feat: add WGAN + AR stop-token config variant

Branches off configs/baseline.toml with both stages on WGAN-GP and
stage-2 n_sec.mode = stop_token, 30 epochs — combines two unbenchmarked
roadmap axes (post-v0.3.0 WGAN, and the AR stop-token multiplicity mode)
into one variant that stays a single edit away from baseline for
attribution.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KzPghrmFcAJYrWUvHApY9N

configs/wgan_ar_stoptoken.toml

@@ -0,0 +1,122 @@
+# GIANT WGAN-GP + AR stop-token variant of configs/baseline.toml.
+#
+# Two roadmap axes, combined into one run: WGAN-GP generators for both
+# stages (unbenchmarked since the 2026-08-03 pre-v0.3.0 failure, which was
+# secondary-species mode collapse — the failure v0.3.0's AR/categorical
+# pivot exists to fix) and the AR stop-token multiplicity mode
+# (stage2_model.n_sec.mode = "stop_token", never benchmarked at all).
+# Everything else is byte-identical to baseline.toml so a rollout compared
+# against baseline's analysis_341dfb14 is attributable to these two axes
+# alone: conditioning (physical/physical), hidden_dim 512 / n_res_blocks 6 /
+# dropout 0.0 per stage, k_max 15, history "markov", teacher_forcing
+# "always", particle_type.target "onehot" (n_classes 32, other_policy
+# "sample"), lr 3e-4, warmup_epochs 3, weight_decay 0.01, ema_decay 0.9999,
+# val_fraction 0.1, num_workers 4, seed 0, validate_steps 10, W&B on.
+#
+# No [stage1_model.wgan] / [stage2_model.wgan] block: the dataclass defaults
+# (noise_dim 64, n_critic 5, gp_weight 10.0, critic_lr 0.0 = inherit
+# train.lr, critic_hidden_dim/critic_n_res_blocks 0 = inherit the stage's
+# 512/6, stage 2's gumbel_tau_start/_end 1.0/0.1) are what the earlier WGAN
+# runs used — writing them out would add keys that don't vary.
+#
+# particle_type.class_weighting stays "none" (the default): config.py's
+# validate_config rejects any other value under stage2_model.generator =
+# "wgan", since that path feeds the type slice to the critic via a
+# straight-through Gumbel relaxation instead of a weighted cross-entropy.
+#
+# Prior WGAN writeup (pre-v0.3.0, describes the failure this run re-tests):
+# /home/lars/knowledge-base/experiments/giant-wgan-physical-rollout-validation.md
+
+[meta]
+# REQUIRED. Without it config.migrate_config reads this file as v0.2 and
+# rewrites it from V02_FIXED_FACTS — silently forcing decoder = "one_shot",
+# particle_type.target = "physical" and the v0.2 default sizes, while still
+# passing validate_config.
+config_version = 3
+
+[conditioning]
+# Physical-property MLPs rather than learned vocab embeddings: computable for
+# any PDG code / material, which is what the held-out-species and
+# held-out-material generalization comparisons need.
+out_dim      = 128
+share_stages = false
+
+# n_layers = 2 rather than the v0.3 default of 1: v0.2's conditioning MLP was
+# always 2 deep (see _migration.V02_FIXED_FACTS), so this keeps the encoder
+# identical to baseline.toml.
+[conditioning.particle]
+type     = "physical"
+emb_dim  = 16
+n_layers = 2
+
+[conditioning.material]
+type     = "physical"
+emb_dim  = 16
+n_layers = 2
+
+[stage1_model]
+generator    = "wgan"
+hidden_dim   = 512
+n_res_blocks = 6
+dropout      = 0.0
+
+[stage2_model]
+# Autoregressive in descending-energy order, as baseline.toml — this variant
+# only swaps the generator (flow -> wgan) and the multiplicity mode
+# (head -> stop_token), not the decoder shape.
+decoder      = "autoregressive"
+generator    = "wgan"
+hidden_dim   = 512
+n_res_blocks = 6
+dropout      = 0.0
+k_max        = 15
+
+[stage2_model.autoregressive]
+history         = "markov"
+teacher_forcing = "always"
+
+[stage2_model.n_sec]
+# EOS-style per-slot stop head on the AR secondary decoder, replacing the
+# n_sec classifier entirely (mutually exclusive — see NSecConfig's
+# docstring in giant/config.py). Requires decoder = "autoregressive" and
+# owner = "stage2" (both already true above/by default); validate_config
+# enforces this.
+mode = "stop_token"
+
+[stage2_model.particle_type]
+target = "onehot"
+# Decoupled from conditioning.particle.emb_dim (gitea #29). 32 classes + the
+# "other" bucket keeps essentially all real secondary species out of "other"
+# without making the head expensive.
+n_classes    = 32
+other_policy = "sample"
+
+[train]
+epochs = 30
+# Halved from baseline's 36864. That figure came from a measured linear fit
+# of the *flow-AR* training step (peak reserved MiB = 0.9736 * batch_size +
+# 115); WGAN invalidates it twice over — each stage gains a critic that by
+# default inherits the stage's own 512/6 body, and gradient_penalty
+# (giant/model/wgan.py, forced fp32 internally) runs a double-backward every
+# batch. 18432 is a conservative choice pending a real memory measurement on
+# this exact config, not a re-derived fit. Throughput is already flat above
+# bs~4096 on the 4070, so this costs occupancy on the L40S, not step
+# efficiency.
+batch_size    = 18432
+lr            = 3e-4
+warmup_epochs = 3
+weight_decay  = 0.01
+ema_decay     = 0.9999
+val_fraction  = 0.1
+num_workers   = 4
+seed          = 0
+# Tightened from baseline's 10: WGANStageTrainer.supports_val_loss = False,
+# and with both stages adversarial there is no per-epoch val loss at all, so
+# validate_every's marginal-KL pass (giant/training/trainers.py's
+# val_objective) is the only comparable-across-epochs best-checkpoint
+# selection signal available. 5 gives 6 evaluations over 30 epochs instead
+# of baseline's 3, at ~6x5000s of extra walltime.
+validate_every = 5
+validate_steps = 10
+wandb          = true
+wandb_project  = "giant"