Review-driven fix campaign: semantic bugs, framework subtraction, hygiene, docs convergence

[zaler23]

Summary

41 commits from a full-repo review (6-domain blind Codex review → PM adjudication → batched fixes), all adversarially re-reviewed to APPROVED over 3 rounds.

• A1 ×6 — semantic bugs: truthful evidence-blocked helpers; missing verification scores 0.5 (not 1.0) with verified-pass > unknown > failed; offspring no longer inherit parent verification state (mutation/crossover/patch_merge); Z3 proof semantics (Not(P) → unsat, satisfiability only when declared); final carry-forward requires matching artifact hash; final synthesis propagates provider interruptions to checkpoint instead of swallowing them.
• B1 ×8 — framework-shell subtraction (runtime net −879 LOC): model_adapter_facets collapsed into core; always-empty adaptive hooks + no-op loop control flow removed; unwired EventBus and VerificationTraceStore deleted; monolith-era imports stripped; fabric config converged to fields with real readers; Makefile fake test layering removed.
• C ×7 — engineering hygiene: uv.lock removed (pip+pyproject single source of truth); test-extra documented; Python contract aligned to CI matrix (3.11–3.13); env files parsed without source; env templates written 0600; dry-run side-effect-free; shellcheck CI gate.
• A2 ×5 — lifecycle/status semantics: route_incomplete/best_current_route pass through instead of folding to completed; resume_available derived from the same checkpoint the /resume endpoint checks; honesty-probe self-certification channel closed; evidence token-overlap stays advisory (never supported); failed patches (INCUBATING / final_answer_blocked_until_repaired) excluded from final-answer fallback.
• D ×9 — docs/code drift: finality language converged to answer-first semantics; seed-not-final, artifact layout, capability taxonomy, API surface, and quickstart demo now match actual behavior.
• F ×5 — adversarial-review fixes: terminal_reject/top_authority scoped to the latest artifact hash (removes a stale-evidence hidden gate); model-requested candidates hard-rejected only for structural blocks or the authorized failed-patch gate — otherwise unbound with the model answer preserved; launcher version predicate synced; status normalization.

Verification

• Full suite: 700 passed, 1 skipped (baseline 679/1; every batch's count reconciled), independently rerun outside the implementation sandbox.
• Adversarial review: 3 rounds on the merged diff (findings → fixes → re-verification with reproducers), final VERDICT: APPROVED.
• Red line held: verification/proof diagnostics remain advisory; the only hard exclusions from final answers are structural blocks and the failed-patch repair gate; objective_solved still always requires external verification.

Notes

• Runtime code net −708 lines; tests net +441 lines (regression locks, no defensive gilding).
• Known deferred items: 3.10/3.14 support claims dropped pending real testing; native-eval-suite.json marked as drifted (rename deferred); fixture-provider demo decision deferred.

🤖 Generated with Claude Code