[0z5a]

The public IQuest loop implementation loses first-loop KV in cache-free training, and its full-prefill logits differ from token-by-token decode. This draft adds explicit-KV functional replay and a reproducible semantic audit against the unmodified token-by-token autograd path. It also freezes the official source and all sixteen checkpoint headers, verifying 39,794,696,320 unique BF16 parameters without downloading weight tensor bodies.

Validation: two initialization seeds, lengths 1/4/8/65, ordinary and recomputed graphs, and two AdamW updates give 32 paired steps per attention backend. AUTO and MATH SDPA pass every output and whole-gradient check, but each passes only 24/32 strict parameter-update checks. Failed deltas, original cache/prefill failures, full matrices, source notices and SHA manifests are retained. The replay is not registered as a learner provider; GPU performance, full-checkpoint updates, recovery and reward convergence are unqualified.

Ruff, formatting, CLI help, reference-file hashes and fixture AST identity pass. New implementation code contains no try/catch, Any or getattr. This is stacked on #31; upstream branches and the existing runtime are unchanged.