docs(releases-maintenance): before approving release, wait until the newest scheduled CI at or before the version commit has completed — 18.0.0 was approved on a green 9 commits back while the nearer run was in flight (it went red) #24889

Filing gate ③: a task the maintainer set (「如何避免类似的问题」, then 「立卡」). Filed by the triage seat (#6015, session_01AavokzJ5DndAwitDXvKy4U). ⛔ Not a claim.

Reader (who acts): the domain:devx seat, in one docs PR on docs/releases-maintenance.md ("Cutting a release", about :297). ⛔ No code, no gate change, no merge freeze.

Measured (18.0.0, version commit 06cab4849a)

• What the gate read: the release gate (job 118623776476, 12:22–12:23Z) found the nearest completed CI run to be 39412016980 on abd5acdaf7, 9 first-parent commits before the version commit.
• What it could not see: the nearer scheduled run, 39419131283 on 81120caac8 (2 commits before), was still in progress, and the gate never waits (by design). That run then concluded failure at 12:30Z (hourly full run: red on main (CI) #24879, a stale dogfood fixture).
• The approval: release was approved, and 18.0.0 published on the older reading.
• The confirmation: the later force: true dispatch (run 39427630088) reads that run and refuses ("CI: the newest full-battery run at or before the release commit concluded failure"). So the gate works on completed readings. The gap is only the window in which a nearer run is in flight.

Done when

• "Cutting a release" adds one maintainer step before the release approval. Read the Release run's Full battery green at the version commit line, and if a scheduled CI run at or before the version commit is still in progress, wait for it to complete (about 25–30 min) before approving.
• The step says why: the gate passed on an older green, and the in-flight run may be red.
• The step names what it is not: not a merge freeze. The snapshot flow (release: snapshot one-click rc cut — a cut-rc workflow that bumps, versions, gates and publishes from ONE SHA, ending the two-treadmill freeze dance #7447, :350–353) chose against coordinated freezes. A Version Packages PR merged behind main is already handled by annotate-shipped (release(next): 18.0.0-next.2 shipped 11 changesets with no CHANGELOG entry (the version PR merged behind main; #21361's class, in pre mode) #23818, PR fix(release): refuse a Version Packages PR whose queue build lands behind main; mark next.2's 11 missed changesets #23833), and that stays as is.

Dedupe

objectstack issues since 2026-09-20 were grepped for freeze or 冻结 near cut/release, and for "stale version PR". The hits are #23818 (closed: the CHANGELOG half, now automated), #24856 (open: H16's premise about changeset-release/*, a different instrument) and #24445 (closed, unrelated). Nothing covers the approval reading. New.

Generated by Claude Code