docs(dispatch-worktree-automation): I-phase Sub-wave C — recursive acceptance (step count 7→2 verified)

[artyhoo]

What

I-phase Sub-wave C of dispatch-worktree-automation — the recursive acceptance test for the substrate landed by PR #279 (WorktreeCreate hook, merge 5278675, MERGED) + PR #282 (skill prose → claude -w, merge 9f8dc84, MERGED). Both verified ancestors of live origin/staging (3fe1336).

Ships: docs/meta-factory/research-patches/2026-05-29-dispatch-worktree-iphase-acceptance.md (acceptance report) + a 1-line CLAUDE.md ## See also pointer (pointer, not duplication — dual-implementation-discipline §7).

Empirical step-count measurement (kickoff §4 criterion 2, target ≤2)

Dispatch	Command	Outcome (command-verified)	Steps
Baseline (pre-#279/#282)	git worktree add+cd+2×ln -s+git checkout -b+open tab+paste	manual	7
Primary — THIS session	claude -w iphase-acceptance + paste prompt	worktree + branch worktree-iphase-acceptance + both node_modules symlinks @18:32	2
Secondary — parked umbrella	claude -w aif-handoff-sw-A-resume (headless)	worktree + symlinks @18:42, PROBE-OK, cleaned up after verify	1

Verdict: criterion 2 MET (7→2, →1 headless). This session itself was spawned by claude -w iphase-acceptance, so the primary data point is the live environment, not a reconstructed trace (T15).

STOP §10 — NOT triggered

Step count ≤2 ✓ · hook fired (symlinks present, both runs) ✓ · claude -w behaved as coded ✓. No revert of #279/#282.

MINOR finding (§3.2) — surfaced, not patched here

claude -w branches from origin/HEAD, which is bb3ecec (stale) while origin/staging is 3fe1336 — both test worktrees were born behind live staging. Root cause: hook base-ref loop prefers origin/HEAD (worktree-setup.sh:73), and that symbolic ref isn't refreshed by plain git fetch. Step-count win is unaffected (one git rebase fixes content). Hardening worktree-setup.sh base-ref resolution is a dedicated follow-up — out of Sub-wave C scope (C = acceptance + docs only).

§1.7 Forward-check applied

• no-paid-llm-in-ci: pipeline = shell flag + deterministic bash hook; the -p probe is a single subscription-bundled call (pre-2026-06-15), not CI cost. ✓
• build-first-reuse-default §1: ADOPT (claude -w) + ADAPT (tfriedel/claude-worktree-hooks), SSOT feat(skill-drift): multi-channel drift detection from D-AuditC-5 #65; this report measures, builds nothing. ✓
• doc-authority-hierarchy §2-§3: patch carries <!-- scope:... --> (principle 10, test green) + Authoritative-for header. ✓
• phase-research-coverage §1.7: every verdict command/file:line-backed; the negative claim (stale origin/HEAD) is git rev-parse-verified. ✓

§1.7 Backward-check applied

• worktree-setup.sh NOT edited (C = acceptance + docs); §3.2 finding surfaced as explicit follow-up, not silently absorbed. ✓
• PR research(dispatch-worktree-automation): R-phase — ADOPT claude -w + ADAPT WorktreeCreate hook #271 R-phase patch implemented/measured, not superseded. ✓
• CLAUDE.md gains a pointer only; no existing content rewritten. ✓
• Kickoff §12 anti-scope respected (no Sub-wave A/B edits, no parked-kickoff edits, no principle test added). ✓

Reproduction

gh pr view 279 --json state --jq .state   # MERGED
gh pr view 282 --json state --jq .state   # MERGED
ls -la .claude/worktrees/iphase-acceptance/node_modules   # symlink → primary node_modules
npx vitest run packages/core/principles/10-research-patch-annotation.test.ts   # 5 passed

🟢 Простыми словами

Было: чтобы запустить изолированную параллельную сессию — ~7 ручных команд. Стало: одна команда claude -w <имя>, хук сам делает папку и симлинки. Проверено на этой самой сессии (2 шага) и на свежем имени для запаркованной задачи (1 шаг, потом убрал). Цель «≤2 шага» достигнута. Нюанс-не-блокер: ветка создаётся от устаревшего origin/HEAD — записал как отдельную доработку хука.