[0z5a]

Independent source, checkpoint and output audits could pass without proving that generated samples used the retained checkpoint tensors. This change binds the seven initial model/tokenizer files to each phase and worker, recomputes native publication digests, and compares all physical publication tensors against the initial state or matching checkpoint. The combined audit requires a useful Adam update and RL signal in the same update.

Stacked on #22 (test/native-source-membership); the diff contains only this qualification increment. Model/training/serving behavior is unchanged. Execution now requires --model-manifest.

Validation: 109 distinct CPU tests pass, including 20 new cases, model/preflight corruption controls and three actual tiny native Ouro/RLTT updates. All 16 physical tensors match across four publications; maximum actor/native score error is 4.77e-7. Full-chain worker records are synthetic, while the tiny test uses actual native generation and CPU AdamW. Neither establishes official Ray/CUDA qualification or reward convergence.

The comparison table, raw logs/XML, retained fixture failures, extracted measurements and file hashes are in docs/native_qualification_chain.md and benchmarks/results/native-qualification-chain/. GPU throughput, memory savings and reward convergence remain unmeasured.