_epoch=max(steps_per_epoch, 1),
+            precision=t.precision,
             lambda_balance=stage_spec.router.lambda_balance,
             lambda_proc=stage_spec.router.lambda_proc,
             lambda_entropy=stage_spec.router.lambda_entropy,
@@ -169,6 +239,9 @@ class StageSpec:
             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,
+            stage1_context=s2_spec.stage1_context if is_stage2 else cls.stage1_context,
+            ctx_p_start=s2_spec.ctx_p_start if is_stage2 else cls.ctx_p_start,
+            ctx_p_end=s2_spec.ctx_p_end if is_stage2 else cls.ctx_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
@@ -186,13 +259,16 @@ class StageSpec:
 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.
+    Reads only the shared `StepBatch` (`giant.data.dataset`) by default — stage
+    2 conditions on the ground-truth `x1_s1` (`stage2_model.stage1_context =
+    "truth"`, stage-level teacher forcing), so "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. `stage2_model.stage1_context = "sampled"` is the one
+    exception — `build_stage_trainers` wires the stage-2 trainer to the
+    stage-1 one via `attach_stage1` so it can draw a real stage-1 sample
+    (`giant.sample.sample_stage1`) instead, scheduled by `ctx_p_start`/
+    `ctx_p_end` (see `_stage1_context`).
 
     Grad-norm clipping is per-stage here — v0.2's single shared optimizer
     clipped both stages' gradients jointly; splitting per stage is a small,
@@ -231,11 +307,18 @@ class StageTrainer:
         self.is_stage2 = spec.is_stage2
         self.generator = spec.generator
         self.decoder = spec.decoder
+        self.frozen = spec.freeze
         self.device = device
         self.model = model.to(device)
         self.router = _stage_router(self.model)
         self._modules = (self.model, *extra_modules)
 
+        # Resolved once (not re-derived every step) — see
+        # giant.training.amp.resolve_autocast (gitea #47).
+        self._autocast_device_type, self._autocast_dtype, self._autocast_enabled = resolve_autocast(
+            spec.precision, device
+        )
+
         self.particle_type_cfg = spec.particle_type
         self.particle_type_n_classes = spec.particle_type_n_classes
         self.ema_decay = spec.ema_decay
@@ -246,6 +329,18 @@ class StageTrainer:
             for p in self.ema_model.parameters():
                 p.requires_grad_(False)
 
+        #: Set by `attach_stage1` when `stage2_model.stage1_context =
+        #: "sampled"` — the stage-1 `StageTrainer` this (stage-2) trainer
+        #: draws its context sample from. `None` for stage 1 itself, and for
+        #: stage 2 under "truth".
+        self.stage1_source: "StageTrainer | None" = None
+
+    def attach_stage1(self, stage1_trainer: "StageTrainer") -> None:
+        """Wires this (stage-2) trainer to the stage-1 trainer it should
+        sample from under `stage2_model.stage1_context = "sampled"` — see
+        `build_stage_trainers`."""
+        self.stage1_source = stage1_trainer
+
     # --- schedule -------------------------------------------------------
 
     def _init_lr_schedule(self, optimizer: optim.Optimizer, warmup_steps: int, total_steps: int) -> None:
@@ -289,6 +384,58 @@ class StageTrainer:
         for module in self._modules:
             module.eval()
 
+    # --- stage-1/stage-2 boundary (shared by both trainer subclasses) ---
+
+    def _stage1_context(
+        self,
+        x1_s1: torch.Tensor,
+        cond_cont: torch.Tensor,
+        cond_cat: torch.Tensor,
+        epoch: int | None,
+    ) -> torch.Tensor:
+        """The stage-1 outcome stage 2 conditions on this batch.
+
+        `epoch=None` means "always ground truth" regardless of
+        `spec.stage1_context` — the same val-loss convention `_ar_inputs`
+        uses, so validation stays a stable, non-stochastic comparison.
+        Otherwise, under `stage1_context = "sampled"`, each example
+        independently uses the ground truth with probability `p_truth`
+        (`_ctx_truth_prob`, ramped by `ctx_p_start`/`ctx_p_end`) and a fresh
+        `giant.sample.sample_stage1` draw from `stage1_source.sampling_model()`
+        otherwise — a real sampling pass, not a cheap proxy, matching
+        `_assemble_stage2_ar_inputs_scheduled`'s precedent for the equivalent
+        in-stage-2 self-sample. Mixed per example (not per-dimension): a row
+        is either the real ground-truth 9D vector or a real sample, never an
+        elementwise blend of the two.
+        """
+        x1_s1 = x1_s1.detach()
+        if self.stage1_source is None or epoch is None:
+            return x1_s1
+        p_truth = _ctx_truth_prob(
+            self.spec.stage1_context,
+            self.spec.ctx_p_start,
+            self.spec.ctx_p_end,
+            epoch,
+            self.spec.epochs,
+        )
+        if p_truth >= 1.0:
+            return x1_s1
+
+        stage1_model = self.stage1_source.sampling_model()
+        was_training = stage1_model.training
+        sampled, _ = sample_stage1(
+            stage1_model,
+            cond_cont,
+            cond_cat,
+            steps=self.spec.ar_sample_steps,
+            ddpm_steps=self.stage1_source.spec.ddpm_n_steps,
+        )
+        if was_training:
+            stage1_model.train()
+
+        use_truth = torch.rand(x1_s1.size(0), 1, device=x1_s1.device) < p_truth
+        return torch.where(use_truth, x1_s1, sampled).detach()
+
     # --- stage-2 secondary assembly (shared by both trainer subclasses) ---
 
     def _ar_inputs(
@@ -434,14 +581,36 @@ class StageTrainer:
         stop_acc = (((logits >= 0).float() == target).float() * mask_f).sum() / denom
         return l_stop, stop_acc
 
-    @staticmethod
-    def _step_optimizer(optimizer: optim.Optimizer, loss: torch.Tensor, params: list) -> float:
+    def _autocast(self) -> torch.autocast:
+        """The training-step autocast region (`train.precision`, gitea #47).
+
+        Only wraps forward/loss computation — `backward()`/`optimizer.step()`
+        stay outside, and `val_loss` never calls this at all, so validation
+        (and the best-checkpoint selection it drives) stays precision-
+        independent and comparable against every fp32-only run recorded so
+        far. `enabled=False` under `precision = "fp32"` (the default) makes
+        this a true no-op, so callers never need to branch on precision
+        themselves."""
+        return torch.autocast(
+            self._autocast_device_type,
+            dtype=self._autocast_dtype,
+            enabled=self._autocast_enabled,
+        )
+
+    def _step_optimizer(self, 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."""
+        the pre-clip grad norm. The one place the grad-clip constant lives.
+
+        `self.frozen` (`stage{1,2}_model.freeze`, gitea #42) skips only the
+        final `optimizer.step()` — backward/clip still run so loss/grad_norm
+        stay meaningful to watch, but the stage's weights (and, for a WGAN
+        stage, its critic's — this same method is both trainers' single
+        optimizer-step choke point) never move."""
         optimizer.zero_grad()
         loss.backward()
         grad_norm = torch.nn.utils.clip_grad_norm_(params, 1.0)
-        optimizer.step()
+        if not self.frozen:
+            optimizer.step()
         return grad_norm.item()
 
     def _extra_state(self) -> dict:
@@ -504,6 +673,12 @@ class FlowDDPMStageTrainer(StageTrainer):
         # "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
+        # gitea #44: None under class_weighting = "none" (the default),
+        # matching F.cross_entropy's own unweighted default — a real tensor
+        # only materializes when the config asked for one.
+        self.type_class_weights = (
+            None if spec.type_class_weights is None else torch.tensor(spec.type_class_weights, device=device)
+        )
 
         self.params = list(self.model.parameters())
         self.optimizer = optim.AdamW(self.params, lr=spec.lr, weight_decay=spec.weight_decay)
@@ -597,8 +772,12 @@ class FlowDDPMStageTrainer(StageTrainer):
         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")
+            ce = F.cross_entropy(
+                type_out.transpose(1, 2), sec_type_idx, weight=self.type_class_weights, reduction="none"
+            )
             l_type = (ce * mask).sum() / denom
+            # Unweighted, deliberately — type_acc is a diagnostic of raw
+            # per-slot correctness, not the (possibly class-weighted) loss.
             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()
@@ -608,9 +787,10 @@ class FlowDDPMStageTrainer(StageTrainer):
 
     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."""
+        forcing (`p_tf=1.0`) and the ground-truth stage-1 context, regardless
+        of `spec.teacher_forcing`/`spec.stage1_context` — validation should
+        stay a stable, non-stochastic ground-truth comparison; only the
+        training `step` path schedules `p_tf`/`p_truth` by epoch."""
         (
             cond_cont,
             cond_cat,
@@ -621,7 +801,7 @@ class FlowDDPMStageTrainer(StageTrainer):
             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()
+        stage1_ctx = self._stage1_context(x1_s1, cond_cont, cond_cat, epoch)
 
         x1_s2 = None
         ar_inputs = None
@@ -689,10 +869,12 @@ class FlowDDPMStageTrainer(StageTrainer):
                 self.spec.gumbel_tau_end,
             )
         epoch = global_step // self.spec.steps_per_epoch
-        out = self._compute(batch, device, epoch=epoch)
+        with self._autocast():
+            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:
+        if not self.frozen:
+            self.lr_sched.step()
+        if self.ema_model is not None and not self.frozen:
             _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
@@ -846,53 +1028,58 @@ class WGANStageTrainer(StageTrainer):
             sec_type_idx,
         ) = _batch_to_device(batch, device)
         B = cond_cont.size(0)
-        stage1_ctx = x1_s1.detach()
+        epoch = global_step // self.spec.steps_per_epoch
+        stage1_ctx = self._stage1_context(x1_s1, cond_cont, cond_cat, epoch)
         grad_probe: dict[str, float] = {}
 
         ar_inputs = None
-        if not self.is_stage2:
-            real = x1_s1
+        with self._autocast():
+            if not self.is_stage2:
+                real = x1_s1
 
-            def critic_fn(x):
-                return self.critic(x, cond_cont, cond_cat)
+                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, ar_inputs = 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(
+                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, ar_inputs = self._stage2_real_and_fake(
+                    _Stage2RealFakeBatch(cond_cont, cond_cat, n_sec, sec_cont, sec_type_idx),
+                    stage1_ctx,
                     global_step,
-                    self.total_steps,
-                    self.spec.type_gumbel_tau_start,
-                    self.spec.type_gumbel_tau_end,
+                    device,
                 )
-                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
+                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)
+            # --- critic step (every batch) ---
+            fake_detached = fake.detach()
+            real_score = critic_fn(real)
+            fake_score = critic_fn(fake_detached)
+
+        # gradient_penalty forces its own fp32 region internally (see its
+        # docstring) regardless of the ambient autocast above.
         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()
@@ -901,8 +1088,9 @@ class WGANStageTrainer(StageTrainer):
 
         # --- 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)
-        l_stop, stop_acc = self._stop_loss(cond_cont, cond_cat, stage1_ctx, n_sec, device, ar_inputs)
+        with self._autocast():
+            l_nsec, nsec_acc = self._n_sec_loss(cond_cont, cond_cat, stage1_ctx, n_sec, device)
+            l_stop, stop_acc = self._stop_loss(cond_cont, cond_cat, stage1_ctx, n_sec, device, ar_inputs)
 
         # On a non-generator-step batch with no n_sec_head/stop_head on this
         # stage (n_sec now defaults to stage 2), there's nothing for
@@ -910,7 +1098,8 @@ class WGANStageTrainer(StageTrainer):
         # 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 and self.model.stop_head is None
         if did_g_step:
-            g_loss_adv = generator_loss(critic_fn, fake)
+            with self._autocast():
+                g_loss_adv = generator_loss(critic_fn, fake)
             g_loss = self.spec.lambda_weight * g_loss_adv + self.spec.n_sec_lambda * (l_nsec + l_stop)
         else:
             g_loss_adv = torch.zeros((), device=device)
@@ -921,8 +1110,9 @@ class WGANStageTrainer(StageTrainer):
             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:
+            if not self.frozen:
+                self.lr_sched.step()
+            if self.ema_model is not None and not self.frozen:
                 _update_ema(self.ema_model, self.model, self.ema_decay)
 
         return {
@@ -994,15 +1184,27 @@ def build_stage_trainers(
     critics: dict[str, torch.nn.Module | None],
     device: torch.device,
     total_train_batches: int,
+    sec_type_class_counts: dict[int, int] | None = None,
 ) -> dict[str, StageTrainer]:
     """One trainer per active stage — `models[name] is None` means that stage
-    is `active = false` and is simply never constructed."""
+    is `active = false` and is simply never constructed.
+
+    `stage2_model.stage1_context = "sampled"` additionally wires the
+    stage-2 trainer to the stage-1 one (`StageTrainer.attach_stage1`) so it
+    can draw a real stage-1 sample instead of only ever seeing the
+    ground-truth stage-1 outcome — `validate_config` already guarantees both
+    stages are active whenever that config value is set.
+
+    `sec_type_class_counts` (`sec_type_topn_map.class_counts`, gitea #44) is
+    the one dataset-derived input `StageSpec.from_config` needs beyond `cfg`
+    — `None`/absent whenever `stage2_model.particle_type.class_weighting =
+    "none"` (the default), which never reads it."""
     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))
+        spec = StageSpec.from_config(cfg, name, is_stage2, max(total_train_batches, 1), sec_type_class_counts)
         if build_objective(spec.generator).is_adversarial:
             critic = critics.get(name)
             assert critic is not None, (
@@ -1011,4 +1213,9 @@ def build_stage_trainers(
             trainers[name] = WGANStageTrainer(spec, model, critic, device)
         else:
             trainers[name] = FlowDDPMStageTrainer(spec, model, device)
+
+    stage2 = trainers.get("stage2")
+    stage1 = trainers.get("stage1")
+    if stage2 is not None and stage1 is not None and stage2.spec.stage1_context == "sampled":
+        stage2.attach_stage1(stage1)
     return trainers

