feat(AC-322): multi-objective advancement contract for generation gating

[jayscambler]

Summary

• AC-322: Canonical multi-objective advancement contract that gates generation progression on composite metrics instead of only local best-score delta

New Module

Module	Purpose
harness/pipeline/advancement.py	AdvancementMetrics, AdvancementRationale, evaluate_advancement()

Key Design Decisions (per comments)

• Confidence and uncertainty are first-class inputs — confidence, sample_agreement as explicit metric fields with risk flagging below thresholds
• Separate proxy from truth — search_proxy_score is informational; resolved_truth_score binds when present and overrides the delta check
• Binding vs proxy — binding_checks list shows what drove the decision; proxy_signals shows what was informational
• Error rate veto — > 20% error rate vetoes advance regardless of score improvement
• Regression → immediate rollback — negative delta never retries

Evaluation Chain

1. Error rate > 20% → rollback (binding veto)
2. Confidence < 0.5 → risk flag
3. Variance > 0.04 → risk flag
4. Resolved truth score present → binds over proxy (truth delta check)
5. Score regression (delta < 0) → rollback
6. Delta >= min_delta → advance
7. Delta < min_delta, retries remaining → retry
8. Delta < min_delta, no retries → rollback

Test plan

• 12 tests in test_advancement_contract.py
• AdvancementMetrics: construction, delta computed, roundtrip
• AdvancementRationale: construction, roundtrip
• evaluate_advancement: advance, rollback regression, retry marginal, rollback high error, low confidence risk, truth overrides proxy, component scores
• ruff clean, mypy clean, full suite 4415 passed

[linear]

AC-322 Define a multi-objective advancement contract for generation gating