/review blindspot on non-application code PRs

I tried `/review` today on a CI change and it missed obvious issues - a Windows `.bat` script was added to a repo where all existing scripts are bash + Makefile, and the CI runners are all Linux.

A human would catch this and a regular claude-code prompt did too.

Overall the `/review` skill seems quite narrow. It runs through the checklist (SQL safety, race conditions, LLM trust boundary, enum completeness) - all categories return N/A - and the structured process creates false confidence that the review was thorough.

**Suggestion**: Add a triage step that classifies the PR type (application code, CI/infra, tooling/scripts, docs-only) and selects relevant review categories. Repo convention checks (existing patterns, tooling choices, naming) should run regardless of PR type. For non-application PRs, add checks for script quality and platform consistency instead of only running the application-code checklist.