feat(evals): record judge identity and refuse cross-identity comparisons #8573

Problem

LLM-as-judge scores are only comparable when the evaluator is the same. The
judge currently returns { scores, rationale, weightedScore, passed } with no
record of who judged or how. If the judge model, rubric, prompt, parser, or
decoding settings change between a baseline and a candidate, a score delta
(e.g. 0.71 → 0.79) can be read as an improvement when it is really a change of
evaluator. The recorded-transcript replay makes the transport deterministic; it
does not make two differently-judged scores comparable.

Proposal

Persist a small judge-identity envelope with every verdict and fail closed on
comparative claims when a material field differs.

• JudgeVerdict gains an identity field: judge provider/model, rubric digest,
  parser version, decoding settings (temperature), and the scenario id.
• The judge check / report surfaces the judge model and rubric digest.
• A comparison helper classifies a baseline→candidate delta as comparable or
  insufficient-evidence when any material identity field differs.
• Add one small, frozen, human-adjudicated calibration slice so judge drift can
  be detected rather than treating the judge as ground truth.

Scope

Follow-up to the judge scorer. Keep it small: identity + digest + a comparison
guard + one calibration fixture. Do not build a general evaluation platform.

Acceptance criteria

• Every verdict carries a stable judge-identity envelope
• The report shows the judge model and rubric digest
• A delta across different judge identity is reported as insufficient
  evidence, not as an improvement
• A frozen calibration slice with known expected scores exists
• Documented how to refresh the calibration slice