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "giant"
-version = "0.3.2"
+version = "0.3.6"
 description = "Geant4 step-function surrogate via conditional flow matching"
 readme = "README.md"
 requires-python = ">=3.12"
@@ -26,6 +26,8 @@ dev = [
     "pytest-cov>=5,<8",
     "ruff>=0.15,<1",
     "ty>=0.0.50,<0.1",
+    "bump-my-version>=1.2,<2",
+    "git-cliff>=2,<3",
     "giant[convert,analysis,geometry,wandb]",
 ]
 geometry = [

tests/test_amp.py

@@ -0,0 +1,136 @@
+"""Tests for giant/training/amp.py (gitea #47)."""
+
+import tempfile
+from pathlib import Path
+
+import pytest
+import torch
+
+from giant.model.routers import EnergyRouter
+from giant.model.wgan import gradient_penalty
+from giant.training.amp import resolve_autocast
+from giant.training.stage2_inputs import _remaining_energy_fraction
+from test_train import _base_cfg, _run_train
+
+# ---------------------------------------------------------------------------
+# resolve_autocast
+# ---------------------------------------------------------------------------
+
+
+def test_resolve_autocast_fp32_is_disabled():
+    device_type, dtype, enabled = resolve_autocast("fp32", torch.device("cpu"))
+    assert device_type == "cpu"
+    assert dtype is torch.float32
+    assert enabled is False
+
+
+def test_resolve_autocast_bf16_on_cpu_is_enabled():
+    """CPU bf16 autocast is what lets the mixed-precision path be tested
+    without a GPU (torch 2.3 supports it)."""
+    device_type, dtype, enabled = resolve_autocast("bf16", torch.device("cpu"))
+    assert device_type == "cpu"
+    assert dtype is torch.bfloat16
+    assert enabled is True
+
+
+def test_resolve_autocast_bf16_on_unsupported_cuda_raises(monkeypatch):
+    monkeypatch.setattr(torch.cuda, "is_bf16_supported", lambda: False)
+    monkeypatch.setattr(torch.cuda, "get_device_capability", lambda device=None: (7, 0))
+    monkeypatch.setattr(torch.cuda, "get_device_name", lambda device=None: "Tesla V100")
+    with pytest.raises(ValueError, match="bf16"):
+        resolve_autocast("bf16", torch.device("cuda"))
+
+
+def test_resolve_autocast_bf16_on_mps_raises():
+    with pytest.raises(ValueError, match="bf16"):
+        resolve_autocast("bf16", torch.device("mps"))
+
+
+def test_resolve_autocast_unknown_precision_raises():
+    with pytest.raises(ValueError, match="fp32.*bf16"):
+        resolve_autocast("fp16", torch.device("cpu"))
+
+
+# ---------------------------------------------------------------------------
+# End-to-end: train() under bf16 on CPU
+# ---------------------------------------------------------------------------
+
+
+def test_train_end_to_end_bf16_cpu_completes_and_stores_fp32_params():
+    """Reuses tests/test_train.py's synthetic-batch harness — train() itself
+    is device-agnostic, and CPU bf16 autocast is real (not mocked) in torch
+    2.3, so this is a genuine exercise of the autocast region added to
+    FlowDDPMStageTrainer.step/WGANStageTrainer.step, not just a config
+    passthrough check.
+
+    Also asserts the checkpoint's stored parameters are fp32: autocast only
+    changes the dtype of intermediate activations, never the model's own
+    stored weights — a regression here would mean something accidentally
+    cast the model itself (e.g. `model.to(dtype=torch.bfloat16)`) rather than
+    using autocast."""
+    cfg = _base_cfg()
+    cfg["train"]["precision"] = "bf16"
+    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)
+        for stage_key in ("model", "sec_decoder"):
+            if stage_key not in ckpt:
+                continue
+            for name, tensor in ckpt[stage_key].items():
+                if tensor.is_floating_point():
+                    assert tensor.dtype == torch.float32, f"{stage_key}.{name} is {tensor.dtype}, expected fp32"
+
+
+@pytest.mark.parametrize("generator", ["wgan", "flow"])
+def test_train_end_to_end_bf16_cpu_stage2_generators(generator):
+    """bf16 covers both trainer subclasses (FlowDDPMStageTrainer and
+    WGANStageTrainer) — the wgan default in _base_cfg exercises the
+    generator-forward/critic-scoring autocast region added to
+    WGANStageTrainer.step, and flow exercises the plain _compute wrap."""
+    cfg = _base_cfg()
+    cfg["train"]["precision"] = "bf16"
+    cfg["stage2_model"]["generator"] = generator
+    with tempfile.TemporaryDirectory() as tmp:
+        _run_train(cfg, Path(tmp) / "run")
+
+
+# ---------------------------------------------------------------------------
+# fp32 guards: correct in fp32, quietly degrade in bf16 — stay fp32 even
+# under an active bf16 autocast region.
+# ---------------------------------------------------------------------------
+
+
+def test_remaining_energy_fraction_stays_fp32_under_bf16_autocast():
+    fraction = torch.rand(4, 5).to(torch.bfloat16)
+    with torch.autocast("cpu", dtype=torch.bfloat16, enabled=True):
+        out = _remaining_energy_fraction(fraction)
+    assert out.dtype == torch.float32
+
+
+def test_gradient_penalty_stays_fp32_under_bf16_autocast():
+    critic = torch.nn.Linear(6, 1)
+
+    def critic_fn(x):
+        return critic(x)
+
+    real = torch.randn(4, 6)
+    fake = torch.randn(4, 6)
+    with torch.autocast("cpu", dtype=torch.bfloat16, enabled=True):
+        gp = gradient_penalty(critic_fn, real, fake)
+    assert gp.dtype == torch.float32
+
+
+def test_router_balance_and_entropy_loss_stay_fp32_under_bf16_autocast():
+    router = EnergyRouter(n_experts=3)
+    cond_cont = torch.randn(8, 15)
+    cond_cat = torch.zeros(8, 2, dtype=torch.long)
+    with torch.autocast("cpu", dtype=torch.bfloat16, enabled=True):
+        balance = router.balance_loss(cond_cont, cond_cat)
+        entropy = router.entropy_loss(cond_cont, cond_cat)
+        weights = router.combine_weights(cond_cont, cond_cat)
+    assert balance.dtype == torch.float32
+    assert entropy.dtype == torch.float32
+    assert weights.dtype == torch.float32

tests/test_analysis_reduce.py

@@ -10,9 +10,11 @@ from giant.analysis import reduce as R
 from giant.analysis.sources import (
     SYNTHETIC_TERMINATION_REASONS,
     Side,
+    open_side,
     physical_steps,
     secondaries,
 )
+from giant.data.loader import EVENT_ID_FILE_STRIDE
 
 
 def _rollout_frame() -> pl.LazyFrame:
@@ -194,3 +196,35 @@ def test_pdg_and_material_labels():
     assert G.pdg_label(22) == "gamma"
     assert G.pdg_label(999999) == "999999"
     assert G.material_label("G4_PbWO4") == "PbWO4"
+
+
+def _write_shard(path, event_ids, edeps):
+    pl.DataFrame({"event_id": event_ids, "pdg": [11] * len(event_ids), "edep": edeps}).write_parquet(path)
+
+
+def test_open_side_reference_offsets_event_ids_across_shards(tmp_path):
+    # Each shard is a separate Geant4 job whose own event_id numbering restarts
+    # from 0 — a naive multi-shard scan collides on event_id across shards.
+    _write_shard(tmp_path / "a.parquet", [0, 1], [1.0, 2.0])
+    _write_shard(tmp_path / "b.parquet", [0, 1], [3.0, 4.0])
+    df = open_side(tmp_path, Side.reference).sort("event_id").collect()
+    assert df["event_id"].to_list() == [0, 1, EVENT_ID_FILE_STRIDE, EVENT_ID_FILE_STRIDE + 1]
+    assert df["edep"].to_list() == [1.0, 2.0, 3.0, 4.0]
+    assert "__source_path" not in df.columns
+
+
+def test_open_side_reference_single_file_unchanged(tmp_path):
+    _write_shard(tmp_path / "only.parquet", [0, 1], [1.0, 2.0])
+    df = open_side(tmp_path / "only.parquet", Side.reference).sort("event_id").collect()
+    assert df["event_id"].to_list() == [0, 1]
+    assert "__source_path" not in df.columns
+
+
+def test_open_side_reference_manifest(tmp_path):
+    _write_shard(tmp_path / "a.parquet", [0, 1], [1.0, 2.0])
+    _write_shard(tmp_path / "b.parquet", [0, 1], [3.0, 4.0])
+    manifest = tmp_path / "shards.manifest"
+    manifest.write_text("a.parquet\nb.parquet\n")
+    df = open_side(manifest, Side.reference).sort("event_id").collect()
+    assert df["event_id"].to_list() == [0, 1, EVENT_ID_FILE_STRIDE, EVENT_ID_FILE_STRIDE + 1]
+    assert df["edep"].to_list() == [1.0, 2.0, 3.0, 4.0]

tests/test_cli_new_run.py

@@ -91,6 +91,31 @@ def test_dry_run_writes_nothing(tmp_path: Path):
     assert not out_dir.exists()
 
 
+def test_stage1_init_from_and_freeze_flags_scaffold_a_partial_retrain_config(tmp_path: Path):
+    """gitea #42."""
+    out_dir = tmp_path / "run5"
+    result = runner.invoke(
+        app,
+        [
+            "new-run",
+            "--out",
+            str(out_dir),
+            "--stage1-init-from",
+            "ckpt/stage1_good/best.pt",
+            "--stage1-freeze",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+
+    with open(out_dir / "config.toml", "rb") as f:
+        cfg = tomllib.load(f)
+
+    assert cfg["stage1_model"]["init_from"] == "ckpt/stage1_good/best.pt"
+    assert cfg["stage1_model"]["freeze"] is True
+    assert cfg["stage2_model"]["init_from"] == ""
+    assert cfg["stage2_model"]["freeze"] is False
+
+
 def test_force_guard_refuses_to_clobber_existing_checkpoints(tmp_path: Path):
     out_dir = tmp_path / "run5"
     out_dir.mkdir()

tests/test_cli_train_overrides.py

@@ -97,6 +97,33 @@ def test_wgan_knobs_split_per_stage(monkeypatch, tmp_path):
     assert cfg["stage2_model"]["wgan"]["gp_weight"] == 2.5
 
 
+def test_stage1_init_from_and_freeze_flags_land_in_cfg_and_dont_touch_stage2(monkeypatch, tmp_path):
+    """gitea #42: --stage{1,2}-init-from/--stage{1,2}-freeze are stage-scoped
+    only. --stage1-freeze alone would fail validate_config (freeze requires
+    init_from or --resume), so both flags are passed together here."""
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        ["--stage1-init-from", "ckpt/stage1_good/best.pt", "--stage1-freeze"],
+    )
+    assert cfg["stage1_model"]["init_from"] == "ckpt/stage1_good/best.pt"
+    assert cfg["stage1_model"]["freeze"] is True
+    assert cfg["stage2_model"]["init_from"] == ""
+    assert cfg["stage2_model"]["freeze"] is False
+
+
+def test_stage2_init_from_and_freeze_flags_land_in_cfg_and_dont_touch_stage1(monkeypatch, tmp_path):
+    cfg = _invoke_and_capture_cfg(
+        monkeypatch,
+        tmp_path,
+        ["--stage2-init-from", "ckpt/stage2_good/best.pt", "--stage2-freeze"],
+    )
+    assert cfg["stage2_model"]["init_from"] == "ckpt/stage2_good/best.pt"
+    assert cfg["stage2_model"]["freeze"] is True
+    assert cfg["stage1_model"]["init_from"] == ""
+    assert cfg["stage1_model"]["freeze"] is False
+
+
 def test_batch_size_invalid_string_errors(monkeypatch, tmp_path):
     monkeypatch.setattr(cli, "run_train_job", lambda *a, **kw: None)
     result = runner.invoke(

tests/test_config.py

@@ -97,6 +97,19 @@ def test_trunk_config_defaults_block_conditioning_to_add_for_both_stages():
     assert gconfig.DEFAULT_CONFIG["stage2_model"]["trunk"]["block_conditioning"] == "add"
 
 
+def test_init_from_freeze_default_to_unset_for_both_stages():
+    """gitea #42: a pre-existing config with no init_from/freeze key must
+    reproduce today's from-scratch, always-training behaviour exactly."""
+    assert gconfig.Stage1ModelConfig().init_from == ""
+    assert gconfig.Stage1ModelConfig().freeze is False
+    assert gconfig.Stage2ModelConfig().init_from == ""
+    assert gconfig.Stage2ModelConfig().freeze is False
+    assert gconfig.DEFAULT_CONFIG["stage1_model"]["init_from"] == ""
+    assert gconfig.DEFAULT_CONFIG["stage1_model"]["freeze"] is False
+    assert gconfig.DEFAULT_CONFIG["stage2_model"]["init_from"] == ""
+    assert gconfig.DEFAULT_CONFIG["stage2_model"]["freeze"] is False
+
+
 def test_heads_config_defaults_reproduce_pre_gitea_36_hardcoded_shape():
     """gitea #36: a pre-existing config with no `heads` key must reproduce
     today's hardcoded `hidden_dim // 2`, one-hidden-layer architecture
@@ -119,7 +132,15 @@ def test_particle_type_config_n_classes_defaults_to_zero_and_round_trips():
     assert gconfig.ParticleTypeConfig().n_classes == 0
     spec = gconfig.ParticleTypeConfig.from_dict({"n_classes": 32})
     assert spec.n_classes == 32
-    assert spec.to_dict()["n_classes"] == 32
+
+
+def test_particle_type_config_class_weighting_defaults_to_none_and_round_trips():
+    """gitea #44: an existing config.toml with no
+    stage2_model.particle_type.class_weighting key must reproduce the
+    pre-#44 unweighted-CE behavior exactly."""
+    assert gconfig.ParticleTypeConfig().class_weighting == "none"
+    spec = gconfig.ParticleTypeConfig.from_dict({"class_weighting": "inverse_freq"})
+    assert spec.class_weighting == "inverse_freq"
 
 
 def test_router_config_extra_round_trips_composed_axis_keys():
@@ -691,6 +712,42 @@ def test_validate_config_embedding_target_passes_with_embedding_conditioning():
     gconfig.validate_config(cfg)  # must not raise
 
