♻️ refactor(skills): slim tmb_planning + tmb_review under 200 LOC (#2904 Phase A+B)

Resolves GL !2904 — v0.7.0 token-burn: slim bro's always-on skill set.

Three commits

1. f0602145 → Phase A: docs(skills) boundary audit + metadata research + placeholder refresh

   • NEW docs/architecture/skill-boundary-audit-v0.7.0.md — per-sentence DETERMINISM classification of all 8 skills
   • NEW docs/architecture/skill-metadata-research.md — industry survey; recommendation (b) no invented field
   • EDIT CODEX.md / CURSOR.md / GEMINI.md / gemini-extension.json — placeholder descriptions refreshed to v0.7.0

2. df5227e6 → Phase B commit 1: test(l6) rewrite step-14 prompt to natural code-touching ask

   • Drops the explicit tmb_planning naming from the prompt so the scenario actually tests bro's description-match judgment

3. 7fd4f5ba → Phase B commit 2: refactor(skills) slim tmb_planning + tmb_review under 200 LOC

   • skills/tmb_planning/SKILL.md: 292 → 189 LOC
   • skills/tmb_review/SKILL.md: 274 → 195 LOC
   • 4 new composites in composites.ts: headless_intent_start, bro_verification_fail_record, pr_review_worktree, reap_and_review_prep
   • 2 new hooks: pr-reviewer-spawn-prompt-shape.sh (PreToolUse[Agent] — 4-anchor + no-rubber-stamp), post-pr-comments-persist.sh (PostToolUse[pr_comments_get])
   • Registered both in hooks/hooks.json

Doctrine basis

• docs/architecture/DETERMINISM.md — boundary test applied to every procedural sentence; DETERMINISM-classified steps migrate to mech 1-6
• feedback_no_negative_prompt_rules.md — positive directives only
• feedback_no_citations_in_prompts.md — no issue # / GL ! in prompt-loaded files
• skill-metadata-research.md recommendation (b) — no new frontmatter fields invented

Rebase note

Branch was rebased onto dev after !2922 + !2923 + !2924 + !2925 merged. One conflict resolved on pr-reviewer-spawn-prompt-shape.sh: combined Phase B's 4-anchor + no-rubber-stamp logic with !2925's helper sourcing + tmb_normalize_role normalization. Stricter validation + prefix-mismatch fix in one hook.

Verification

bash tests/run-all.sh exit 0 on rebased branch.

• L1: 26 lints all PASS
• L2: 438 pass / 0 fail (+15 from new composite tests + lint)
• L3: 55 pass / 0 fail
• L4: all 5 workflow-sim flows PASS

pr-reviewer

Verdict: PASS on both Phase A (validation_attempts id=22) and Phase B (id=23). 5-phase review notes:

• 6 new mechanisms structurally sound (atomic transactions, role-gated, with tests)
• All JUDGMENT sentences preserved per boundary audit
• New hooks follow established convention; composites mirror existing pattern
• One MUST/MUST NOT pair in tmb_review §C is now hook-backed (mech 3 promotion) — permitted

Held for manual review — no auto-merge per Human request.