eval: versioned golden replay corpus gating every gate/prompt/model/config change

[JSONbored]

Context

Parent: #8828 (Phase 2 — regression gate). ORB's decision policy ships without an evaluation gate, in a repo that enforces 99% patch coverage on everything else. The existing threshold/logic backtests (#8082 epic) replay rule changes; this extends that into a versioned golden corpus gating every gate-affecting change.

Requirements

• A versioned golden corpus of historical PR decision inputs with adjudicated expected outcomes, seeded from real failures: the reversal_superseded class, the CI-retrigger loophole cases, near-threshold holds, and the 126 human-closed-despite-merge-verdict disagreements identified on gate_decision can contradict the action taken — stale 'merge' verdicts recorded after CI-failure closes distort calibration #8825.
• Corpus entries store the full decision input snapshot (the same fields the Epic: provable decision accuracy — measurement, labels, calibrated abstention, and a legible decision record #8828 decision record pins) so replay is deterministic; the model-call output is recorded verbatim as an input, not re-queried.
• CI job on gate/prompt/ruleset/config changes (path-filtered like backtest-logic-check): replay the corpus, fail on any known-bad case flipping to merge, and fail on merge-precision regression beyond a stated tolerance.
• Corpus additions follow the loop: production failure → label → corpus entry; document the workflow in the backtest-calibration docs page.
• Tiered: fast subset on every PR; full corpus on model-version or prompt-template changes (the highest-risk uncontrolled inputs).

Test Coverage Requirements

99%+ patch coverage, branch-counted, on the runner and corpus loader; the corpus format gets a schema validation test.

Expected Outcome

No gate-affecting change ships without demonstrating it does not reintroduce a known failure — the same discipline the repo already applies to code, applied to judgment.

[linear-code]

JSO-6116