 
+def test_validate_config_bad_class_weighting_rejected():
+    cfg = _cfg_with(**{"stage2_model.particle_type.class_weighting": "effective_num"})
+    with pytest.raises(ValueError, match="class_weighting"):
+        gconfig.validate_config(cfg)
+
+
+def test_validate_config_class_weighting_requires_onehot_target():
+    cfg = _cfg_with(
+        **{
+            "stage2_model.particle_type.class_weighting": "inverse_freq",
+            "stage2_model.particle_type.target": "physical",
+        }
+    )
+    with pytest.raises(ValueError, match="onehot"):
+        gconfig.validate_config(cfg)
+
+
+def test_validate_config_class_weighting_incompatible_with_wgan_generator():
+    # stage2_model.generator defaults to "wgan" and particle_type.target
+    # defaults to "onehot", so only class_weighting needs overriding here.
+    cfg = _cfg_with(**{"stage2_model.particle_type.class_weighting": "inverse_freq"})
+    with pytest.raises(ValueError, match="wgan"):
+        gconfig.validate_config(cfg)
+
+
+def test_validate_config_class_weighting_passes_with_onehot_and_flow():
+    cfg = _cfg_with(
+        **{
+            "stage2_model.particle_type.class_weighting": "inverse_freq",
+            "stage2_model.particle_type.target": "onehot",
+            "stage2_model.generator": "flow",
+        }
+    )
+    gconfig.validate_config(cfg)  # must not raise
+
+
 def test_validate_config_mixed_particle_material_conditioning_is_valid():
     """The particle and material conditioning axes are configured
     independently and may mix freely — e.g. material
@@ -735,6 +792,29 @@ def test_validate_config_tie_to_stage1_requires_stage1_active():
         assert "tie_to_stage1" in str(e)
 
 
+@pytest.mark.parametrize("stage_name", ["stage1_model", "stage2_model"])
+def test_validate_config_freeze_without_init_from_or_resume_rejected(stage_name):
+    cfg = _cfg_with(**{f"{stage_name}.freeze": True})
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "init_from" in str(e)
+        assert "--resume" in str(e)
+
+
+@pytest.mark.parametrize("stage_name", ["stage1_model", "stage2_model"])
+def test_validate_config_freeze_with_init_from_passes(stage_name):
+    cfg = _cfg_with(**{f"{stage_name}.freeze": True, f"{stage_name}.init_from": "ckpt/best.pt"})
+    gconfig.validate_config(cfg)  # must not raise
+
+
+@pytest.mark.parametrize("stage_name", ["stage1_model", "stage2_model"])
+def test_validate_config_freeze_without_init_from_passes_under_resume(stage_name):
+    cfg = _cfg_with(**{f"{stage_name}.freeze": True})
+    gconfig.validate_config(cfg, resume=True)  # must not raise
+
+
 def test_validate_config_stop_token_accepted_under_autoregressive():
     """DEFAULT_CONFIG's stage2_model.decoder is already "autoregressive"
     (see test_stage2_model_config_defaults_match_documented_v030_intent), so
@@ -780,13 +860,98 @@ def test_validate_config_bad_stop_sampling_rejected():
         assert "stop_sampling" in str(e)
 
 
-def test_validate_config_stage1_context_sampled_not_implemented():
-    cfg = _cfg_with(**{"stage2_model.stage1_context": "sampled"})
+def test_validate_config_default_precision_is_fp32():
+    assert gconfig.DEFAULT_CONFIG["train"]["precision"] == "fp32"
+
+
+def test_validate_config_bf16_precision_accepted():
+    cfg = _cfg_with(**{"train.precision": "bf16"})
+    gconfig.validate_config(cfg)  # no raise
+
+
+@pytest.mark.parametrize("bad", ["fp16", "bogus", ""])
+def test_validate_config_bad_precision_rejected(bad):
+    cfg = _cfg_with(**{"train.precision": bad})
     try:
         gconfig.validate_config(cfg)
         assert False, "expected ValueError"
     except ValueError as e:
-        assert "sampled" in str(e)
+        assert "precision" in str(e)
+
+
+def test_validate_config_stage1_context_sampled_accepted_with_both_stages_active():
+    """gitea #41: 'sampled' is now implemented, so DEFAULT_CONFIG's
+    stage1_model/stage2_model.active = true (both) must let it through."""
+    cfg = _cfg_with(**{"stage2_model.stage1_context": "sampled"})
+    gconfig.validate_config(cfg)  # must not raise
+
+
+def test_validate_config_bad_stage1_context_rejected():
+    cfg = _cfg_with(**{"stage2_model.stage1_context": "bogus"})
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "stage1_context" in str(e)
+
+
+def test_validate_config_stage1_context_sampled_requires_stage1_active():
+    cfg = _cfg_with(
+        **{
+            "stage2_model.stage1_context": "sampled",
+            "stage1_model.active": False,
+        }
+    )
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "sampled" in str(e) and "stage1_model.active" in str(e)
+
+
+def test_validate_config_stage1_context_sampled_requires_stage2_active():
+    cfg = _cfg_with(
+        **{
+            "stage2_model.stage1_context": "sampled",
+            "stage2_model.active": False,
+        }
+    )
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "sampled" in str(e) and "stage2_model.active" in str(e)
+
+
+@pytest.mark.parametrize("key", ["ctx_p_start", "ctx_p_end"])
+@pytest.mark.parametrize("value", [-0.1, 1.1])
+def test_validate_config_ctx_p_out_of_range_rejected(key, value):
+    cfg = _cfg_with(
+        **{
+            "stage2_model.stage1_context": "sampled",
+            f"stage2_model.{key}": value,
+        }
+    )
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert key in str(e)
+
+
+def test_validate_config_stage1_context_sampled_always_truth_rejected_as_noop():
+    cfg = _cfg_with(
+        **{
+            "stage2_model.stage1_context": "sampled",
+            "stage2_model.ctx_p_start": 1.0,
+            "stage2_model.ctx_p_end": 1.0,
+        }
+    )
+    try:
+        gconfig.validate_config(cfg)
+        assert False, "expected ValueError"
+    except ValueError as e:
+        assert "ctx_p_start" in str(e) and "ctx_p_end" in str(e)
 
 
 def test_validate_config_n_sec_truth_rejected_for_rollout_capable_checkpoint():
@@ -1075,6 +1240,11 @@ def test_overrides_from_flags_train_block_passthrough():
     assert overrides == {"train": {"epochs": 5, "lr": 1e-3}}
 
 
+def test_overrides_from_flags_precision_passthrough():
+    overrides = gconfig.overrides_from_flags({"precision": "bf16"})
+    assert overrides == {"train": {"precision": "bf16"}}
+
+
 @pytest.mark.parametrize(
     ("shorthand", "explicit", "path_key"),
     [
@@ -1193,6 +1363,21 @@ def test_overrides_from_flags_critic_sizing_is_stage_scoped_only(stage_flag, sta
     assert overrides == {stage_model: {"wgan": {path_key: 32}}}
 
 
+@pytest.mark.parametrize(
+    ("init_from_flag", "freeze_flag", "stage_model"),
+    [
+        ("stage1_init_from", "stage1_freeze", "stage1_model"),
+        ("stage2_init_from", "stage2_freeze", "stage2_model"),
+    ],
+)
+def test_overrides_from_flags_init_from_freeze_is_stage_scoped_only(init_from_flag, freeze_flag, stage_model):
+    """gitea #42: no shared alias — a checkpoint has one set of weights per
+    stage, so "freeze both stages from the same file" has no sensible
+    meaning."""
+    overrides = gconfig.overrides_from_flags({init_from_flag: "ckpt/best.pt", freeze_flag: True})
+    assert overrides == {stage_model: {"init_from": "ckpt/best.pt", "freeze": True}}
+
+
 # ---------------------------------------------------------------------------
 # checkpoint config-mismatch warnings (unchanged surface, still exercised)
 # ---------------------------------------------------------------------------

tests/test_config_consumed_keys.py

@@ -28,6 +28,7 @@ import ast
 from pathlib import Path
 
 from giant.config import DEFAULT_CONFIG
+from giant.config import leaf_paths as _leaf_paths
 
 _REPO_ROOT = Path(__file__).resolve().parents[1]
 
@@ -53,13 +54,6 @@ _EXCLUDED_FILES = ("giant/model/_legacy.py",)
 # 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 "
@@ -72,19 +66,6 @@ _KNOWN_UNUSED = {
 _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)

tests/test_loader.py

@@ -195,6 +195,10 @@ def test_build_topn_map_from_files_keeps_most_frequent(tmp_path):
     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}
+    # class_counts (gitea #44): per resulting index, "other" is the sum of
+    # everything folded into it (2 + 1 = 3), and the total equals row count.
+    assert m.class_counts == {0: 5, 1: 3, 2: 3}
+    assert sum(m.class_counts.values()) == len(materials)
 
 
 def test_build_topn_map_from_files_fewer_values_than_n_classes(tmp_path):
@@ -205,6 +209,8 @@ def test_build_topn_map_from_files_fewer_values_than_n_classes(tmp_path):
 
     assert m.class_map == {"G4_AIR": 0, "PbWO4": 1}
     assert m.other_members == {}
+    # No "other" bucket ever populated -> no entry for its index either.
+    assert m.class_counts == {0: 1, 1: 1}
 
 
 def test_build_pdg_topn_map_from_files_pools_primary_and_secondary_pdg(tmp_path):
@@ -224,6 +230,7 @@ def test_build_pdg_topn_map_from_files_pools_primary_and_secondary_pdg(tmp_path)
     # pooled: 11 -> 5, 22 -> 1 (primary) + 10 (secondary) = 11
     assert m.class_map[22] == 0
     assert m.class_map[11] == 1
+    assert m.class_counts == {0: 11, 1: 5}
 
 
 def test_build_pdg_topn_map_from_files_missing_sec_pdg_list_column(tmp_path):

tests/test_model_summary.py

@@ -0,0 +1,137 @@
+"""Tests for `giant model summary` (gitea #46)."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+from typer.testing import CliRunner
+
+from giant import config as gconfig
+from giant.cli import app
+from giant.materials import MATERIAL_PROPERTIES
+from giant.model.summary import _NOT_BUILD_TIME, _built_modules, _vocab_caveats, summarize_model
+
+runner = CliRunner()
+
+_PDG_VOCAB = 300
+_MAT_VOCAB = len(MATERIAL_PROPERTIES)
+
+
+def _cfg(overrides: dict | None = None) -> dict:
+    return gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, None, overrides or {})
+
+
+@pytest.fixture(scope="module")
+def default_summary():
+    return summarize_model(_cfg(), pdg_vocab=_PDG_VOCAB, mat_vocab=_MAT_VOCAB)
+
+
+def test_default_config_builds_both_stages_with_a_real_tree(default_summary):
+    assert set(default_summary.modules) >= {"stage1", "stage2"}
+    for module in default_summary.modules.values():
+        assert sum(p.numel() for p in module.parameters()) > 0
+    stage1 = default_summary.modules["stage1"]
+    assert hasattr(stage1, "cond_enc")
+    assert hasattr(stage1, "trunk")
+    assert {"input_proj", "blocks", "out_proj"} <= {n for n, _ in stage1.trunk.named_children()}
+
+
+def test_every_in_scope_leaf_is_classified(default_summary):
+    in_scope = {
+        p
+        for p in gconfig.leaf_paths(gconfig.DEFAULT_CONFIG)
+        if p.split(".", 1)[0] in ("conditioning", "stage1_model", "stage2_model")
+    }
+    classified = set(default_summary.consumed) | set(default_summary.inert) | set(default_summary.elsewhere)
+    assert classified == in_scope
+
+
+def test_not_build_time_allow_list_has_no_stale_entries():
+    in_scope = set(gconfig.leaf_paths(gconfig.DEFAULT_CONFIG))
+    stale = set(_NOT_BUILD_TIME) - in_scope
+    assert not stale, f"_NOT_BUILD_TIME entries no longer in DEFAULT_CONFIG: {sorted(stale)}"
+
+
+def test_router_disabled_by_default_so_its_fields_are_inert(default_summary):
+    assert "stage1_model.router.n_experts" in default_summary.inert
+    assert "stage1_model.router.temperature" in default_summary.inert
+
+
+def test_markov_history_leaves_attention_dims_inert_but_history_itself_consumed(default_summary):
+    assert "stage2_model.autoregressive.attn_n_heads" in default_summary.inert
+    assert "stage2_model.autoregressive.attn_n_layers" in default_summary.inert
+    assert "stage2_model.autoregressive.history" in default_summary.consumed
+
+
+def test_single_literal_branch_fields_are_correctly_seen_as_consumed(default_summary):
+    """Regression guard: n_sec.owner ("stage2"), n_sec.mode ("head") and
+    particle_type.target ("onehot") each branch as `== "one specific other
+    literal"` in giant/model/builders.py|models.py. A naive single generic
+    sentinel probe lands in the same "not that literal" bucket as the
+    current value and never crosses the boundary that actually matters --
+    this is exactly what _STRING_ALTERNATIVES exists to fix."""
+    assert "stage2_model.n_sec.owner" in default_summary.consumed
+    assert "stage2_model.n_sec.mode" in default_summary.consumed
+    assert "stage2_model.particle_type.target" in default_summary.consumed
+
+
+def test_stage1_wgan_generator_swaps_flow_time_dim_for_critic_dims():
+    summary = summarize_model(_cfg({"stage1_model": {"generator": "wgan"}}), pdg_vocab=_PDG_VOCAB, mat_vocab=_MAT_VOCAB)
+    assert "stage1_model.flow.time_dim" in summary.inert
+    assert "stage1_model.wgan.noise_dim" in summary.consumed
+    assert "stage1_model.wgan.critic_hidden_dim" in summary.consumed
+
+
+def test_stage2_one_shot_decoder_makes_autoregressive_block_inert():
+    summary = summarize_model(
+        _cfg({"stage2_model": {"decoder": "one_shot"}}), pdg_vocab=_PDG_VOCAB, mat_vocab=_MAT_VOCAB
+    )
+    assert "stage2_model.autoregressive.history" in summary.inert
+    assert "history_encoder" not in {n for n, _ in summary.modules["stage2"].named_children()}
+
+
+def test_energy_router_enabled_consumes_core_fields_but_not_process_only_fields():
+    summary = summarize_model(
+        _cfg({"stage1_model": {"router": {"enabled": True, "type": "energy", "n_experts": 4}}}),
+        pdg_vocab=_PDG_VOCAB,
+        mat_vocab=_MAT_VOCAB,
+    )
+    assert "stage1_model.router.n_experts" in summary.consumed
+    assert "stage1_model.router.temperature" in summary.consumed
+    # emb_dim/hidden_dim are pdg/process-router-only kwargs -- build_router's
+    # signature filter drops them for an energy router.
+    assert "stage1_model.router.hidden_dim" in summary.inert
+    assert "stage1_model.router.emb_dim" in summary.inert
+
+
+def test_vocab_caveat_text_for_embedding_particle_conditioning():
+    cfg = _cfg({"conditioning": {"particle": {"type": "embedding"}}})
+    caveats = _vocab_caveats(cfg)
+    assert any("pdg_vocab" in c and "embedding" in c for c in caveats)
+    assert not any("mat_vocab" in c for c in caveats)
+
+
+def test_pdg_vocab_flag_changes_embedding_table_size():
+    cfg = _cfg({"conditioning": {"particle": {"type": "embedding"}}})
+    small = _built_modules(cfg, pdg_vocab=10, mat_vocab=_MAT_VOCAB)
+    big = _built_modules(cfg, pdg_vocab=1000, mat_vocab=_MAT_VOCAB)
+    assert big["stage1"].cond_enc.pdg_emb.weight.numel() > small["stage1"].cond_enc.pdg_emb.weight.numel()
+
+
+def test_invalid_combo_exits_nonzero_with_validate_config_message(tmp_path: Path):
+    config_path = tmp_path / "bad.toml"
+    config_path.write_text('[meta]\nconfig_version = 3\n\n[stage2_model.particle_type]\ntarget = "embedding"\n')
+    result = runner.invoke(app, ["model", "summary", "--config", str(config_path)])
+    assert result.exit_code == 1
+    assert "requires conditioning.particle.type = 'embedding'" in result.output
+
+
+def test_cli_default_smoke():
+    result = runner.invoke(app, ["model", "summary"])
+    assert result.exit_code == 0, result.output
+    assert "stage1" in result.output
+    assert "stage2" in result.output
+    assert "parameters" in result.output
+    assert "trunk" in result.output
+    assert "inert under this config" in result.output

