fix(running-in-ci): dismiss a standing approval invalidated by another PR merging

[tend-agent]

Every dismissal site tend has is keyed on an event on the approved PR — a review round (review step 6), a rewrite (weekly step 2), a red check (review step 7) — so an approval invalidated from outside the PR, most often by a different PR merging and superseding it, reaches none of them and stands until a human clears it. This adds one generic rule in running-in-ci, the skill every session that can reach that conclusion already loads, keyed on the conclusion rather than on posting.

The reported case (#1138) is a dependency PR that a later merge turned into a downgrade: two tend-nightly sessions stated the supersession in their own summaries and left the bot APPROVED standing, correctly declining to post a second deferral comment — which is exactly why a rule keyed on a posting site never fired. Merging was blocked there only because the branch happened to be conflicting; a supersession that doesn't touch the same manifest lines leaves the PR mergeable under a stale bot approval.

Details

Placement. One site rather than a copy in nightly and resolve-conflicts: both load running-in-ci (nightly Step 0, resolve-conflicts line 10), as do weekly, review, triage and the mention paths, so the rule is in context wherever the conclusion is reachable. It also keeps bot-review-state.sh as the single mechanism — the recipe reads the existing standing_approval_id field from #1136, which reads "" once cleared, so a re-run dismisses nothing.

Scope guard. The rule states the test as the merge, not tidiness, so it can't be read as "dismiss whenever you have a finding" — that would withdraw verdicts the code still earns, which is the failure #1136's own scoping sentence guards against.

Test. test_a_dismissal_path_exists_for_an_invalidation_that_is_not_an_event in generator/tests/test_bot_review_state.py, alongside the existing dismissal contract tests. It fails on main (the skill contains no standing_approval_id reference) and passes with the change. Full uv run pytest (916 passed) and pre-commit run --all-files are green.

---

Closes #1138 — automated triage

[tend-agent]

Self-review — the gap is real and the placement is right: nightly (Step 0) and resolve-conflicts (line 10) both load running-in-ci, so one section reaches every session that can reach the conclusion, and standing_approval_id reads "" after a dismissal (the jq keeps only APPROVED/CHANGES_REQUESTED, and a dismissal rewrites the state to DISMISSED), so a re-run is a genuine no-op. The test's four assertions are all absent from the base file, so it fails on main as claimed.

Two things in the prose, both inline.

One observation that isn't for this PR: this is the third copy of the same standing_approval_id → dismissals recipe (review step 6, review step 7, now here), and review loads running-in-ci, so those two could become pointers to this section plus their own trigger conditions. Worth folding into whichever PR next touches review, rather than growing this one.