[P1] rescore: stale judge caches are silently attributed to a newly requested judge model

Problem

Both pending selection and rescore_one reuse any cached non-null verdict, without comparing judge identity/configuration, rubric revision, instruction or intercepted evidence. The aggregate then unconditionally writes roll['judge_model'] = args.judge_model. Requesting a new judge can therefore produce a report labeled with that judge even when every verdict came from an older judge.

Sources: cache reuse, pending selection, aggregate attribution.

Verified reproduction

Create one intercepted run with judge.json = {"match": true, "judge_model": "old-judge"}; run rescore.main() with --judge-model new-judge --rubric strict, a temporary dummy model configuration, and a mocked judge function.

Observed: exit 0, judge calls 0, rescore-summary.json.judge_model = "new-judge", n_match_strict = 1. Direct rescore_one(..., judge_model='new-judge', force=False) returns the old-judge object unchanged. No network call was made.

Acceptance criteria

• Define a non-secret evaluation fingerprint covering judge identity/config, rubric/prompt revision, instruction and intercepted input.
• Reuse only a valid, matching cache; handle legacy cache entries explicitly (recompute or report unknown provenance).
• Preserve actual per-verdict provenance and reject or explicitly describe mixed-provenance aggregates.
• Keep --force as an explicit recomputation option; ordinary model changes must not silently mislabel results.
• Test changing model, input and rubric revision, and unchanged-cache reuse.

Suggested priority P1: misleading benchmark attribution. Related to provenance PR #347, but run metadata alone does not fix this evaluator cache contract. #313 fixes paths/config loading; #319 fixes judge transports; neither addresses cache identity.

@Perry2004 please review the cache/provenance contract and reporting requirements.

Start in src/clawbench/eval/rescore.py at the cache reuse, pending selection, and aggregate paths linked in the issue; trace rescore_one and rescore.main with the reproduced old-judge/new-judge case. Add focused coverage for changed model, intercepted input, rubric revision, and unchanged-cache reuse, then verify rescore-summary.json preserves actual provenance and handles legacy or mixed results explicitly.