tests/test_network.py

@@ -8,8 +8,12 @@ from giant.model.network import (
     HISTORY_REGISTRY,
     AttentionHistory,
     ConditionEncoder,
+    CriticModel,
+    FilmResBlock,
     HistoryEncoder,
+    LinearTrunk,
     MarkovHistory,
+    NoHistory,
     SinusoidalEmbedding,
     Stage1Model,
     Stage2Autoregressive,
@@ -465,16 +469,52 @@ def test_attention_history_step_matches_forward():
     assert torch.allclose(stepped, expected, atol=1e-5)
 
 
+# --- NoHistory (gitea #45) ----------------------------------------------------
+
+
+def test_no_history_shape():
+    hist = NoHistory(in_dim=7, out_dim=12)
+    B, K = 3, 5
+    feat = torch.randn(B, K, 7)
+    has_prev = (torch.arange(K) >= 1).unsqueeze(0).expand(B, -1)
+    out = hist(feat, has_prev)
+    assert out.shape == (B, K, 12)
+
+
+def test_no_history_ignores_feat_and_has_prev():
+    hist = NoHistory(in_dim=4, out_dim=6)
+    B, K = 2, 3
+    has_prev_a = (torch.arange(K) >= 1).unsqueeze(0).expand(B, -1)
+    has_prev_b = torch.zeros(B, K, dtype=torch.bool)
+    feat_a = torch.randn(B, K, 4)
+    feat_b = torch.randn(B, K, 4) * 100
+    out_a = hist(feat_a, has_prev_a)
+    out_b = hist(feat_b, has_prev_b)
+    assert torch.equal(out_a, torch.zeros(B, K, 6))
+    assert torch.equal(out_a, out_b)
+
+
+def test_no_history_uses_base_class_o1_defaults():
+    hist = NoHistory(in_dim=4, out_dim=6)
+    assert hist.init_cache() is None
+    feat = torch.randn(2, 1, 4)
+    has_prev = torch.ones(2, 1, dtype=torch.bool)
+    out, cache = hist.step(feat, has_prev, "unused-cache")
+    assert torch.equal(out, torch.zeros(2, 1, 6))
+    assert cache == "unused-cache"
+
+
 # --- HISTORY_REGISTRY / build_history (gitea #35) ----------------------------
 
 
-def test_history_registry_has_exactly_the_two_known_histories():
-    assert set(HISTORY_REGISTRY) == {"markov", "attention"}
+def test_history_registry_has_exactly_the_known_histories():
+    assert set(HISTORY_REGISTRY) == {"markov", "attention", "none"}
 
 
 def test_build_history_returns_correct_concrete_type():
     assert isinstance(build_history("markov", 4, 6), MarkovHistory)
     assert isinstance(build_history("attention", 4, 8), AttentionHistory)
+    assert isinstance(build_history("none", 4, 6), NoHistory)
 
 
 def test_build_history_unknown_name_raises():
@@ -605,7 +645,7 @@ def test_stage2_autoregressive_n_sec_head_and_type_head_cfg_control_hidden_width
 
 @pytest.mark.parametrize("target", ["physical", "onehot", "embedding"])
 @pytest.mark.parametrize("generator", ["wgan", "flow"])
-@pytest.mark.parametrize("history", ["markov", "attention"])
+@pytest.mark.parametrize("history", ["markov", "attention", "none"])
 def test_stage2_autoregressive_forward_shape(target, generator, history):
     B, K, emb_dim = 4, 5, 6
     model = _build_stage2_ar(target, generator, emb_dim=emb_dim, k_max=K, history=history)
@@ -907,7 +947,7 @@ def test_build_critics_particle_type_n_classes_overrides_conditioning_emb_dim():
     assert wider_critic is not None
     # k_max=3 slots, each CONT_SLOT_DIM + n_classes wide under wgan folding —
     # widening n_classes alone (emb_dim stays 4) must widen the critic input.
-    assert wider_critic.input_proj.in_features > default_n_classes_critic.input_proj.in_features
+    assert wider_critic.trunk.input_proj.in_features > default_n_classes_critic.trunk.input_proj.in_features
 
 
 # ── build_models/build_critics: DEFAULT_CONFIG fallback drift (issues.md #1) ─
@@ -984,12 +1024,12 @@ def test_build_critics_omitted_particle_type_matches_default_config():
     cfg["stage2_model"]["generator"] = "wgan"
     onehot_critic = build_critics(cfg)["stage2"]
     assert onehot_critic is not None
-    onehot_in_dim = onehot_critic.input_proj.in_features
+    onehot_in_dim = onehot_critic.trunk.input_proj.in_features
 
     cfg["stage2_model"]["particle_type"] = {"target": "physical"}
     physical_critic = build_critics(cfg)["stage2"]
     assert physical_critic is not None
-    physical_in_dim = physical_critic.input_proj.in_features
+    physical_in_dim = physical_critic.trunk.input_proj.in_features
 
     # onehot's per-slot type width is emb_dim classes vs. physical's fixed
     # (log-mass, charge) pair — different unless emb_dim happens to be 2, so
@@ -1009,15 +1049,15 @@ def test_build_critics_stage1_critic_hidden_dim_and_n_res_blocks_override_genera
 
     inherited = build_critics(cfg)["stage1"]
     assert inherited is not None
-    assert inherited.input_proj.out_features == 8
-    assert len(inherited.blocks) == 1
+    assert inherited.trunk.input_proj.out_features == 8
+    assert len(inherited.trunk.blocks) == 1
 
     cfg["stage1_model"]["wgan"]["critic_hidden_dim"] = 16
     cfg["stage1_model"]["wgan"]["critic_n_res_blocks"] = 3
     overridden = build_critics(cfg)["stage1"]
     assert overridden is not None
-    assert overridden.input_proj.out_features == 16
-    assert len(overridden.blocks) == 3
+    assert overridden.trunk.input_proj.out_features == 16
+    assert len(overridden.trunk.blocks) == 3
 
 
 def test_build_critics_stage2_critic_hidden_dim_and_n_res_blocks_override_generator_size():
@@ -1028,15 +1068,15 @@ def test_build_critics_stage2_critic_hidden_dim_and_n_res_blocks_override_genera
 
     inherited = build_critics(cfg)["stage2"]
     assert inherited is not None
-    assert inherited.input_proj.out_features == 8
-    assert len(inherited.blocks) == 1
+    assert inherited.trunk.input_proj.out_features == 8
+    assert len(inherited.trunk.blocks) == 1
 
     cfg["stage2_model"]["wgan"]["critic_hidden_dim"] = 16
     cfg["stage2_model"]["wgan"]["critic_n_res_blocks"] = 3
     overridden = build_critics(cfg)["stage2"]
     assert overridden is not None
-    assert overridden.input_proj.out_features == 16
-    assert len(overridden.blocks) == 3
+    assert overridden.trunk.input_proj.out_features == 16
+    assert len(overridden.trunk.blocks) == 3
 
 
 # ── StageModel base (gitea #39): Stage1Model/Stage2OneShot/Stage2Autoregressive
@@ -1197,3 +1237,34 @@ def test_stagemodel_time_emb_matches_objective_needs_time(cls, generator):
     assert model.generator_kind == generator
     assert model.noise_dim == 8
     assert (model.time_emb is not None) == build_objective(generator).needs_time
+
+
+# ── CriticModel uses the trunk/block registries + StageModel base (gitea #57) ─
+
+
+def test_critic_model_is_stagemodel_subclass():
+    assert issubclass(CriticModel, StageModel)
+
+
+@pytest.mark.parametrize("stage", ["stage1", "stage2"])
+def test_build_critics_threads_trunk_type_from_generator_config(stage):
+    cfg = _minimal_model_config(share_stages=False)
+    cfg["stage1_model"]["generator"] = "wgan"
+    cfg["stage2_model"]["generator"] = "wgan"
+    cfg[f"{stage}_model"]["trunk"] = {"type": "linear"}
+
+    critic = build_critics(cfg)[stage]
+    assert critic is not None
+    assert isinstance(critic.trunk, LinearTrunk)
+
+
+@pytest.mark.parametrize("stage", ["stage1", "stage2"])
+def test_build_critics_threads_block_conditioning_from_generator_config(stage):
+    cfg = _minimal_model_config(share_stages=False)
+    cfg["stage1_model"]["generator"] = "wgan"
+    cfg["stage2_model"]["generator"] = "wgan"
+    cfg[f"{stage}_model"]["trunk"] = {"block_conditioning": "film"}
+
+    critic = build_critics(cfg)[stage]
+    assert critic is not None
+    assert all(isinstance(block, FilmResBlock) for block in critic.trunk.blocks)

tests/test_pipeline.py

@@ -231,7 +231,7 @@ def test_run_train_job_builds_caches_and_persists_material_topn_map(tmp_path, da
 
 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}
+    cfg["stage2_model"]["particle_type"].update({"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)
 

tests/test_release_tooling.py

@@ -0,0 +1,77 @@
+"""Config-correctness tests for the CI version-bump/tag/changelog automation
+(gitea #50). The workflow YAML itself can only be exercised by a real push to
+master, so these check the two config files it drives (.bumpversion.toml,
+cliff.toml) against real repo content instead.
+"""
+
+import re
+import shutil
+import subprocess
+import tomllib
+from pathlib import Path
+
+import pytest
+
+_ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_bumpversion_search_pattern_matches_pyproject():
+    bump_config = tomllib.loads((_ROOT / ".bumpversion.toml").read_text())["tool"]["bumpversion"]
+    current_version = bump_config["current_version"]
+    search = bump_config["files"][0]["search"].format(current_version=current_version)
+
+    pyproject = (_ROOT / "pyproject.toml").read_text()
+    assert search in pyproject, (
+        f"bumpversion search pattern {search!r} (rendered from .bumpversion.toml's "
+        f"current_version={current_version!r}) not found in pyproject.toml — "
+        "the bump would silently edit nothing"
+    )
+
+
+@pytest.mark.skipif(shutil.which("git-cliff") is None, reason="git-cliff binary not on PATH")
+def test_cliff_config_groups_and_links_commits(tmp_path):
+    repo = tmp_path / "repo"
+    repo.mkdir()
+    subprocess.run(["git", "init", "-q"], cwd=repo, check=True)
+    subprocess.run(["git", "config", "user.name", "test"], cwd=repo, check=True)
+    subprocess.run(["git", "config", "user.email", "test@example.com"], cwd=repo, check=True)
+
+    _commit(repo, "Add class-balanced secondary particle-type loss (gitea #44)")
+    _commit(repo, "Fix leaking secondary energy budget")
+    _commit(repo, "Merge pull request 'Add X' (#1) from fix/issue-1 into master")
+    _commit(repo, "chore: bump version 0.3.3 -> 0.3.4 [skip ci]")
+
+    result = subprocess.run(
+        [
+            "git-cliff",
+            "--config",
+            str(_ROOT / "cliff.toml"),
+            "--repository",
+            str(repo),
+            "--tag",
+            "v0.3.4",
+            "--unreleased",
+        ],
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    changelog = result.stdout
+
+    assert "## [0.3.4]" in changelog
+    assert "### Added" in changelog
+    assert "### Fixed" in changelog
+    assert re.search(
+        r"\[gitea #44\]\(https://git\.larsbogner\.de/lars/giant/issues/44\)",
+        changelog,
+    )
+    assert "Add class-balanced secondary particle-type loss" in changelog
+    assert "Fix leaking secondary energy budget" in changelog
+    assert "Merge pull request" not in changelog
+    assert "skip ci" not in changelog
+
+
+def _commit(repo: Path, message: str) -> None:
+    (repo / "f.txt").write_text(message)
+    subprocess.run(["git", "add", "f.txt"], cwd=repo, check=True)
+    subprocess.run(["git", "commit", "-q", "-m", message], cwd=repo, check=True)

tests/test_router.py

@@ -13,6 +13,8 @@ from giant.model.network import (
     EnergyRouter,
     ExpertTrunk,
     FilmResBlock,
+    LinearTrunk,
+    NoneRouter,
     PdgRouter,
     ProcessRouter,
     ROUTER_REGISTRY,
@@ -73,6 +75,34 @@ def test_energy_router_registered():
     assert ROUTER_REGISTRY["energy"] is EnergyRouter
 
