[0z5a]

Changed weights and nonzero gradients can reflect Adam decay or old momentum without a new useful update. This adds a fresh DCP state comparison and an FP32 AdamW counterfactual check, joined with both runs’ raw reward/advantage/gradient evidence. It compares rollout counters and all stored checkpoint state except explicitly excluded launch arguments; cached chunk fingerprints are not reused.

Stacked on #18 (16def6e0), with no production runtime change. Validation: 155 CPU checks pass, including actual AdamW/DCP controls at LR 1e-3 and 1e-6, six state corruptions and combined CLI checks. Fresh reads also match 17 chunks and counters from the retained earlier tiny Ray/PPO/SGD run; that checkpoint is explicitly rejected for Adam qualification.

The result table and scope retain all raw receipts and SHA256 manifests. This does not execute a new official model or fresh Ray worker, prove held-out quality, or measure speed. Older/fused optimizer formats without explicit stored AdamW identity still need actual-worker identity and layout qualification.