[hsliuustc0106]

Companion to .claude/skills/review-pr/SKILL.md: review-pr covers the Verifier stage; this walks the whole loop an agent runs for one issue, as proven on #71 → PR #74:

1. Establish the facts (crop the screenshot; file the issue first if needed), confirm scope.
2. Worktree + environment, including the gotchas a fresh worktree hits (harness node_modules symlink or DSH_HARNESS, HTTPS push timeouts → SSH URL).
3. Fail-without-change: test red at base, green at the tip, evidence kept for mechanical re-verification.
4. Every gate at the linked tip (build, test, --check, lint + arch-docs, patches generate/check), topics.json for first-touch files, lint warnings compared against base.
5. PR shape with Verifier-honest verification section.
6. Independent review loop per review-pr — tip advance voids the PASS; outside-the-diff root causes become follow-up issues at once.
7. Merge gate (same tip, agreement or waiver, maintainer's go) and cleanup.

Also adds item 8 to the AGENTS.md iteration loop pointing at it.

Docs-only (.claude/ + AGENTS.md): no gates apply locally; CI runs on this PR. Not run: everything code-shaped — nothing code-shaped changed.