 
+# ── NoneRouter (gitea #45) ──────────────────────────────────────────────────
+
+
+def test_none_router_registered():
+    assert ROUTER_REGISTRY["none"] is NoneRouter
+
+
+def test_none_router_gate_is_uniform():
+    router = NoneRouter(n_experts=4)
+    cond_cont, cond_cat = _cond(16)
+    g = router.gate(cond_cont, cond_cat)
+    assert g.shape == (16, 4)
+    torch.testing.assert_close(g, torch.full((16, 4), 0.25))
+
+
+def test_none_router_gate_ignores_conditioning():
+    router = NoneRouter(n_experts=3)
+    cond_cont_a, cond_cat_a = _cond(8)
+    cond_cont_b, cond_cat_b = _cond(8)
+    torch.testing.assert_close(router.gate(cond_cont_a, cond_cat_a), router.gate(cond_cont_b, cond_cat_b))
+
+
+def test_none_router_top1_always_expert_zero():
+    router = NoneRouter(n_experts=4)
+    cond_cont, cond_cat = _cond(16)
+    assert torch.equal(router.top1(cond_cont, cond_cat), torch.zeros(16, dtype=torch.long))
+
+
 def test_energy_router_gate_partition_of_unity():
     router = EnergyRouter(n_experts=4)
     cond_cont, cond_cat = _cond(16)
@@ -168,6 +198,47 @@ def test_build_expert_body_unknown_type_raises():
     raise AssertionError("expected ValueError for unknown trunk type")
 
 
+# ── LinearTrunk (gitea #45) ──────────────────────────────────────────────────
+
+
+def test_trunk_registry_has_linear():
+    assert "linear" in TRUNK_REGISTRY
+    assert TRUNK_REGISTRY["linear"] is LinearTrunk
+
+
+def test_linear_trunk_forward_shape():
+    trunk = build_expert_body("linear", in_dim=9, out_dim=9, hidden_dim=64, n_blocks=6, cond_dim=12)
+    assert trunk.in_dim == 9
+    assert trunk.out_dim == 9
+    x = torch.randn(5, 9)
+    cond = torch.randn(5, 12)
+    out = trunk(x, cond)
+    assert out.shape == (5, 9)
+
+
+def test_linear_trunk_depends_on_x_and_cond():
+    trunk = build_expert_body("linear", in_dim=9, out_dim=9, hidden_dim=64, n_blocks=6, cond_dim=12)
+    x = torch.randn(5, 9)
+    cond_a = torch.randn(5, 12)
+    cond_b = torch.randn(5, 12)
+    assert not torch.allclose(trunk(x, cond_a), trunk(x, cond_b))
+
+
+def test_routed_linear_trunk_is_mixture_of_trivial_experts():
+    """trunk.type = 'linear' composes for free with router.enabled = true
+    (gitea #33's comment on this issue) — a RoutedTrunk of n_experts linear
+    bodies."""
+    router = build_router("energy", n_experts=3)
+    trunk = RoutedTrunk(router, "linear", in_dim=9, out_dim=9, hidden_dim=64, n_res_blocks=6, cond_dim=12)
+    assert len(trunk.experts) == 3
+    assert all(isinstance(e, LinearTrunk) for e in trunk.experts)
+    x = torch.randn(5, 9)
+    cond = torch.randn(5, 12)
+    cond_cont, cond_cat = _cond(5)
+    out = trunk(x, cond, cond_cont, cond_cat)
+    assert out.shape == (5, 9)
+
+
 # ── BLOCK_REGISTRY / build_block (gitea #34) ────────────────────────────────
 
 
@@ -1195,3 +1266,37 @@ def test_build_models_routed_pair_is_drop_in_for_sample_flow():
     )
     assert sec_cont.shape == (B, K_MAX, 4)
     assert sec_valid.shape == (B, K_MAX)
+
+
+def test_routed_and_unrouted_trunk_agree_on_dtype_under_bf16_autocast():
+    """gitea #47 regression: `_route_forward`'s accumulator (giant/model/
+    trunks.py) used to be a hard-fp32 `torch.zeros`, so under autocast a
+    `RoutedTrunk` returned fp32 while an unrouted `ExpertTrunk` returned
+    bf16 — `router.enabled` alone silently changed the model's output dtype.
+    Checked in both train mode (the differentiable mixture sum) and eval
+    mode (the masked `out[mask] = expert(...)` dispatch) — the two branches
+    of `_route_forward` had independent copies of the bug."""
+    torch.manual_seed(0)
+    cond_cont, cond_cat = _cond(B=6)
+    x = torch.randn(6, X_DIM)
+    t = torch.rand(6)
+
+    unrouted = Stage1Model(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=PARTICLE_CFG,
+        material_cfg=MATERIAL_CFG,
+        hidden_dim=16,
+        n_res_blocks=2,
+    )
+    routed = _routed_stage1(n_experts=3)
+
+    for train_mode in (True, False):
+        unrouted.train(train_mode)
+        routed.train(train_mode)
+        with torch.autocast("cpu", dtype=torch.bfloat16, enabled=True):
+            out_unrouted = unrouted(x, cond_cont, cond_cat, t=t)
+            out_routed = routed(x, cond_cont, cond_cat, t=t)
+        assert out_unrouted.dtype == out_routed.dtype, (
+            f"train={train_mode}: unrouted returned {out_unrouted.dtype}, routed returned {out_routed.dtype}"
+        )

tests/test_setup_cache.py

@@ -99,7 +99,7 @@ def test_save_load_round_trip_topn_maps(tmp_path):
 
     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}
+        class_map={22: 0, 11: 1, 2212: 2}, other_members={2212: 5}, class_counts={0: 100, 1: 50, 2: 5}
     )
     cache.topn_maps[setup_cache.topn_key("material", 2)] = TopNMap(
         class_map={"G4_AIR": 0, "PbWO4": 1}, other_members={}
@@ -114,9 +114,21 @@ def test_save_load_round_trip_topn_maps(tmp_path):
     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)
+    # class_counts (gitea #44) round-trips too, keyed by class index (always
+    # int, independent of the pdg/material axis's own key type).
+    assert pdg_m.class_counts == {0: 100, 1: 50, 2: 5}
+    assert all(isinstance(k, int) for k in pdg_m.class_counts)
 
     mat_m = loaded.topn_maps[setup_cache.topn_key("material", 2)]
     assert mat_m.class_map == {"G4_AIR": 0, "PbWO4": 1}
+    assert mat_m.class_counts == {}
+
+
+def test_topnmap_from_json_missing_class_counts_defaults_empty():
+    """A checkpoint's topn map predating gitea #44 has no class_counts key at
+    all — must decode to {}, not raise, since inference never reads it."""
+    m = setup_cache.topnmap_from_json({"class_map": {"11": 0}, "other_members": {}}, axis="pdg")
+    assert m.class_counts == {}
 
 
 def test_topn_key_unknown_axis_raises():

tests/test_train.py

@@ -5,10 +5,12 @@ import csv
 import math
 import tempfile
 from pathlib import Path
-from unittest.mock import MagicMock
+from unittest.mock import MagicMock, patch
 
+import numpy as np
 import pytest
 import torch
+import torch.nn.functional as F
 
 from giant.config import ParticleTypeConfig
 from giant.constants import (
@@ -19,16 +21,22 @@ from giant.constants import (
     SEC_SLOT_DIM,
     X_DIM,
 )
+from giant.checkpoint_io import load_for_inference
 from giant.data.dataset import StepBatch
+from giant.data.transforms import Normalizer
 from giant.model.network import Stage2Autoregressive, build_critics, build_models
+from giant.sample import sample_stage1 as trainers_sample_stage1
 from giant.training import (
     FlowDDPMStageTrainer,
     StageSpec,
     WGANStageTrainer,
+    build_checkpoint,
     build_stage_trainers,
+    init_stages_from_checkpoints,
     train,
 )
 from giant.training.metrics import _wandb_run_config
+from giant.training.trainers import _type_class_weight_vector
 from giant.training.stage2_inputs import (
     _ar_has_prev,
     _assemble_stage2_ar_inputs,
@@ -351,7 +359,7 @@ def _model_config(cfg):
     }
 
 
-def _run_train(cfg, out_dir, resume_path=None):
+def _run_train(cfg, out_dir, resume_path=None, normalizer_dict=None):
     model_config = _model_config(cfg)
     models = build_models(model_config)
     critics = build_critics(model_config)
@@ -365,7 +373,7 @@ def _run_train(cfg, out_dir, resume_path=None):
         val_loader=val_loader,
         device=torch.device("cpu"),
         out_dir=out_dir,
-        normalizer_dict={"cond": {}, "target": {}, "sec_phys": {}},
+        normalizer_dict=normalizer_dict or {"cond": {}, "target": {}, "sec_phys": {}},
         pdg_map={"22": 0},
         mat_map={"G4_AIR": 0},
         proc_map=None,
@@ -593,6 +601,313 @@ def test_flow_stage_trainer_ddpm_not_implemented_for_stage2():
         FlowDDPMStageTrainer(spec, torch.nn.Linear(1, 1), torch.device("cpu"))
 
