[0z5a]

Native checkpoint equality does not show whether resume used new Ray workers or the same loaded implementation. Add opt-in startup records for the native learner and rollout engine, including worker/process IDs, source hashes, actual optimizer configuration and the rollout runtime epoch. The qualification launcher enables a separate directory for each phase; ordinary runs remain disabled by default.

The new audit reads the phase receipts and raw rollout dumps, rejects reused workers and mismatched sources/contracts, and binds each rollout to its recorded epoch. Loader iteration is recorded separately from completed rollout count. This does not relax the stored AdamW gate or establish frozen-source, weight or held-out equivalence.

Stacked on #19 with test/native-checkpoint-recovery as the base. One incremental commit, authored and signed off by 0z5a.

Validation	Result
Focused CPU regression	217 passed; 28 new cases
New capture checks	Real PyTorch models/Adam/AdamW, explicit Ray/transport doubles
Resume evidence checks	Six synthetic worker records; 12 corruption cases rejected
Existing MCore config/chain handling	3 passed; no environment changes
Actual Ray/CUDA E2E, speedup, peak-memory reduction, reward convergence	Not run / not established

Results and the speed-comparison table: native_runtime_identity.md. Raw logs/XML, validation receipt and hashes are in benchmarks/results/native-runtime-identity/. The full framework import still needs Triton in an admitted existing runtime; these CPU controls are not model-training results.