feat(bench): held-out rubric evaluator + dep bump to ai-consensus-core 0.11.1

[marceloceccon]

Summary

The bench was measuring the wrong thing. On architecture_v2 (4 cases × 3 runs × seed 42), with grok-4.3 as both judge and baseline and claude-opus-4-5 as a held-out evaluator:

Metric	Consensus	Baseline	Δ
Self-reported (consensus score vs. baseline confidence)	60.0	75.4	−15.4
Held-out rubric (judged by claude-opus-4-5, blind)	83.3	48.0	+35.3

Consensus wins on the held-out rubric in 12/12 runs (100%). Same 12 runs, the self-reported-confidence metric says consensus loses 11/12. The two metrics invert. Without a quality measurement, the bench was producing a measurement artifact, not a result about the panel.

This PR adds a held-out LLM-as-judge rubric evaluator. Pass --evaluator-model + --evaluator-provider; when the panel declares a rubric, both consensus and baseline outputs are scored blind by a third model (neither side of the comparison). architecture_v2 ships a 5-criterion rubric (quantification, single-recommendation, reversibility-weighing, tripwire-specificity, failure-mode-realism). Adding a rubric to any other panel is purely additive.

Also bumps ai-consensus-core to ^0.11.1 — fixes the silent extractJudgeConfidence 50 default that previously made judge-confidence numbers in this repo's bench reports meaningless. Judge confidence on this 12-run suite now reports a real distribution (μ=66.9, σ=5.2) instead of (μ=50.0, σ=0.0).

See the new README section "Quality benchmark — held-out evaluator" for the full methodology, per-case Δ-rubric table, exact reproduction command, and honest caveats.

Per-case Δ rubric (12 runs, seed=42)

Case	Runs	Mean Δ
arch-microservices-day-one	+36, +28, +32	+32
arch-event-sourcing-billing	+44, +52, +56	+51
arch-sync-vs-async-fanout	+24, +40, +40	+35
arch-db-multi-tenant	+12, +32, +28	+24

Baseline scored 28/100 on every event-sourcing-billing run — a reproducible single-model blind spot that the panel surfaces every time.

Implementation surface

• New module src/benchmark/rubric.ts: structured JSON-emitting evaluator prompt, tolerant bracket-scanning JSON extractor (no regex backtracking), zod schema validation, score clamping, sentinel errorMessage on any failure path. Never throws — a rubric eval failure is data quality, not a suite-fatal error (matches the contract baseline.ts uses).
• New Preset.rubric?: readonly RubricCriterion[] field. Panel-declared because criteria are domain-specific.
• New CLI flags --evaluator-model + --evaluator-provider (validated as a pair). CLI warns when the evaluator coincides with baseline (self-grading) or judge (same brain producing and grading the consensus output) — the held-out contract is the bench's only guarantee that the comparison isn't self-graded.
• New report metrics: consensusRubricNormalizedMean, baselineRubricNormalizedMean, consensusBeatsBaselineRubricRate, rubricRunsCounted. Markdown report grows a "Held-out rubric" block and 3 per-case table columns (Rubric C / Rubric B / Δ rubric).
• Bumps ai-consensus-core ^0.10.0 → ^0.11.1 (upstream JUDGE_CONFIDENCE directive contract fix, entropyvortex/consensus-core#4).

Test plan

• npm run check clean — typecheck, lint, format, tests, coverage all green
• 341/341 tests pass (+25 from this branch: 16 rubric unit + 3 runner integration + 2 format + restore existing on new shape).
• Coverage above all thresholds (stmts 79.4%, branches 66.1%, fns 86.4%, lines 80.9%).
• End-to-end quick bench (bench --quick --evaluator-model claude-opus-4-5 --evaluator-provider anthropic) produces real rubric scores.
• End-to-end full suite (12 runs, ~$1) reported in README — reproduces deterministically with --seed 42.
• CLI warning logic exercised manually: warns when evaluator coincides with judge or baseline; rejects --evaluator-model without --evaluator-provider and vice versa.

Notes for the reviewer

• Stacks on top of feat(v0.12): expert panels, benchmarking suite, persistent project memory #3. Re-target to main after feat(v0.12): expert panels, benchmarking suite, persistent project memory #3 merges, or merge into the v0.12 branch first.
• No breaking changes. Callers who don't pass --evaluator-model see the existing report unchanged — with the side effect that judge confidence now reports a real number rather than the silent 50.
• Bench JSON artifacts and progress logs added to .gitignore (bench-*.json, bench-*.log) so reproduction outputs don't trip future format checks.