 
+# --- gitea #44: class-balanced secondary particle-type loss -----------------
+
+
+def test_type_class_weight_vector_none_scheme_returns_none():
+    assert _type_class_weight_vector({0: 100, 1: 5}, n_classes=2, scheme="none") is None
+
+
+def test_type_class_weight_vector_raises_without_counts():
+    with pytest.raises(ValueError, match="class_counts"):
+        _type_class_weight_vector({}, n_classes=4, scheme="inverse_freq")
+
+
+def test_type_class_weight_vector_inverse_freq_favors_rare_class_and_has_mean_one():
+    weights = _type_class_weight_vector({0: 1000, 1: 10, 2: 1, 3: 1}, n_classes=4, scheme="inverse_freq")
+    assert weights is not None
+    assert len(weights) == 4
+    assert weights[1] > weights[0]  # rarer class -> larger weight
+    assert math.isclose(sum(weights) / len(weights), 1.0, rel_tol=1e-9)
+
+
+def test_type_class_weight_vector_missing_index_clamps_to_count_one():
+    # n_classes=3 but only index 0 was ever observed (e.g. a tiny dataset) —
+    # indices 1/2 must not divide by zero.
+    weights = _type_class_weight_vector({0: 10}, n_classes=3, scheme="inverse_freq")
+    assert weights is not None
+    assert all(math.isfinite(w) for w in weights)
+
+
+def _onehot_flow_stage2_setup():
+    """A built stage-2 model + a batch, under target='onehot' + generator='flow'
+    (mirrors the 'stage2_onehot_target_flow' case in test_train_end_to_end)."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["generator"] = "flow"
+    cfg["stage2_model"]["particle_type"] = {"target": "onehot", "lambda": 1.0}
+    model_config = _model_config(cfg)
+    model = build_models(model_config)["stage2"]
+    assert model is not None
+    batch = _fake_batches(1, 8)[0]
+    device = torch.device("cpu")
+    cond_cont, cond_cat, x1_s1 = batch.cond_cont, batch.cond_cat, batch.target_s1
+    # Mostly class 0 (common), a few slot 1's set to class 1 (rare) —
+    # PARTICLE_CFG's emb_dim=8, n_classes=0 (inherit) -> 8 type classes.
+    sec_type_idx = torch.zeros(8, K_MAX, dtype=torch.long)
+    sec_type_idx[:, :2] = 1
+    sec_mask = torch.ones(8, K_MAX, dtype=torch.bool)
+    return model, cond_cont, cond_cat, x1_s1, sec_type_idx, sec_mask, device
+
+
+def test_flow_ddpm_trainer_type_loss_none_leaves_weight_unset():
+    model, *_ = _onehot_flow_stage2_setup()
+    spec = StageSpec(
+        name="stage2",
+        is_stage2=True,
+        generator="flow",
+        particle_type=ParticleTypeConfig(target="onehot", class_weighting="none"),
+        particle_type_n_classes=8,
+        ema_decay=0.0,
+    )
+    trainer = FlowDDPMStageTrainer(spec, model, torch.device("cpu"))
+    assert trainer.type_class_weights is None
+
+
+def test_flow_ddpm_trainer_type_loss_matches_manual_weighted_cross_entropy():
+    model, cond_cont, cond_cat, x1_s1, sec_type_idx, sec_mask, device = _onehot_flow_stage2_setup()
+    class_counts = {0: 1000, 1: 10, 2: 1, 3: 1, 4: 1, 5: 1, 6: 1, 7: 1}
+    weights = _type_class_weight_vector(class_counts, n_classes=8, scheme="inverse_freq")
+    spec = StageSpec(
+        name="stage2",
+        is_stage2=True,
+        generator="flow",
+        particle_type=ParticleTypeConfig(target="onehot", class_weighting="inverse_freq"),
+        particle_type_n_classes=8,
+        type_class_weights=weights,
+        ema_decay=0.0,
+    )
+    trainer = FlowDDPMStageTrainer(spec, model, device)
+    assert trainer.type_class_weights is not None
+    stage1_ctx = trainer._stage1_context(x1_s1, cond_cont, cond_cat, epoch=None)
+
+    with torch.no_grad():
+        type_out = model.predict_type(cond_cont, cond_cat, stage1_ctx)
+        weight_t = torch.tensor(weights)
+        ce = F.cross_entropy(type_out.transpose(1, 2), sec_type_idx, weight=weight_t, reduction="none")
+        expected = (ce * sec_mask.float()).sum() / sec_mask.float().sum().clamp(min=1)
+
+        l_type, _ = trainer._type_loss(cond_cont, cond_cat, stage1_ctx, sec_type_idx, sec_mask, device)
+
+    assert torch.allclose(l_type, expected, atol=1e-6)
+
+    # Unweighted trainer, same model/batch — the two losses must differ
+    # (the batch mixes the common and rare classes, so weighting changes the
+    # per-slot contributions), confirming the weight is actually plumbed in.
+    spec_none = StageSpec(
+        name="stage2",
+        is_stage2=True,
+        generator="flow",
+        particle_type=ParticleTypeConfig(target="onehot", class_weighting="none"),
+        particle_type_n_classes=8,
+        ema_decay=0.0,
+    )
+    trainer_none = FlowDDPMStageTrainer(spec_none, model, device)
+    with torch.no_grad():
+        l_type_none, _ = trainer_none._type_loss(cond_cont, cond_cat, stage1_ctx, sec_type_idx, sec_mask, device)
+    assert not torch.allclose(l_type, l_type_none)
+
+
+def test_build_stage_trainers_threads_sec_type_class_counts_into_weights():
+    cfg = _base_cfg()
+    cfg["stage2_model"]["generator"] = "flow"
+    cfg["stage2_model"]["particle_type"] = {
+        "target": "onehot",
+        "lambda": 1.0,
+        "class_weighting": "inverse_freq",
+    }
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    class_counts = {i: 100 for i in range(8)}
+    class_counts[1] = 1  # one rare class
+    trainers = build_stage_trainers(
+        cfg, models, critics, torch.device("cpu"), total_train_batches=4, sec_type_class_counts=class_counts
+    )
+    stage2_trainer = trainers["stage2"]
+    assert isinstance(stage2_trainer, FlowDDPMStageTrainer)
+    weights = stage2_trainer.type_class_weights
+    assert weights is not None
+    assert weights[1] > weights[0]
+
+
+def test_build_stage_trainers_no_class_counts_with_none_weighting_is_fine():
+    """The overwhelmingly common case (class_weighting = 'none', the
+    default): build_stage_trainers must not require sec_type_class_counts at
+    all."""
+    cfg = _base_cfg()
+    cfg["stage2_model"]["generator"] = "flow"
+    cfg["stage2_model"].__setitem__("particle_type", {"target": "onehot", "lambda": 1.0})
+    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)
+    stage2_trainer = trainers["stage2"]
+    assert isinstance(stage2_trainer, FlowDDPMStageTrainer)
+    assert stage2_trainer.type_class_weights is None
+
+
+# --- gitea #42: freeze / init_from -------------------------------------------
+
+
+def _state_dict_clone(module):
+    return {k: v.clone() for k, v in module.state_dict().items()}
+
+
+def _assert_state_dicts_equal(before, after, label):
+    for key, value in before.items():
+        assert torch.equal(value, after[key]), f"{label}: {key} changed while frozen"
+
+
+def test_frozen_flow_stage_trainer_step_does_not_update_model_or_ema():
+    cfg = _base_cfg()
+    model_config = _model_config(cfg)
+    model = build_models(model_config)["stage1"]
+    assert model is not None
+    spec = StageSpec(name="stage1", is_stage2=False, generator="flow", freeze=True, ema_decay=0.999, steps_per_epoch=4)
+    trainer = FlowDDPMStageTrainer(spec, model, torch.device("cpu"))
+    assert trainer.ema_model is not None
+    model_before = _state_dict_clone(trainer.model)
+    ema_before = _state_dict_clone(trainer.ema_model)
+    for batch in _fake_batches(4, 8):
+        trainer.step(batch, torch.device("cpu"), global_step=1)
+    _assert_state_dicts_equal(model_before, trainer.model.state_dict(), "frozen flow model")
+    _assert_state_dicts_equal(ema_before, trainer.ema_model.state_dict(), "frozen flow ema")
+
+
+def test_frozen_wgan_stage_trainer_step_does_not_update_generator_or_critic():
+    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",
+        freeze=True,
+        n_critic=1,  # a generator step every batch, so a bug would surface immediately
+        ema_decay=0.999,
+        steps_per_epoch=4,
+    )
+    trainer = WGANStageTrainer(spec, models["stage1"], critics["stage1"], torch.device("cpu"))
+    assert trainer.ema_model is not None
+    model_before = _state_dict_clone(trainer.model)
+    critic_before = _state_dict_clone(trainer.critic)
+    ema_before = _state_dict_clone(trainer.ema_model)
+    for global_step, batch in enumerate(_fake_batches(4, 8)):
+        trainer.step(batch, torch.device("cpu"), global_step=global_step)
+    _assert_state_dicts_equal(model_before, trainer.model.state_dict(), "frozen wgan generator")
+    _assert_state_dicts_equal(critic_before, trainer.critic.state_dict(), "frozen wgan critic")
+    _assert_state_dicts_equal(ema_before, trainer.ema_model.state_dict(), "frozen wgan ema")
+
+
+@pytest.mark.parametrize("stage1_generator", ["flow", "wgan"])
+def test_train_end_to_end_frozen_stage1_unchanged_while_stage2_trains(stage1_generator):
+    cfg = _base_cfg()
+    cfg["stage1_model"]["generator"] = stage1_generator
+    cfg["stage1_model"]["freeze"] = True
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    assert models["stage1"] is not None and models["stage2"] is not None
+    stage1_before = _state_dict_clone(models["stage1"])
+    stage2_before = _state_dict_clone(models["stage2"])
+    with tempfile.TemporaryDirectory() as tmp:
+        train(
+            cfg=cfg,
+            models=models,
+            critics=critics,
+            train_loader=_fake_batches(4, cfg["train"]["batch_size"]),
+            val_loader=_fake_batches(2, cfg["train"]["batch_size"], seed=1),
+            device=torch.device("cpu"),
+            out_dir=Path(tmp) / "run",
+            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,
+        )
+    _assert_state_dicts_equal(stage1_before, models["stage1"].state_dict(), "frozen stage1")
+    stage2_after = models["stage2"].state_dict()
+    assert any(not torch.equal(v, stage2_after[k]) for k, v in stage2_before.items()), (
+        "unfrozen stage2 should have trained"
+    )
+
+
+def test_init_stages_from_checkpoints_loads_matching_stage_and_ema_weights(tmp_path):
+    cfg = _base_cfg()
+    model_config = _model_config(cfg)
+    source_models = build_models(model_config)
+    source_critics = build_critics(model_config)
+    source_trainers = build_stage_trainers(cfg, source_models, source_critics, torch.device("cpu"), 4)
+    source_stage1_ema = source_trainers["stage1"].ema_model
+    assert source_stage1_ema is not None
+    # Diverge the source's EMA from its raw weights so a same-vs-different
+    # check below actually distinguishes the two copy paths.
+    for p in source_stage1_ema.parameters():
+        p.data.add_(1.0)
+    ckpt_path = tmp_path / "source.pt"
+    ckpt = build_checkpoint(source_trainers, epoch=1, global_step=1, best_val_loss=0.0, extras={})
+    torch.save(ckpt, ckpt_path)
+
+    cfg2 = copy.deepcopy(cfg)
+    cfg2["stage1_model"]["init_from"] = str(ckpt_path)
+    dest_models = build_models(_model_config(cfg2))
+    dest_critics = build_critics(_model_config(cfg2))
+    dest_trainers = build_stage_trainers(cfg2, dest_models, dest_critics, torch.device("cpu"), 4)
+
+    loaded = init_stages_from_checkpoints(dest_trainers)
+    assert len(loaded) == 1 and "stage1" in loaded[0]
+    dest_stage1_ema = dest_trainers["stage1"].ema_model
+    assert dest_stage1_ema is not None
+
+    _assert_state_dicts_equal(
+        source_trainers["stage1"].model.state_dict(), dest_trainers["stage1"].model.state_dict(), "init_from raw"
+    )
+    _assert_state_dicts_equal(
+        source_stage1_ema.state_dict(),
+        dest_stage1_ema.state_dict(),
+        "init_from ema",
+    )
+    # stage2 has no init_from set -- untouched fresh init, not the source's.
+    stage2_matches_source = all(
+        torch.equal(v, dest_trainers["stage2"].model.state_dict()[k])
+        for k, v in source_trainers["stage2"].model.state_dict().items()
+    )
+    assert not stage2_matches_source
+
+
+def test_run_train_job_stage1_init_from_freeze_produces_rollout_capable_checkpoint(tmp_path):
+    """The exact scenario gitea #42 exists for: retrain stage 2 alone against
+    a fixed, known-good stage 1, and still get a checkpoint giant rollout can
+    load (checkpoint_io.load_for_inference with require_stage2=True)."""
+    normalizer_dict = {
+        "cond": Normalizer().fit(np.zeros((1, COND_DIM), dtype=np.float32)).to_dict(),
+        "target": Normalizer().fit(np.zeros((1, X_DIM), dtype=np.float32)).to_dict(),
+        "sec_phys": Normalizer().fit(np.zeros((1, 2), dtype=np.float32)).to_dict(),
+    }
+
+    cfg = _base_cfg()
+    source_out = tmp_path / "source"
+    _run_train(cfg, source_out, normalizer_dict=normalizer_dict)
+    source_ckpt = torch.load(source_out / "best.pt", weights_only=False)
+
+    cfg2 = copy.deepcopy(cfg)
+    cfg2["stage1_model"]["init_from"] = str(source_out / "best.pt")
+    cfg2["stage1_model"]["freeze"] = True
+    retrain_out = tmp_path / "retrain"
+    _run_train(cfg2, retrain_out, normalizer_dict=normalizer_dict)
+
+    ctx = load_for_inference(retrain_out / "best.pt", torch.device("cpu"), "rollout", require_stage2=True)
+    assert ctx.stage1 is not None and ctx.stage2 is not None
+
+    retrain_ckpt = torch.load(retrain_out / "best.pt", weights_only=False)
+    for key, value in source_ckpt["model"].items():
+        assert torch.equal(value, retrain_ckpt["model"][key]), f"frozen stage1 {key} drifted across the retrain"
+
+
 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
@@ -831,3 +1146,127 @@ def test_wgan_physical_omits_grad_norm_slice_columns():
         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
+
+
+# --- stage2_model.stage1_context = "sampled" (gitea #41) --------------------
+
+
+def _sampled_ctx_cfg(ema_decay=0.999):
+    cfg = _base_cfg()
+    cfg["stage1_model"]["generator"] = "flow"
+    cfg["stage2_model"]["generator"] = "flow"
+    cfg["stage2_model"]["stage1_context"] = "sampled"
+    cfg["stage2_model"]["ctx_p_start"] = 0.0
+    cfg["stage2_model"]["ctx_p_end"] = 0.0
+    cfg["train"]["ema_decay"] = ema_decay
+    return cfg
+
+
+def _build_sampled_trainers(cfg):
+    model_config = _model_config(cfg)
+    models = build_models(model_config)
+    critics = build_critics(model_config)
+    return build_stage_trainers(cfg, models, critics, torch.device("cpu"), total_train_batches=4)
+
+
+def test_build_stage_trainers_attaches_stage1_only_under_sampled():
+    trainers = _build_sampled_trainers(_sampled_ctx_cfg())
+    assert trainers["stage2"].stage1_source is trainers["stage1"]
+    assert trainers["stage1"].stage1_source is None
+
+
+def test_build_stage_trainers_leaves_stage1_source_none_under_truth():
+    """Regression guard for the old silent no-op: 'truth' (the default) must
+    never attach a stage1_source, so _stage1_context short-circuits without
+    ever calling sample_stage1."""
+    cfg = _base_cfg()
+    trainers = _build_sampled_trainers(cfg)
+    assert trainers["stage2"].stage1_source is None
+
+
+def test_stage1_context_sampled_calls_sample_stage1_and_differs_from_truth():
+    cfg = _sampled_ctx_cfg()
+    trainers = _build_sampled_trainers(cfg)
+    stage1, stage2 = trainers["stage1"], trainers["stage2"]
+    batch = _fake_batches(1, 4)[0]
+    cond_cont, cond_cat, x1_s1 = batch.cond_cont, batch.cond_cat, batch.target_s1
+
+    with patch("giant.training.trainers.sample_stage1", wraps=trainers_sample_stage1) as spy:
+        ctx = stage2._stage1_context(x1_s1, cond_cont, cond_cat, epoch=0)
+    assert spy.call_count == 1
+    assert spy.call_args.args[0] is stage1.sampling_model()
+    assert not torch.equal(ctx, x1_s1)
+
+
+def test_stage1_context_truth_default_never_calls_sample_stage1():
+    cfg = _base_cfg()
+    trainers = _build_sampled_trainers(cfg)
+    stage2 = trainers["stage2"]
+    batch = _fake_batches(1, 4)[0]
+    cond_cont, cond_cat, x1_s1 = batch.cond_cont, batch.cond_cat, batch.target_s1
+
+    with patch("giant.training.trainers.sample_stage1", wraps=trainers_sample_stage1) as spy:
+        ctx = stage2._stage1_context(x1_s1, cond_cont, cond_cat, epoch=0)
+    assert spy.call_count == 0
+    assert torch.equal(ctx, x1_s1)
+
+
+def test_stage1_context_val_epoch_none_uses_ground_truth_even_under_sampled():
+    cfg = _sampled_ctx_cfg()
+    trainers = _build_sampled_trainers(cfg)
+    stage2 = trainers["stage2"]
+    batch = _fake_batches(1, 4)[0]
+    cond_cont, cond_cat, x1_s1 = batch.cond_cont, batch.cond_cat, batch.target_s1
+
+    with patch("giant.training.trainers.sample_stage1", wraps=trainers_sample_stage1) as spy:
+        ctx = stage2._stage1_context(x1_s1, cond_cont, cond_cat, epoch=None)
+    assert spy.call_count == 0
+    assert torch.equal(ctx, x1_s1)
+
+
+def test_stage1_context_sampled_preserves_stage1_training_mode():
+    """Every sampler in giant/sample.py flips its model to .eval() as a side
+    effect with no restore of its own (see sample_flow). Sampling from the
+    RAW stage-1 model (ema_decay=0, so sampling_model() returns self.model,
+    the same weights the stage-1 trainer is actively training on) must not
+    silently leave it in eval mode for the rest of the epoch's stage-1
+    updates."""
+    cfg = _sampled_ctx_cfg(ema_decay=0.0)
+    trainers = _build_sampled_trainers(cfg)
+    stage1, stage2 = trainers["stage1"], trainers["stage2"]
+    stage1.train_mode()
+    assert stage1.model.training
+    batch = _fake_batches(1, 4)[0]
+    cond_cont, cond_cat, x1_s1 = batch.cond_cont, batch.cond_cat, batch.target_s1
+
+    stage2._stage1_context(x1_s1, cond_cont, cond_cat, epoch=0)
+    assert stage1.model.training
+
+
+@pytest.mark.parametrize("stage2_generator", ["wgan", "flow"])
+def test_build_stage_trainers_sampled_step_runs(stage2_generator):
+    """Both trainer subclasses' call sites (FlowDDPMStageTrainer._compute,
+    WGANStageTrainer.step) must run end to end under 'sampled' and produce a
+    finite loss."""
+    cfg = _sampled_ctx_cfg()
+    cfg["stage2_model"]["generator"] = stage2_generator
+    trainers = _build_sampled_trainers(cfg)
+    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])
+
+
+def test_train_end_to_end_stage1_context_sampled():
+    """Full train() run with stage1_context='sampled' must complete and
+    write a checkpoint + metrics.csv with finite losses throughout."""
+    cfg = _sampled_ctx_cfg()
+    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"]
+        assert all(math.isfinite(float(r["stage2/train/loss"])) for r in rows)

tests/test_wgan.py

@@ -2,7 +2,7 @@ import torch
 
 from giant.config import ConditioningAxisConfig
 from giant.constants import COND_DIM, K_MAX, SEC_DIM, SEC_SLOT_DIM, X_DIM
-from giant.model.network import CriticModel, Stage1Model, Stage2OneShot
+from giant.model.network import CriticModel, LinearTrunk, Stage1Model, Stage2OneShot
 from giant.model.wgan import critic_loss, generator_loss, gradient_penalty
 from giant.sample import sample_secondaries_wgan, sample_wgan
 
