feat(0.31.0): JudgeScoresRecord on RunRecord.outcome

[tangletools]

The gap this closes

agent-builder PR #179 just landed a forge-chat judge rubric that
computes three per-dim scores (helpfulness, clarity, on_topic) per
cell, but only persists the composite to records.jsonl because
RunRecord.outcome doesn't have a per-judge / per-dim slot. The same
problem hits every consumer that wants ensemble scoring across the
five product agents (tax / creative / legal / gtm / agent-builder).

Consumers were either dropping the breakdown on the floor or
smuggling it through stringly-typed outcome.raw keys like
judge_kimi_helpfulness — neither survives a corpus-IRR run.
corpusInterRaterAgreement (0.27.2) expects structured per-judge
per-dim records, not parsed strings.

What this ships

• JudgeScoresRecord type (src/run-record.ts):
  • perJudge[judgeId][dim]: number — canonical store
  • perDimMean[dim]: number — convenience projection
  • composite: number — mirrors the score the gate sees
  • failedJudges?: string[] — explicit dead-judge ids
  • notes?: string — panel prose
• RunOutcome.judgeScores?: JudgeScoresRecord — additive on the
  outcome; existing single-judge runs leave it unset.
• CampaignRunOutcome.judgeScores? wired through runEvalCampaign
  so per-cell ensemble outcomes land on RunRecord.outcome.judgeScores
  unchanged.
• Validator extended in validateRunRecord: per-judge / per-dim /
  composite scores must be finite (no silent NaN-as-zero);
  failedJudges entries must be non-empty strings.
• Tests in tests/run-record.test.ts and tests/eval-campaign.test.ts
  cover all four shapes (full, partial with failedJudges, missing,
  with notes) plus a fail-loud case where one judge throws and the
  record carries the dead-judge id, not a silent zero.
• Consumer contract (tests/consumer-contract.test.ts) pins
  JudgeScoresRecord as a type-level export so consumer code stops
  compiling if the field gets renamed.

Design tradeoffs

• perDimMean and composite are precomputed projections of
  perJudge. Storing both costs a few bytes per record but spares
  every reporter and IRR primitive a re-aggregation; the trade is on
  the right side for the read-heavy access pattern.
• failedJudges?: string[] is the typed-outcome answer to partial
  failures. Missing keys in perJudge would be ambiguous (silent
  zero vs not run); the explicit list is fail-loud.
• Field is optional on RunOutcome so the 0.30.0 surface is
  preserved. Scalar-only runs leave it unset.

What consumers gain

• Forge-chat (agent-builder) stops dropping per-dim scores;
  corpusInterRaterAgreement consumes the records directly.
• Tax / creative / legal / gtm agents inherit the same typed slot
  without each implementing their own conventions on outcome.raw.

Version bumps (lockstep)

• package.json 0.30.0 → 0.31.0
• clients/python/pyproject.toml 0.30.0 → 0.31.0
• clients/python/src/agent_eval_rpc/__init__.py 0.30.0 → 0.31.0

Test plan

• pnpm typecheck clean
• pnpm test — 1208 tests pass (5 new in eval-campaign, 5 new in run-record, 1 new in consumer-contract)
• pnpm build clean (tsup + openapi)
• After merge: human tags v0.31.0 and pushes to fire the publish workflow.