skills: gate sweep-found fixes on reachability, and stop tests pinning undefined behavior

[tend-agent]

A nightly sweep on PRQL opened PRQL/prql#6250 for a real vector-sizing mismatch that no in-repo caller can trigger, and the collaborator rejected it: "While the change might fix theoretical future problems, it's not fixing a current problem." Nothing in the bundled skills asked the sweep to check that. nightly's review checklist leads with "Bugs, logic errors, unhandled edge cases" and carries no reachability test; running-in-ci's Weighing a Fix raises the bar only for compute-saving changes, and triage's reproduction gate never applies because a sweep has no report to reproduce. Reported by the PRQL bot in #1133.

Three lines of guidance, no new machinery:

• nightly — a gate on the checklist's bug line. Before a sweep-found defect becomes a PR, name the in-repo call path that reaches it. If none does, the finding is a note in the Step 9 summary. public visibility doesn't clear the bar on its own; a documented public surface on a published library does, and the PR body should name the contract rather than the keyword.
• triage — a fix-quality gate beside the existing "Don't 'fix' tests by adding skip guards". When a doc claim and the code disagree and which is wrong is still open, the finding is that question; a test pinning the current output turns unspecified behavior into a contract, so the eventual fix arrives looking like a regression. This is the second finding on the same PR: the bot answered a doc/code mismatch by asserting current output in a test, and the reviewer flagged it.
• review — the reviewer-side counterpart under Testing, one line: don't suggest that test either. Distinct action from the triage gate (what to suggest vs. what to write), and a wrong inline suggestion is an outward action, so it's worth the line rather than relying on the author-side gate alone.

Verification is pre-commit on the three files (all hooks pass, including the bang-backtick and install-tend mirror guards) plus generator/tests/test_cross_file_contracts.py (15 passed). No regression test: nothing in the suite asserts skill prose, and these are prompt-text changes whose effect shows up in adopter run outcomes, not in CI.

---

Closes #1133 — automated triage

[tend-agent]

The three additions read cleanly against the flows they sit in — the triage section fits beside the existing skip-guard gate, and the review line is a distinct enough action (what to suggest vs. what to write) to be worth its own line rather than a duplicate. One inline suggestion on the nightly gate.