@@ -115,6 +115,33 @@ def test_critic_output_shape():
     assert out.shape == (B,)
 
 
+def test_critic_model_honours_trunk_type_and_block_conditioning():
+    """gitea #57: CriticModel routes its body through build_trunk/build_block
+    like every generator stage model, instead of hand-rolling a plain
+    ResBlock stack."""
+    B = 8
+    critic = CriticModel(
+        pdg_vocab=3,
+        mat_vocab=2,
+        particle_cfg=PARTICLE_CFG,
+        material_cfg=MATERIAL_CFG,
+        in_dim=X_DIM,
+        hidden_dim=32,
+        n_res_blocks=2,
+        stage="stage1",
+        trunk_type="linear",
+        block_conditioning="adaln",
+    )
+    assert isinstance(critic.trunk, LinearTrunk)
+    cond_cont, cond_cat = _cond(B)
+    real = torch.randn(B, X_DIM)
+    fake = torch.randn(B, X_DIM)
+    loss = critic_loss(lambda x: critic(x, cond_cont, cond_cat), real, fake.detach(), gp_weight=10.0)
+    loss.backward()
+    for name, p in critic.named_parameters():
+        assert p.grad is not None, f"no grad for {name}"
+
+
 def test_sample_wgan_shape():
     B = 6
     model = _small_generator()

uv.lock

@@ -5,19 +5,19 @@ resolution-markers = [
     "python_full_version >= '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
     "python_full_version >= '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
     "python_full_version >= '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
-    "python_full_version < '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
-    "python_full_version < '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
-    "python_full_version < '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
     "python_full_version >= '3.14' and sys_platform == 'win32' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version >= '3.14' and sys_platform == 'emscripten' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version >= '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
+    "python_full_version >= '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
+    "python_full_version >= '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
+    "python_full_version >= '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
+    "python_full_version < '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
+    "python_full_version < '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
+    "python_full_version < '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda'",
     "python_full_version < '3.14' and sys_platform == 'win32' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version < '3.14' and sys_platform == 'emscripten' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "(python_full_version == '3.13.*' and platform_machine == 'aarch64' and platform_python_implementation == 'CPython' and sys_platform == 'linux' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda') or (python_full_version < '3.14' and platform_machine != 'aarch64' and sys_platform == 'linux' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda') or (python_full_version < '3.14' and platform_python_implementation != 'CPython' and sys_platform == 'linux' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda') or (python_full_version < '3.14' and sys_platform != 'darwin' and sys_platform != 'emscripten' and sys_platform != 'linux' and sys_platform != 'win32' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda')",
     "(python_full_version < '3.13' and platform_machine == 'aarch64' and platform_python_implementation == 'CPython' and sys_platform == 'linux' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda') or (python_full_version < '3.14' and sys_platform == 'darwin' and extra == 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda')",
-    "python_full_version >= '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
-    "python_full_version >= '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
-    "python_full_version >= '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version < '3.14' and sys_platform == 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version < '3.14' and sys_platform == 'emscripten' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
     "python_full_version < '3.14' and sys_platform != 'emscripten' and sys_platform != 'win32' and extra != 'extra-5-giant-cpu' and extra != 'extra-5-giant-cuda'",
@@ -45,6 +45,19 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/99/91/8acff4f5e50511b911bbccb72b8628a49c68ce14148cd9f6431094859a90/annotated_types-0.8.0-py3-none-any.whl", hash = "sha256:f072f4d804ea359e4eaf198b1af7a8b0943881a87f31bb764f8bf219bb9419e0", size = 13427, upload-time = "2026-07-23T20:16:12.938Z" },
 ]
 
+[[package]]
+name = "anyio"
+version = "4.14.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "idna", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "typing-extensions", marker = "(python_full_version < '3.13' and sys_platform != 'emscripten') or (python_full_version >= '3.13' and extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda') or (sys_platform == 'emscripten' and extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/61/cc/a381afa6efea9f496eff839d4a6a1aed3bfafc7b3ab4b0d1b243a12573dd/anyio-4.14.2.tar.gz", hash = "sha256:cfa139f3ed1a23ee8f88a145ddb5ac7605b8bbfd8592baacd7ce3d8bb4313c7f", size = 260176, upload-time = "2026-07-12T20:29:07.082Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/da/35/f2287558c17e29fafc8ef3daf819bb9834061cfa43bff8014f7df7f63bdc/anyio-4.14.2-py3-none-any.whl", hash = "sha256:9f505dda5ac9f0c8309b5e8bd445a8c2bf7246f3ce950121e45ea15bc41d1494", size = 125813, upload-time = "2026-07-12T20:29:05.763Z" },
+]
+
 [[package]]
 name = "appnope"
 version = "0.1.4"
@@ -134,6 +147,35 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/44/a1/70ebfffd6c6edc6034a547838ee46287c65ed89f710592ddc39c76b4a5a8/awkward_cpp-53-cp314-cp314t-win_arm64.whl", hash = "sha256:1be0c1d87d9f4fdf94b767a061df849f1bb21579d302b2996fb101527fc80a97", size = 551257, upload-time = "2026-06-08T12:31:56.319Z" },
 ]
 
+[[package]]
+name = "bracex"
+version = "3.0.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ac/01/5f394b8bcd6e5b92f73130990960423bbb19711f906bd9fe9ea5557c667c/bracex-3.0.1.tar.gz", hash = "sha256:4e38e32392e4a4780fe15d644bfc7c8514057cfc3861e060b11814ce829c25e4", size = 44019, upload-time = "2026-07-20T13:43:00.335Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b8/8f/6f7273a7adb8d73fc8d21ede4376a3e475e52f98435c6007f69100dec8ca/bracex-3.0.1-py3-none-any.whl", hash = "sha256:6523ad83aeb5098a4ee597cff0f964442ff74e460bd3fafaffab6a013ff2288c", size = 11940, upload-time = "2026-07-20T13:42:59.268Z" },
+]
+
+[[package]]
+name = "bump-my-version"
+version = "1.5.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "httpx2" },
+    { name = "pydantic" },
+    { name = "pydantic-settings" },
+    { name = "questionary" },
+    { name = "rich" },
+    { name = "rich-click" },
+    { name = "tomlkit" },
+    { name = "wcmatch" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/23/09/5b09ac74962eca809cbf7010a08ea6ad405852bdd53489209a9f473d775c/bump_my_version-1.5.1.tar.gz", hash = "sha256:5079e443ab8c9a9903f140b427ff9f6fe8dd54013a55a4cf48b89326f3a71c07", size = 1132060, upload-time = "2026-08-06T14:26:38.9Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b3/0b/5885530f79d4400368b9d4dcb9b39274c0d52e633f7871e7fc6feceea1e3/bump_my_version-1.5.1-py3-none-any.whl", hash = "sha256:df3e2989d0d7fe704718feb24a5880f089b6b6369e427a4445b89c3adebfcff1", size = 65090, upload-time = "2026-08-06T14:26:37.083Z" },
+]
+
 [[package]]
 name = "certifi"
 version = "2026.7.22"
@@ -633,7 +675,7 @@ wheels = [
 
 [[package]]
 name = "giant"
-version = "0.3.2"
+version = "0.3.6"
 source = { editable = "." }
 dependencies = [
     { name = "numpy" },
@@ -666,6 +708,8 @@ cuda = [
 ]
 dev = [
     { name = "awkward" },
+    { name = "bump-my-version" },
+    { name = "git-cliff" },
     { name = "ipykernel" },
     { name = "matplotlib" },
     { name = "plotstyle" },
@@ -688,7 +732,9 @@ wandb = [
 [package.metadata]
 requires-dist = [
     { name = "awkward", marker = "extra == 'convert'", specifier = ">=2.6,<3" },
+    { name = "bump-my-version", marker = "extra == 'dev'", specifier = ">=1.2,<2" },
     { name = "giant", extras = ["convert", "analysis", "geometry", "wandb"], marker = "extra == 'dev'" },
+    { name = "git-cliff", marker = "extra == 'dev'", specifier = ">=2,<3" },
     { name = "ipykernel", marker = "extra == 'analysis'", specifier = ">=7.3.0" },
     { name = "matplotlib", marker = "extra == 'analysis'", specifier = ">=3.8,<4" },
     { name = "numpy", specifier = ">=1.26,<3" },
@@ -713,6 +759,35 @@ requires-dist = [
 ]
 provides-extras = ["cpu", "cuda", "dev", "geometry", "wandb", "convert", "analysis"]
 
+[[package]]
+name = "git-cliff"
+version = "2.13.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/62/57/b12494e2cbc3c9154c942e64659b5aec2b1ce9f12d07f6dc6167e2c63ae5/git_cliff-2.13.1.tar.gz", hash = "sha256:e949ea9c3951ba6037b99eec465162be2584f27f0836ace45f44d6f45650f8c6", size = 113119, upload-time = "2026-04-26T10:33:42.331Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8f/dd/24768c3c0030710d36706c17b997d06aee27cb76b27ab2abb058ae254175/git_cliff-2.13.1-py3-none-macosx_10_12_x86_64.whl", hash = "sha256:08a9cb0ec760e165210ed22fefa295b6549a3520b420db995ccbb3620cbb1fbe", size = 7260035, upload-time = "2026-04-26T10:33:16.269Z" },
+    { url = "https://files.pythonhosted.org/packages/22/df/842973ead79d27a58cd1eccd167191c0a71513e5c1e9dc30337dafdb7d36/git_cliff-2.13.1-py3-none-macosx_11_0_arm64.whl", hash = "sha256:e92cf470ecbe73f7d2963dfa80e8961a6b76888d0c19949b0f028a28a0a0470c", size = 6854384, upload-time = "2026-04-26T10:33:18.808Z" },
+    { url = "https://files.pythonhosted.org/packages/7c/4d/6d6efa7d61be8632563990ccd402e401695e4a85a0bb1002f28730d03268/git_cliff-2.13.1-py3-none-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2e8d8e420adf6a36b97e0fbdbf2b07e47712199a9baf3675c9a759430243ea26", size = 7308164, upload-time = "2026-04-26T10:33:20.869Z" },
+    { url = "https://files.pythonhosted.org/packages/f0/4a/98b8d2f53a2d0b7d313e98ab363ad7e0d6a514e878a8d678f22402cb0ec7/git_cliff-2.13.1-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1ab059d671565189faa4f3858b2fb42535aa39fe265ad95d84d5c116a2724fc7", size = 7687163, upload-time = "2026-04-26T10:33:23.096Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/07/cdd149b3909644aa3f0be7960406d9bbb38598f8618e039ac48cbb43ded6/git_cliff-2.13.1-py3-none-manylinux_2_28_aarch64.whl", hash = "sha256:a93db30da45967c42df607fbbc2092111fcd576b0ca9e2fbddd3f653d8c71be7", size = 7317670, upload-time = "2026-04-26T10:33:25.323Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/11/6d377a7f3113f6e26d87a28a32013eb05bd62bce176d6f8ee808e4868c1f/git_cliff-2.13.1-py3-none-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:df9f5a2bd16e5225030c9c2362e6ac70b34a906c0fbb83f8cbf5ae46932ba0d2", size = 7502294, upload-time = "2026-04-26T10:33:27.315Z" },
+    { url = "https://files.pythonhosted.org/packages/80/6b/e1da9acf3aec99e6600be02b6ce0c9e8bd42d072e3120384514c905231e2/git_cliff-2.13.1-py3-none-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:c12276784d280aa6a7148d3e52ff139e891f4c97720cd9be581b19892ea39fe0", size = 7927258, upload-time = "2026-04-26T10:33:29.554Z" },
+    { url = "https://files.pythonhosted.org/packages/ed/ea/9f2188a5e474e5f02193d9c1cdf7028773d483a3f508a1df4f1d93c9cc80/git_cliff-2.13.1-py3-none-musllinux_1_2_aarch64.whl", hash = "sha256:17da93dc605cbc48c762770402067fa726437cedbd62514f9caef6e0ccb58a43", size = 7308153, upload-time = "2026-04-26T10:33:31.589Z" },
+    { url = "https://files.pythonhosted.org/packages/b3/70/5e2b2a0e42c07956f911e1eccd6dc6d79b96fc6c8a604a526c1ee0474c84/git_cliff-2.13.1-py3-none-musllinux_1_2_i686.whl", hash = "sha256:cd4a08cf3f638ec71d2ed451aa8673bef99e1107a36366153db97ca18d981655", size = 7502287, upload-time = "2026-04-26T10:33:33.894Z" },
+    { url = "https://files.pythonhosted.org/packages/05/50/cc4c1d3d360621c0235d66d2c74472d9993d8aadf23ffa311eaf29d7a3aa/git_cliff-2.13.1-py3-none-musllinux_1_2_x86_64.whl", hash = "sha256:c5bb87f6e1db18be09e50c9d59234dc949713e20c2700e94eca378d22ad79719", size = 7927253, upload-time = "2026-04-26T10:33:36.205Z" },
+    { url = "https://files.pythonhosted.org/packages/86/5d/717d30f37dad65a6cc5220b04a3ef1bc44c31fb56d0ce2895114e159df4f/git_cliff-2.13.1-py3-none-win32.whl", hash = "sha256:c8878972e0a6c26d9137fc406a611116239333578d95ac05064d2807920bd83c", size = 6718261, upload-time = "2026-04-26T10:33:38.1Z" },
+    { url = "https://files.pythonhosted.org/packages/99/b2/99fac50978b9a90bfec0f1b89354a667ec83f4990301f6c708abce05484e/git_cliff-2.13.1-py3-none-win_amd64.whl", hash = "sha256:856d831a0bede9c258229dbd4d4c2b1c0810d8fce3d3882729669e8dc09c72bf", size = 7714969, upload-time = "2026-04-26T10:33:40.163Z" },
+]
+
+[[package]]
+name = "h11"
+version = "0.16.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/01/ee/02a2c011bdab74c6fb3c75474d40b3052059d95df7e73351460c8588d963/h11-0.16.0.tar.gz", hash = "sha256:4e35b956cf45792e4caa5885e69fba00bdbc6ffafbfa020300e549b208ee5ff1", size = 101250, upload-time = "2025-04-24T03:35:25.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/04/4b/29cac41a4d98d144bf5f6d33995617b185d14b22401f75ca86f384e87ff1/h11-0.16.0-py3-none-any.whl", hash = "sha256:63cf8bbe7522de3bf65932fda1d9c2772064ffb3dae62d55932da54b31cb6c86", size = 37515, upload-time = "2025-04-24T03:35:24.344Z" },
+]
+
 [[package]]
 name = "hepunits"
 version = "2.4.6"
@@ -722,6 +797,45 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/85/10/7f9c58d1ec6a0b7f7783fe552f3593f39cda30c2e1d7a9d148ae711e748d/hepunits-2.4.6-py3-none-any.whl", hash = "sha256:089c52c3b84ef67a159b5e9ee9bdd50e1a442e3fd0c101303cc409c1e9011c4d", size = 17090, upload-time = "2026-06-16T09:23:35.35Z" },
 ]
 
