Follow-ups from the review of #747 (#737). None blocks the merge (maintainer's two-round cap, 2026-09-27). The review reproduced every number in the body (CI shape: main 5 FAIL / 23 of 25 subtests; head -count=3 18 top-level and 75 subtests PASS; revert 5 FAIL; merged with #746 on 698bed6, root 382/0/1).
Handled in round 2, no code change:
Open (test text and process):
- The
/proc/stat test comments overclaim isolation. start_failure_release_linux_test.go:24 says that counting only /proc/stat links means descriptors other tests open or close "cannot move it". Every celeris server in the same binary opens and closes a /proc/stat monitor in doPrepare. The tests are safe only because they are non-parallel top-level tests. Reword the comment, or it invites a t.Parallel later.
after != before should be after > before (:64, :129). A monitor closed by something else in the window, for example a finalizer queued before SetGCPercent(-1), would fail the check. The leak direction is >. No flake observed: main's first subtest starts at 2 and ends at 22, and every run passes at the head.- The dial assertion in
TestFailedStartClosesTheSuppliedListener cannot fail. The test closes the listener itself (start_failure_release_test.go:79, the ErrClosed check) before it dials (:82), so the dial is refused whether or not the fix closed it. It could only catch a surviving duplicate of the fd. Under the revert arm, left the supplied listener open fired 12 times and still connected 0 times. Dial first, or drop the claim from the test's doc.
- The fd-delta probe is not committed. The body's scenario table comes from
737/probe/ in the lane's evidence root, which is not version-controlled, so only the maintainer can re-run it. The committed tests reproduce the /proc/stat half. Either commit the probe (as a test, or in probatorium) or mark the table maintainer-local.
Follow-ups from the review of #747 (#737). None blocks the merge (maintainer's two-round cap, 2026-09-27). The review reproduced every number in the body (CI shape: main 5 FAIL / 23 of 25 subtests; head
-count=318 top-level and 75 subtests PASS; revert 5 FAIL; merged with #746 on 698bed6, root 382/0/1).Handled in round 2, no code change:
breakinglabel (minor): a failedStartWithListenernow closes the caller's listener, so reusing that listener for a fallback server fails. fix(server): Shutdown runs the OnShutdown hooks, and returns, only after the drain on every engine (#703) #746 carriesbreakingfor a behaviour-only change, as fix(server): shut down on every context cancel, and return only after it (celeris#673) #692 did, so fix(server): a Start that never serves releases the caller's listener, the CPU monitor and the settle re-opener (#737) #747 now carries it too. The body's "Behaviour change" section already describes the change.TestAdaptiveSettledRouteRetime592cannot tell the two sides apart when the rig misses about 2^-7 of runs. The attribution is the failing run's own log: its reference window saw/pingfast, so the rig's premise did not hold. The body now says that.Open (test text and process):
/proc/stattest comments overclaim isolation.start_failure_release_linux_test.go:24says that counting only/proc/statlinks means descriptors other tests open or close "cannot move it". Every celeris server in the same binary opens and closes a/proc/statmonitor indoPrepare. The tests are safe only because they are non-parallel top-level tests. Reword the comment, or it invites at.Parallellater.after != beforeshould beafter > before(:64,:129). A monitor closed by something else in the window, for example a finalizer queued beforeSetGCPercent(-1), would fail the check. The leak direction is>. No flake observed: main's first subtest starts at 2 and ends at 22, and every run passes at the head.TestFailedStartClosesTheSuppliedListenercannot fail. The test closes the listener itself (start_failure_release_test.go:79, theErrClosedcheck) before it dials (:82), so the dial is refused whether or not the fix closed it. It could only catch a surviving duplicate of the fd. Under the revert arm,left the supplied listener openfired 12 times andstill connected0 times. Dial first, or drop the claim from the test's doc.737/probe/in the lane's evidence root, which is not version-controlled, so only the maintainer can re-run it. The committed tests reproduce the/proc/stathalf. Either commit the probe (as a test, or in probatorium) or mark the table maintainer-local.