[codex] require confirmed failures for eval export

[divo12]

Stack

Depends on #13. Review against codex/offline-e2e-release.

Research finding

Production evidence can detect a failure automatically, but root-cause attribution is not authoritative enough to create tuning or holdout data without confirmation. TRACE reports 72.7% root-cause attribution; AgentEval reports 72% against an 81% human ceiling. LangSmith/Langfuse dataset workflows likewise place curation between production traces and reusable eval data.

The full primary-source synthesis and gap audit is included in docs/research/2026-08-22-failure-mining-evals-tuning-ab-research.md.

Implementation

• Adds immutable ClusterReview records bound to cluster ID, revision, content digest, reviewer, decision, and timestamp.
• Adds deterministic DiagnosisReview, producing a new content-addressed diagnosis artifact rather than mutating the original.
• Adds explicit CONFIRMED and REJECTED review transitions while preserving prior review history across diagnosis revisions.
• Reopens a rejected cluster only when new diagnosis content changes its revision.
• Reopens any previously trusted CONFIRMED/TARGETED cluster when its content changes, forcing a new review.
• Routes proposed, reopened, rejected, or missing failure clusters to REVIEW.
• Allows only CONFIRMED or TARGETED clusters into memory, frontier, regression, selection, or admission partitions.
• Applies the confirmation gate before historical partition reuse, preventing old unsafe exports from bypassing it.
• Allows a confirmed cluster to graduate from a prior REVIEW partition while preserving established non-review partitions.
• Updates the offline release to confirm the exact four cluster revisions before eval export.

TDD evidence

• Proposed clusters stay in REVIEW and cannot populate any eval or holdout suite.
• A matching review deterministically confirms the exact cluster revision.
• A stale content digest fails with REVIEW_INVALID.
• Rejected clusters remain review artifacts.
• Changed confirmed content returns to REOPENED; prior REVIEW data graduates only after exact confirmation.
• Existing confirmed-cluster exports and leakage checks remain green.
• Full trace-to-review release still produces one governed winner.

Verification

• pytest -q — 156 tests in the full suite; focused Mine/diagnosis/export/E2E verification is green
• Ruff and strict mypy across src and tests — clean
• forbidden construct scan — no Any, raw dict, object, getattr, setattr, lambdas, or __import__
• Bandit — zero findings
• source distribution and wheel build successfully

Deliberate scope

• Implicit dissatisfaction remains review evidence, not verified failure truth.
• No content capture, vector database, or LLM judge is added.