+[[package]]
+name = "httpcore2"
+version = "2.11.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "h11", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "truststore", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/30/7d/ee6787efd5fe675d7cfd5eb149e40ccb5bdfc7e7c9252edcf7825c38986f/httpcore2-2.11.0.tar.gz", hash = "sha256:82e6fc95d784e6ee22ebd4b2cb57df53a2efb13ad6a11260a236ecebbc5f50c7", size = 67532, upload-time = "2026-08-18T08:03:53.008Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f3/54/e84a5c82ac0959d5e55b3970d326fd95306446b2bf0702302888745f7e5c/httpcore2-2.11.0-py3-none-any.whl", hash = "sha256:c7c899fbc6b8abb6e747dda427aa6f52934c45e191eabb986965164ebb02a908", size = 83061, upload-time = "2026-08-18T08:03:50.894Z" },
+]
+
+[[package]]
+name = "httpx2"
+version = "2.11.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "httpcore2", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "httpx2-jsfetch", marker = "sys_platform == 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "idna" },
+    { name = "truststore", marker = "sys_platform != 'emscripten' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "typing-extensions", marker = "python_full_version < '3.13' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/d5/4d/b3fcae38f29bfb0f300517d085c488f41f65e5b0a73023976b2122f568cc/httpx2-2.11.0.tar.gz", hash = "sha256:ea01b2e8febfb026e2601814c77ecb1e64fff114a87bc789cb520e67f27e7809", size = 99617, upload-time = "2026-08-18T08:03:53.691Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/39/29/f60bcf54028601920c0ce3da81537bcd0af4ee19dd9672af85dc9dfd60e0/httpx2-2.11.0-py3-none-any.whl", hash = "sha256:c9790f62a327110f52a099f1e2030cbe32f78b28781ad68fe58bfd6f23e73ab0", size = 95043, upload-time = "2026-08-18T08:03:52.004Z" },
+]
+
+[[package]]
+name = "httpx2-jsfetch"
+version = "1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/c4/0e5636363151a2a1795e0a77617168b9ca438e1748ec05fc9b5687f93d64/httpx2_jsfetch-1.0.tar.gz", hash = "sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60", size = 6872, upload-time = "2026-08-07T00:13:07.492Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9b/43/832f631d32e4f1211caa2ba368317739fe71f0b8530e4c9d15dc454bac2a/httpx2_jsfetch-1.0-py3-none-any.whl", hash = "sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32", size = 6382, upload-time = "2026-08-07T00:13:06.567Z" },
+]
+
 [[package]]
 name = "idna"
 version = "3.18"
@@ -1780,6 +1894,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/fa/c3/7c8b240552251faf6b3a957db200fcfbbcec36763c050428b601e0c9b83b/pydantic_core-2.46.4-graalpy312-graalpy250_312_native-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:00c603d540afdd6b80eb39f078f33ebd46211f02f33e34a32d9f053bba711de0", size = 2147590, upload-time = "2026-05-06T13:39:29.883Z" },
 ]
 
+[[package]]
+name = "pydantic-settings"
+version = "2.15.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pydantic" },
+    { name = "python-dotenv" },
+    { name = "typing-inspection" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/68/ca/31c57507b13119d7d3cfa1576dad2911a4861e3be07b579395f4e9d393f9/pydantic_settings-2.15.0.tar.gz", hash = "sha256:694b793e84f766ba76a90ebdefc01d0a9a045dab0382bee70393da93712ad117", size = 261253, upload-time = "2026-08-07T09:24:57.419Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/30/a4/2bffa9f8e804325a09867f0e9d30795c80ea9f8d62560bd1b6ad6220eb2f/pydantic_settings-2.15.0-py3-none-any.whl", hash = "sha256:0ba092c291c94baceb5eff768aa0d56400a457585bc0175925a5a5510303da42", size = 69413, upload-time = "2026-08-07T09:24:55.839Z" },
+]
+
 [[package]]
 name = "pygments"
 version = "2.20.0"
@@ -1840,6 +1968,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" },
 ]
 
+[[package]]
+name = "python-dotenv"
+version = "1.2.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6a/53/ed9d74092561d4b01a2ef1349d52cdbc135e526c245f366b089cfca6de49/python_dotenv-1.2.3.tar.gz", hash = "sha256:a20a594dabeaa385725aa239d5244871c143ecb356add8a20fcf23773a6c3a35", size = 58945, upload-time = "2026-08-16T16:54:54.067Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0d/17/c5c6b53ddc18f297992099b3d9ec16c855c0ccc83263a21fe4d1c625ec6c/python_dotenv-1.2.3-py3-none-any.whl", hash = "sha256:904552145e8bfed22162c09dab1c2b9b54fefa7b23ba780f4f26ca0316b0f0d9", size = 22780, upload-time = "2026-08-16T16:54:52.473Z" },
+]
+
 [[package]]
 name = "pyyaml"
 version = "6.0.3"
@@ -1929,6 +2066,18 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/81/d6/4bfbb40c9a0b42fc53c7cf442f6385db70b40f74a783130c5d0a5aa62228/pyzmq-27.1.0-cp314-cp314t-win_arm64.whl", hash = "sha256:dc5dbf68a7857b59473f7df42650c621d7e8923fb03fa74a526890f4d33cc4d7", size = 575170, upload-time = "2025-09-08T23:09:01.418Z" },
 ]
 
+[[package]]
+name = "questionary"
+version = "2.1.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "prompt-toolkit" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f6/45/eafb0bba0f9988f6a2520f9ca2df2c82ddfa8d67c95d6625452e97b204a5/questionary-2.1.1.tar.gz", hash = "sha256:3d7e980292bb0107abaa79c68dd3eee3c561b83a0f89ae482860b181c8bd412d", size = 25845, upload-time = "2025-08-28T19:00:20.851Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3c/26/1062c7ec1b053db9e499b4d2d5bc231743201b74051c973dadeac80a8f43/questionary-2.1.1-py3-none-any.whl", hash = "sha256:a51af13f345f1cdea62347589fbb6df3b290306ab8930713bfae4d475a7d4a59", size = 36753, upload-time = "2025-08-28T19:00:19.56Z" },
+]
+
 [[package]]
 name = "requests"
 version = "2.34.2"
@@ -1957,6 +2106,20 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/82/3b/64d4899d73f91ba49a8c18a8ff3f0ea8f1c1d75481760df8c68ef5235bf5/rich-15.0.0-py3-none-any.whl", hash = "sha256:33bd4ef74232fb73fe9279a257718407f169c09b78a87ad3d296f548e27de0bb", size = 310654, upload-time = "2026-04-12T08:24:02.83Z" },
 ]
 
+[[package]]
+name = "rich-click"
+version = "1.9.8"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click" },
+    { name = "colorama", marker = "sys_platform == 'win32' or (extra == 'extra-5-giant-cpu' and extra == 'extra-5-giant-cuda')" },
+    { name = "rich" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f7/ea/21e4867ea0ef881ffd4c0550fc21a061435e50d6324bcd034396633cbc18/rich_click-1.9.8.tar.gz", hash = "sha256:4008f921da88b5d91646c134ec881c1500e5a6b3f093e90e8f29400e09608371", size = 75363, upload-time = "2026-05-28T19:54:59.144Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6d/97/a87901aef6b7e7e4a34c6dd6cc17dca8594a592ef9d9dd765fca2b7facf7/rich_click-1.9.8-py3-none-any.whl", hash = "sha256:12873865396e6927835d4eabb1cc3996edcd65b7ac9b2391a29eca4f335a2f93", size = 72189, upload-time = "2026-05-28T19:54:57.867Z" },
+]
+
 [[package]]
 name = "ruff"
 version = "0.15.17"
@@ -2149,6 +2312,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/32/d5/f9a850d79b0851d1d4ef6456097579a9005b31fea68726a4ae5f2d82ddd9/threadpoolctl-3.6.0-py3-none-any.whl", hash = "sha256:43a0b8fd5a2928500110039e43a5eed8480b918967083ea48dc3ab9f13c4a7fb", size = 18638, upload-time = "2025-03-13T13:49:21.846Z" },
 ]
 
+[[package]]
+name = "tomlkit"
+version = "0.15.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/94/96/e07752635b98536177fa1f37671c8f3cdde2e724c6bcf6034b2cfb571565/tomlkit-0.15.1.tar.gz", hash = "sha256:e25bbf38843005246210a12982776f27f99cb9be67160e14434d0c0d21ee1e97", size = 180129, upload-time = "2026-07-17T01:48:04.562Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/13/bc/8c13eb66537dce1d2bd3a57132902f38d0e7f5bb46fa9f4daed9fe9d76ee/tomlkit-0.15.1-py3-none-any.whl", hash = "sha256:177a05aece5a8ca5266fd3c448abb47b8d352f09d477d3ca8332db4d89b24304", size = 49449, upload-time = "2026-07-17T01:48:05.728Z" },
+]
+
 [[package]]
 name = "torch"
 version = "2.3.1"
@@ -2270,6 +2442,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/96/8d/1080ee4c231f361b6ce4470d556c8c435b67c7e0753aaa641497ee92f88b/traitlets-5.15.1-py3-none-any.whl", hash = "sha256:770a53705f84b81ac107e83a1b3328ff2dae16094d8fc3cfc004e4b22dfd8e92", size = 85858, upload-time = "2026-06-03T12:26:04.395Z" },
 ]
 
+[[package]]
+name = "truststore"
+version = "0.10.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/a3/1585216310e344e8102c22482f6060c7a6ea0322b63e026372e6dcefcfd6/truststore-0.10.4.tar.gz", hash = "sha256:9d91bd436463ad5e4ee4aba766628dd6cd7010cf3e2461756b3303710eebc301", size = 26169, upload-time = "2025-08-12T18:49:02.73Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/19/97/56608b2249fe206a67cd573bc93cd9896e1efb9e98bce9c163bcdc704b88/truststore-0.10.4-py3-none-any.whl", hash = "sha256:adaeaecf1cbb5f4de3b1959b42d41f6fab57b2b1666adb59e89cb0b53361d981", size = 18660, upload-time = "2025-08-12T18:49:01.46Z" },
+]
+
 [[package]]
 name = "ty"
 version = "0.0.50"
@@ -2394,6 +2575,18 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/07/78/75b6827a6665337a715c5347c5edbd84eca660f7a0f48d8d6d24d1f66bee/wandb-0.28.1-py3-none-win_arm64.whl", hash = "sha256:4aa07f13dd3bcac2c0524c8d0f49f76e83ab5c1054fd09f3b1a436cfcde146a6", size = 22299006, upload-time = "2026-07-16T18:47:02.71Z" },
 ]
 
+[[package]]
+name = "wcmatch"
+version = "11.0.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "bracex" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/57/43/30e407989e313677dbb9d5f045f966549a7254834571e342eaa4b55cc67b/wcmatch-11.0.1.tar.gz", hash = "sha256:1ea2b4fa678b8ca268253798d5963935df39132d47c3e241c0a0732224005e7d", size = 144662, upload-time = "2026-08-14T15:20:40.477Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ce/77/7a02b0f05b3ffcdbef9719ce3ee0b508d6a29b58e95299f1580055671db3/wcmatch-11.0.1-py3-none-any.whl", hash = "sha256:fd149ecddb9f0a88ea780017d6dde17c994e494e7f7303d4e3c9d6251f978f4b", size = 43449, upload-time = "2026-08-14T15:20:39.379Z" },
+]
+
 [[package]]
 name = "wcwidth"
 version = "0.8.1"