feat(eval): aggregateJudgeVerdicts — generic judge-ensemble reducer

[drewstone]

The one substrate primitive the shared agent-app eval-campaign scaffold needs: a generic judge-ensemble reducer. legal (aggregateEnsemble, tests/eval/lib/scoring.ts), creative (production-loop/judges.ts), and tax (judge-ensemble.ts) implement the same reduction three times — fan out N uncorrelated judges, mean each dimension over the survivors, report inter-rater spread, sum cost. This lifts it to the substrate, keyed by the caller's rubric (D extends string).

Behavior (fail-loud)

• Survivor mean per dimension; out-of-range judge scores clamped to [0,1].
• A failed judge (perDimension: null) is recorded in failedJudges, never folded into a zero. All-failed throws — a silent zero here corrupts the gate's number (CLAUDE.md "No fallbacks. Fail loud.").
• Cost sums over ALL verdicts, failed included — a failed judge still burned tokens; counting only survivors under-reports spend (the cost-ledger silent-zero class).
• maxDisagreement = max over dims of (max−min) across survivors — the inter-rater signal gates/analyzeRuns consume.
• composite reuses the substrate's weightedComposite (no re-implementation). No weights ⇒ uniform; a partial weights map selects-and-weights exactly those dims.

Verification

• pnpm typecheck clean · pnpm lint exit 0
• pnpm test — 1889 passed, 2 skipped (1879 prior + 10 new: survivor-mean, uniform/weighted composite, failed-judge-not-zero, all-failed-throws, cost-includes-failures, disagreement-spread, clamp, empty-input, rationale)
• Additive only

Next

• agent-app /eval-campaign façade composes this in buildEnsembleJudge; legal/creative/tax delete their copies and call it.

[tangletools]

Pure reducer, fail-loud (failed judge ≠ zero, all-failed throws, cost includes failures), composes weightedComposite. 10 new tests, 1889 green. Approving.