feat(rex): land the review-experiment epic work on main (PMAT-4524)

[noahgift]

Summary

Lands the review-experiment (rex) epic work on main as one commit (PMAT-4524).

• aprender-review-experiment: new and changed modules, their tests,
  and the rex and replay examples (Qwen3.5 as the model, llama.cpp as
  the runtime). replay run --out writes its file whole, so a rerun
  replaces the s1-v2 rows and never appends to them.
• contracts: new and amended contracts, with the ONT witness
  regenerated for the new relation set.
• aprender-serve: top-K logprobs per generated step on the Qwen3.5
  hybrid chat path, and POST /v1/chat/prompt-ids mounted on that
  path only. The infer path refuses --logprobs for an APR or
  SafeTensors model before the load.
• apr run: --logprobs K (at most 20, as top_logprobs on HTTP)
  records the K most likely next tokens and the encoded prompt ids
  in the --json and --stream documents (apr run: payload trace / --logprobs must record per-step top-k logits (no CLI path today; found on #4004) #4026). It is refused by name
  wherever it cannot be reported: text output, --benchmark (with or
  without --json) and --batch-jsonl. Recorded steps whose chosen
  tokens are not this run's generated tokens are refused, not
  reported. One exception: a turn that ended on a stop token may
  carry one trailing step more than its reply, the stop token the
  dense GGUF route drops from the reply; that one step is dropped.
• Gate change: check_append_only_ledgers.sh now accepts a file under
  docs/audits/ that .gitattributes declares audit-golden (written
  whole by one writer) without a union merge, and FAILs one declared
  both ways. Only the bare attribute declares a golden. .gitattributes
  declares the review corpus and the PRM-S1 v2 rows golden and unions
  the rex-001 admission ledgers by glob. The self-test proves a real
  ledger is not golden, that the scan FAILs a golden+union tree and
  passes the golden-only control, and that the value forms are not
  golden.
• The binding scanner now sees FALSIFY-TDC-009: a trailing comment on
  its attribute line hid the test fn.
• Three functions split below the complexity bar, with no change in
  behaviour: apr's model resolver, its run and diagnostic
  dispatch, and the orchestrate tool-call envelope parser.
• docs, evidence and scripts: the audit receipts and roadmap entries
  behind the above, and a read-only transcript-retention audit. The
  prm-04 parity capture carries <models>/ in place of two absolute
  weight paths; its receipt names that edit and both hashes.

Gate change (narrow)

scripts/check_append_only_ledgers.sh gains a declared audit-golden class. docs/audits/review-corpus/*.jsonl moves from must-union to audit-golden: the corpus is written whole by a single writer (one commit, 989cb01), and the guard never had a history check. New FAIL rows: a file declared both golden and union. Signed as a narrow gate change in its C314 round.

Test plan

• cargo test -p aprender-serve --lib: 16277 passed, 0 failed
• cargo test -p apr-cli --lib, the rex crate tests, and the apr run: payload trace / --logprobs must record per-step top-k logits (no CLI path today; found on #4004) #4026 CLI test
• clippy -D warnings on aprender-serve, apr-cli and the rex crate
• contracts gate: 8 of 8 steps ran, 0 failed
• README counts and ONT witness regenerated on this base

Refs #4354
keep-open: #4354 is the rex epic; this PR lands one ticket of it (PMAT-4524), other rows stay open
ont-delta: resolves agent-trace-terms-v1 lane-independence-v1 review-b2-loop-v1 review-champion-challenger-v1 review-lane-perf-ratchet-v1 review-ledger-v1 review-ledger-v2 review-replay-v1 rex-pilot-projection-v1 rex-promotion-ladder-v1 rex-promotion-ladder-v2 sparse-logits-v1 trace-admission-prov-v1 trace-admission-secret-v1 trace-datacard-v1 trace-dedup-v1 trace-outcome-join-v1 trace-split-guard-v1 workload-profile-v1 (19 schema contracts enter contracts.nt; Contract nodes 1838 → 1857)

🤖 Generated with Claude Code