[trinity-ability]

Fixes #2994

Problem

dependabot-auto-merge.yml's Wait for build + pytest to pass step greps every check-run named build or pytest* and treats any completed non-success conclusion as a failed gate. Since the #2945 tiering every PR carries pytest (base, inline fallback) and pytest (push, absolute failures) as completed / skipped (they run on other events), so the gate has answered a gating check failed on every Dependabot PR since mid-September — with build and pytest (head) green right beside them. The Dependabot queue filled with green patch/minor bumps nobody had looked at; today's merge train merged #3100, #2967, #2965 and #2830 by hand after re-running their flaky journey-smoke and reading the auto-merge log.

Fix

One predicate: skipped is exempt. Every other non-success conclusion (failure, cancelled, timed_out, action_required) still blocks.

Test

tests/unit/test_dependabot_automerge_skipped_gate.py extracts the step's run: from the workflow YAML and executes it under bash with gh stubbed to answer the rows — the script that runs on the runner is the script under test. With the fix reverted, exactly the skipped-rows case fails (verified in a python:3.13-slim container; the test skips on macOS's bash 3.2, which has no mapfile).

Lane C (.github/workflows/), not deploy-dev.yml. No behaviour change for majors — they are still held for a human by the eligibility step.

🤖 Generated with Claude Code

[vybe]

merge-train: on this train. I pushed one commit (7c0c3e745) and added Fixes #2994 to the body.

Why: pytest (head) runs its suite with || true and reports green whatever happens. The unit verdict is regression diff, which the ^build$|pytest filter never selected. With skipped rows no longer failing the gate, a patch/minor bump that breaks unit tests would have auto-merged.

What changed:

• The filter now also selects ^regression diff$.
• New test: test_the_regression_verdict_is_a_gating_check runs the workflow's own jq filter under real jq.

Not addressed here: the job is guarded on github.actor rather than the PR author's login. That's older than this PR.

[vybe]

Correction to the note above: this PR rides the next train (with 3148, 3157's sibling batch), not this one. The pushed fix and keyword stand.

[vybe]

merge-train: pushed a5497edbf. This is our own PR, so the finding was fixed here instead of sending it back.

Finding: regression diff has needs: [test, baseline, base-run], so GitHub creates its check-run only after pytest (head) finishes. In that window every gating row is green or skipped. The gate read "nothing pending" and passed without the unit verdict, and the PR's own test asserted green on exactly that row set.

Fix:

• The jq projection now carries the check name.
• A missing regression diff row counts as pending, bounded by the existing 30-minute ceiling.
• A poll where every row is skipped no longer counts as green.
• Two new tests use a stub that changes its answer between polls, so they prove the gate did not decide on the first poll.

The suite is 8/8 locally (bash 3.2 with a mapfile shim), and 8/8 fail against the previous workflow.

Not addressed:

• gh pr merge --auto merges a later head without re-gating it. --match-head-commit is one option.
• The 5-minute "no rows → exit 0" fail-open.

Both pre-date this PR.

[vybe]

merge-train: batch validated on train PR #3179 (all checks green).