test(hooks): paired-negative test for end-of-turn-reminder Stop hook (M.4.5)

[artyhoo]

Summary

Track M.4.5 — paired-negative coverage for .claude/hooks/end-of-turn-reminder.sh (Stop hook), closing one of the 6 untested-hook gaps surfaced in the post-PR #183 channel-earliness audit. Per umbrella kickoff .claude/orchestrator-prompts/m4-bash-hook-tests/kickoff.md §1 row 5.

Single new test file: packages/core/hooks/end-of-turn-reminder.test.ts (314 LOC, 12 tests). Zero production code change.

Test shape

• ❌ trigger turns must emit JSON {decision:"block", reason, systemMessage}:
  Branch A (long markdown), Branch B (bare question), Branch C (long+question),
  Branch D (file:line claim), AskUserQuestion-only tool_use.
• ✅ skip conditions exit 0 silent: stop_hook_active, missing transcript, no
  assistant turn, short non-claim turn.
• Boundary (precision): file:line claim inside a fenced code block does NOT
  count as a claim (hook:71-74 cleaner-fix guard).
• Regression guard (B2): AskUserQuestion-only turn after a prior ## 🟢 recap
  must FIRE, not be eaten by grep -qF "" matching everything (hook:128-131 fix).

Per T-M4-B the test asserts payload SHAPE — decision, reason (non-trivial string), systemMessage matching 🎯 <anchor> — not just exit code. Exit-only would have missed both the #81 broken-systemMessage-delivery bug and the round-1 AUQ false-suppress regression that prior cold-review caught.

Mutation-sanity (T19)

Verified the test has teeth: mutating production hook line 256 decision: "block" → "allow" makes 6 trigger tests fail; the 6 silent-exit tests correctly stay green. Hook restored before commit; diff -q against backup = 0.

Test plan

• npx vitest run hooks/end-of-turn-reminder.test.ts — 12/12 green at packages/core/hooks/end-of-turn-reminder.test.ts:1
• tsc -p packages/core --noEmit — clean
• Mutation-sanity: 6 trigger tests fail on decision: "block"→"allow"; restored
• pnpm test full suite — 825 passed / 4 skipped / 1 failed; the 1 failure is
  pre-existing on origin/staging in hooks/pre-push.test.ts:54-61 (asserts
  audit-ai-docs.test.sh / hook-stub-completeness.test.sh literal paths, but
  Wave 10.4 / 10.6 migrated those to .ts — see packages/core/hooks/pre-push.ts
  inline comment «audit-ai-docs.test.ts (Wave 10.4): run via vitest (replaces
  audit-ai-docs.test.sh)»). CI on origin/staging is green (audit-self workflow
  filters its vitest scope), confirming this failure is outside the gating set.
  Surfaced as observation per kickoff §2 + T5 (no fix bundled into M.4.5).

§1.7 Forward-check applied

• no-paid-llm-in-ci.md §1 — test uses
  deterministic vitest + spawnSync of the bash hook with on-disk JSONL fixture
  (packages/core/hooks/end-of-turn-reminder.test.ts:65-72); zero API-billed call.
• build-first-reuse-default.md §1
  verdict REFERENCE: test file leading docstring + spawnSync pattern at
  packages/core/hooks/end-of-turn-reminder.test.ts:30-48 mirrors precedent
  packages/core/hooks/check-hook-marker.test.ts:50-64. No new framework, no new
  dep, no SSOT row needed.
• paired-negative principle 02 —
  ❌ trigger tests at packages/core/hooks/end-of-turn-reminder.test.ts:118-218
  paired with ✅ silent-exit tests at :222-264 and a precision boundary at :271-285.
• ai-laziness-traps.md §2 — active T-traps:
  T3 (file:line cites against .claude/hooks/end-of-turn-reminder.sh:7-15,128-131,249-259
  throughout assertions), T5 (pre-push.test.ts pre-existing failure surfaced
  as observation only — not fixed in this PR), T11 (REFERENCEs precedent at
  packages/core/hooks/check-hook-marker.test.ts:50-64 in the docstring), T13
  (asserts payload SHAPE per T-M4-B at :127-131,143-148,159-167,180-187,199-204,
  not exit code only), T15 (test files are leaves — meta-tests-of-tests would
  be infinite regress; noted here pre-emptively), T19 (own mutation-sanity ran
  before push: 6/6 trigger tests failed on decision:"block"→"allow" mutation,
  hook restored, diff -q = 0).
• Domain-specific (kickoff §7): T-M4-A (TS + spawnSync, no bats), T-M4-B
  (payload-shape assertions on decision/reason/systemMessage, not exit-only),
  T-M4-C (Stop-hook contract specifically — reason → MODEL, systemMessage →
  USER per memory project_eot_hook_redesign_approved 2026-05-22; verified via
  hook source .claude/hooks/end-of-turn-reminder.sh:249-254 cited contract).

§1.7 Backward-check applied

• Scope-additive: exactly 1 new file (packages/core/hooks/end-of-turn-reminder.test.ts:1-314);
  zero file deletion; zero modification of .claude/hooks/end-of-turn-reminder.sh —
  verified via diff -q /tmp/eot-hook-backup.sh .claude/hooks/end-of-turn-reminder.sh
  returning identical after mutation-sanity restore.
• No existing artefact silently superseded. The 3 already-tested-hook precedents
  (check-hook-marker.test.ts, check-kickoff-traps.test.ts,
  inject-matching-rule.test.ts) remain unchanged; this PR adds a 4th sibling, not
  a replacement.
• No dual-implementation markers added to legacy hooks in this PR
  (per kickoff §6 — that is a maintainer-batched separate item per
  dual-implementation-discipline.md §9
  forward-going convention).
• Pre-existing CI-failure observation (pre-push.test.ts:54-61) surfaced in
  test-plan section above; not fixed here (T5 — no implementation bundled into
  test-add PR; maintainer decides scope).

Unblocks

Tracks 4/6 of M.4 once siblings M.4.1, M.4.2, M.4.3, M.4.4, M.4.6 also merge → unblocks Track M.5a (top-3 edit-time hooks: manifest render-drift / research-patch §1.7-substance / actionlint per-yml) per kickoff §3 downstream gate.