feat(eot-hook): Branch A — recommendation-laziness self-check nudge

[artyhoo]

Summary

Extends .claude/hooks/end-of-turn-reminder.sh Branch A (long-text recap, no question) with a recommendation-laziness self-check nudge — mirrors the existing Branch B/C «настоящая развилка» framing for the case where there is no AskUserQuestion.

Why: Defer-reflex Stage 2 benchmark (PR #232) REJECTed the regex-based detection mechanism (precision 3%, recall 0/3). The underlying problem (agent says «PR ready, ждёт твой клик» instead of merging — 3 documented incidents in feedback_no_human_verification_ai_self_verifies.md) is real and uncovered. Maintainer pointed out: Branch B/C already has the «один вариант явно лучше → НЕ спрашивай, сделай» nudge, but it only fires on asked=true. Defer-reflex incidents happen in Branch A turns (long recap, no explicit question). Extending Branch A with the same nudge closes the gap without a regex gate — text-recall reminder fires always, no precision/recall trade-off.

Sibling, not regex replacement: This is the text-recall complement to T20 (inline-verdict-without-evidence) — same family, different action surface. Per recommendation-laziness-discipline.md §1.12 framing.

Changes

File	LOC delta	What
.claude/hooks/end-of-turn-reminder.sh	+1	New bullet inside Branch A cat <<EOF block, between «Следующий шаг» and «Любой пункт не выходит конкретным»
packages/core/hooks/end-of-turn-reminder.test.ts	+4 (1 assertion + 3-line comment)	Inside existing Branch A it(...) block: `expect(payload.reason).toMatch(/рекомендовал

Not a capability commit

Hook edit (existing file) + test extension (existing file). No new dep, no new ≥80 LOC packages/ source, no new file under new packages/ subdir. Pre-push detectCapabilityReason returns null → no Prior-art trailer required.

How to verify

• All 14 hook tests pass — verified: pnpm vitest run hooks/end-of-turn-reminder.test.ts green, including new Branch A nudge assertion
• Pre-push hook ran tests on push — verified: 115 tests passed (18 files), pre-push exit 0
• Hook smoke — verified: bash hook <<< '{"transcript_path":"/dev/null","stop_hook_active":false}' exits 0 silently
• Diff scope — verified: 2 files, +5 lines total, no drive-by changes
• Branch A wording mirrors B/C — verified: identical imperative tone, «настоящая развилка» / «сделай и скажи что сделал» action verb
• @cc-only-rationale annotation preserved — verified: hook line 2 unchanged (internal dev tooling, not shipped to consumer projects)
• No regex gate added — verified: this is text-recall reminder only, fires always when Branch A fires; orthogonal to PR research(defer-reflex-detection): Stage 2 benchmark — REJECT UMBRELLA verdict #232's REJECTed mechanism
• Empirical effect — owner: maintainer + future sessions; falsifier: defer-reflex incidents continue WITH this nudge active in Branch A → text-recall insufficient → escalate to AI-agnostic sub-agent (option 3 from Stage 2 §5)

§1.7 Forward/Backward-check applied

Forward-check:

• Complies with no-paid-llm-in-ci.md §1: deterministic bash, no LLM, no API calls — file:line evidence: .claude/hooks/end-of-turn-reminder.sh:228 (text-only insertion in heredoc)
• Complies with dual-implementation-discipline.md §3: hook already carries @cc-only-rationale: internal dev tooling — end-of-turn reminder injection for maintainer's CC session; not shipped to consumer projects via install.sh at line 2; extension inherits — no annotation change needed
• Complies with parallel-subwave-isolation.md §1: Worker ran in harness-allocated worktree via isolation: \"worktree\"
• Complies with recommendation-laziness-discipline.md §1.12 — this is exactly the «delivery-channel» implementation of §1.12 «Lead with a reasoned recommendation; act when the best path is clear» for the non-AskUserQuestion case

Backward-check:

• Does NOT supersede T20 (ai-laziness-traps.md) — same family, different surface; T20 = inline-verdict-without-evidence (recommendation moment with no preceding tool call); this nudge = «check yourself for laziness in recommendation/deferral within recap moment»
• Does NOT modify recommendation-laziness-discipline.md — that rule stays Class C; this is an additional always-on delivery channel
• Does NOT touch ai-laziness-traps.md — no new T-trap added; T21 slot remains free for next claimant
• Does NOT reverse Stage 2 REJECT (PR research(defer-reflex-detection): Stage 2 benchmark — REJECT UMBRELLA verdict #232) — regex mechanism stays REJECT; this is the orthogonal text-recall path

Rollback plan

Revert single PR — Branch A reverts to pre-nudge wording; recommendation-laziness discipline reverts to T20 + Branch B/C-only coverage (the pre-this-PR state).