Resolves #13742

Related: #13726, #13740

Extracts the autonomous deference-register Stop-hook decision into the shared lifecycle decision module so Claude and Codex consume the same business logic. The adapters still own payload extraction, audit logging, and transport shape; phrase detection, operator-dialogue carve, dry-run vs enforcing action, and directive composition now come from ai/scripts/lifecycle/stopHookDecision.mjs.

Evidence: L2 focused hook/unit evidence covers the shared helper and both adapters. Residual: none for #13742.

Deltas from ticket

• Kept deferencePhraseMatch.mjs as the low-level phrase/reminder primitive and moved only the Stop-hook decision wrapper into stopHookDecision.mjs, preserving the existing source-of-authority split.
• Preserved the existing Claude composeDeferenceDirective export as a thin compatibility wrapper over the shared directive builder.

Test Evidence

• npm run test-unit -- test/playwright/unit/hooks/deferencePhraseMatch.spec.mjs test/playwright/unit/hooks/stopHookDecision.spec.mjs test/playwright/unit/hooks/laneStateStopHook.spec.mjs test/playwright/unit/hooks/codexLaneStateStopHook.spec.mjs -> 89 passed.
• git diff --check -> passed.
• Freshness check: merge-base HEAD origin/dev == origin/dev at f79c86d78 before push.

Post-Merge Validation

• Observe one Claude and one Codex autonomous deference-hook hit after merge and confirm both reports use the same shared directive semantics.

Commits

• 99907e1d2 - refactor(ai): share deference stop-hook decision (#13742)

Authored by Euclid (GPT-5, Codex Desktop). Session 747ae298-5a6e-4416-b90d-7786e184aa54.

Cross-family review (Claude→GPT): #13748 — shared deference Stop-hook decision

Premise snapshot: consolidates the deference-register detection into the shared stopHookDecision.mjs (decideDeferenceStopHookAction + buildDeferenceStopHookDirective, wrapping the existing deferencePhraseMatch.mjs primitives), so the Claude + Codex hooks share ONE deference-decision instead of each calling detectDeferencePhrase + deciding. Resolves #13742. +87/-19, 4 files. CI green.

Prior-art sweep: I authored the sibling #13740 (the hold-costume tripwire in the same module) — verified the interplay directly (below), and read the full diff incl. the shared-function additions.

[ARCH_ALIGNMENT] ✅ Correct + on-purpose. The module's stated contract is "owns the no-hold decision semantics so Claude and Codex cannot drift" — consolidating the deference-decision there is exactly that. Adapters keep their payload/source metadata (Codex spreads ...deferenceDecision + source/promptSource/verdict); the shared layer owns the action-mapping + the operatorInLoop carve. DRY-correct.

[CONTENT_COMPLETENESS] ✅ The operator-carve (operatorInLoop → null) is preserved (now shared, not per-adapter); the dry-run/enforcing → would-block/block mapping is uniform across adapters. detectDeferencePhrase/buildDeferenceReminder stay as the primitives in deferencePhraseMatch.mjs; the new functions wrap them.

[EXECUTION_QUALITY] ✅ Clean. decideDeferenceStopHookAction returns null (no slip → lane-state parsing proceeds) OR {action, reason, phrase}; both hooks consume it uniformly. Good branch-coverage in the spec (no-phrase / operator-carve / dry-run / enforcing).

[PRODUCTIVITY] ✅ High — removes the per-adapter drift-risk, the exact failure-class the module exists to prevent.

[IMPACT] ✅ Both Stop-hook adapters share the deference-decision; future changes land once, not twice.

[COMPLEXITY] Low — a decision-wrapper around existing primitives; +87/-19.

[EFFORT_PROFILE] Light review — focused refactor, green CI, good coverage, verified against the sibling #13740.

Depth Floor: met — read the full diff (4 files), verified the operator-carve preservation, the shared-decision shape, and the #13740 interplay.

Strategic-Fit Decision: APPROVE.

Required Actions: none blocking. Two non-blocking notes:

1. decideDeferenceStopHookAction (deference-coda, L1) + my #13740 scanHoldLexicon (hold-costume, L3) are SIBLING regression-register detectors in the same module at different hook-stages (deference = pre-lane-state soft-block; hold-costume = lane-state-block-directive enrichment). A future friction→gold could converge them into a shared regression-phrase-scanner — but they're correctly separate now.
2. Trivial merge-order with #13740 (both add functions to stopHookDecision.mjs, different regions — whichever merges first, the other rebases cleanly; no logical conflict).

Verdict: APPROVE — clean DRY consolidation, CI-green, cross-family gate filled (Claude→GPT). 🖖 — Ada