feat(qa-review): advisory review job that complements existing PR checks

[lihsaa591]

Summary

Step 2 of the QA-review build. Adds the review job to qa-review.yml (runs after the gate), still manual-dispatch only.

Re-scoped after inspecting the pilot repo: user-registration-pro already has PHPCS-on-PR, an AI security scan, and claudegrill's E2E suite + .themegrill-qa/ data (and claudegrill deliberately keeps AI off the PR path). So this does NOT run its own static tools or a sandbox. It reads the diff + the base branch's .themegrill-qa/ (area map, test-case titles, knowledge) + other checks' results, makes ONE model call, and posts one advisory sticky comment.

• Never checks out or runs PR code (API reads only).
• Deterministic: areas touched (from suite.json), other checks' status.
• Verified, not trusted: a risk is shown only if its file:line is an added line in the diff; a suggested test only if the title exists verbatim in the index. Others are dropped and counted. No confidence number.
• Model output is stripped of URLs, @mentions, HTML, backticks; untrusted text can't close the prompt's data blocks; unparseable/empty model output fails the job instead of posting a clean-looking review.
• post_comment=false (default for manual runs) writes the result to the run summary only.

Testing

• npm test: 50/50 (12 new: glob/area mapping, line-number annotation, diff budget, sanitizer, citation verification, prompt-injection escaping, end-to-end with a fake model incl. stale head / empty summary / repo without QA data).
• Gather half run against the real PR (themegrill/user-registration-pro#1765) with a stubbed model: prompt about 94k chars (~24k tokens).
• NOT verified: a real model call (quality, cost; pricing.mjs has no row for the default model) and the comment posting path in a live run.

Notes

Uses the shared OPENAI_API_KEY. A dedicated spend-capped key is advisable before n8n makes this fire automatically. Model override: QA_REVIEW_MODEL repo variable.