ce-babysit-pr review-gate evals test the reasoning but not the lookup

The gap

ce-babysit-pr's readiness decision now rests on a judgment the agent makes from live GitHub state: whether a review is still on its way. skills/ce-babysit-pr/references/settle.md tells it to make one lookup across four surfaces — reactions on the PR body, top-level comments, check runs on the current head, and reviews against it — and reason from what comes back.

Five eval cells cover that reasoning (tests/skill-eval-cell/catalog.ts, ids starting ce-babysit-pr/). Every one of them hands the agent a prepared description of what the lookup would have returned and forbids gh. They are read_only: true by design, which is what makes them cheap and deterministic.

So the cells validate the semantic classification — a terminal check that accounts for the announced review clears the wait, a timed-out one does not read as approval, an unrelated check is not the review finishing, a silent reviewer of an earlier head still waits. They validate none of the acquisition:

• Head attribution. Reactions and top-level comments are PR-scoped and carry no commit identity; check runs and reviews are commit-scoped. Nothing states how a PR-scoped announcement is bound to the current head.
• Identity correlation. Bot logins differ across APIs — REST returns cursor[bot] and chatgpt-codex-connector[bot], GraphQL returns cursor and chatgpt-codex-connector. A lookup that assumes one shape silently misses reviewers.
• Pagination. No stated behavior for a PR with more reactions, comments, or check runs than one page.
• Probe failure. No stated behavior when part of the lookup fails, which is the case most likely to produce a false ready.

Why it matters

This is the seam most likely to fail in production, and it is the one place the current tests say nothing. Two independent cross-model reviews of the change raised it separately, which is the reason it is filed rather than argued.

Direction, not a prescription

A cell that permits gh against a real fixture PR would exercise acquisition, but it trades determinism for realism and needs a stable subject. An alternative is to keep the reasoning cells as they are and add deterministic coverage of the correlation rules alone — the [bot] suffix variance and the head-attribution rule are mechanical and belong in bun test rather than an eval.

Worth deciding which before adding either.

Context: #1611.