[0z5a]

Checkpoint equality does not establish that a resumed worker generated the same samples or evaluated every held-out question. Add an output audit that reopens both runs, binds evaluation to the recorded worker and publication, re-decodes saved tokens, regrades rewards, and compares training tensors plus all four held-out rounds.

The frozen qualification covers 96 training and 32 held-out samples per run. Tokens and semantic traces must match; numeric comparisons use the predeclared FP32 tolerance and report maximum differences. Identical outputs with no useful learning signal do not qualify. This change leaves training and serving behavior unchanged.

Stacked on #20 with feature/native-runtime-identity as the base. One incremental commit, authored and signed off by 0z5a.

Validation	Result
CPU regression	245 passed; 20 new audit cases
Actual tiny Ouro / Nanbeige / Huginn sampling and Adam recovery	8 paired held-out outputs per family; exact tokens/rewards/traces, zero score difference
Official retained Ouro tokenizer	40 answer-replay controls; six config/tokenizer files match the pinned manifest
Full official Ray/CUDA E2E, speedup, memory reduction, reward convergence	Not run / not established

The tiny model tests restore fresh objects in one process and use tokenizer/transport doubles. CLI tests use synthetic worker/process records; official tokenizer controls use known answers and invented scores. These are not fresh Ray-worker or reasoning-convergence results. Weight/checkpoint/optimizer state and frozen-source membership remain separate gates.

Results and speed-comparison table, with raw logs/XML, tokenizer controls and hashes in benchmarks/results/native-output-recovery/.