feat(m2-harness): judge reliability validator + paired stats, calibration, replication, distractor scoring runners

Add evals/validate-judge.js with:
- agreement(): compute percent agreement + Cohen's kappa from vote fixtures
- biasDetection(): detect format/length preference in judge votes
- main(): CLI for live droid-based judge reliability runs

Add evals/tests/validate-judge.test.js with 11 offline tests covering:
- 3-vote fixture percent agreement = 2/3, kappa = 0.4
- Single judge reports cannot estimate agreement
- Perfect/total disagreement edge cases
- Length bias detection on planted fixtures

Includes previously uncommitted M2 harness work:
- lib/stats.js: pairedDiff, mcnemarFull/midp, scoreDistractor
- run-calibration.js, run-replication.js, run-distractor-scoring.js
- evals/tests/stats.test.js (7 tests)

All 18 tests pass; backward-compat preserved (legacy exports, summarize fields, validate-skills)