[mybelovedstar123]

Fixes #6279

Test-only; the gateway is unchanged. This implements the fix the issue suggests.

Problem

In test_a_reply_waits_for_its_rooms_earlier_send_and_another_room_does_not, the test checked fake.calls == [b1] as soon as _post_ready_results returned. With DELIVERY_PASS_WAIT_S patched to 0.2 s, the pass waits for room B's worker only that long, then lets it "go on alone".

On a loaded runner the post lands after the deadline. The check then sees [], and the late send posts into the next test's FakeGateway, because setUp patches a new gw._req before the worker sends. That is what made test_a_result_that_posts_nothing_is_not_held_behind_its_rooms_send fail with a leading 'task-slow-b1'.

Change (tests/gateway-ordered-result-delivery.test.py)

• Poll for room B's post. Before the existing assertion, the check now waits with _until(lambda: self.fake.calls == [b1], 10). assertFalse(second.is_alive()), which proves the pass doesn't wait on room A, is kept.
• Wait for every send to settle. Harness registers a cleanup that waits until nothing is delivering (_DELIVERING and _DELIVERING_ROOMS empty, read under _DELIVERING_MUTEX) and fails if a delivery outlives its test. It is registered after mock.patch.stopall, so it runs before the patches are undone.

Evidence (WSL2)

The issue notes that forcing the pass deadline to 0.0 s reproduces both CI assertions every time. I ran each tree as committed, and again with DELIVERY_PASS_WAIT_S patched to 0.0 in a temporary copy of the test:

main, as committed (0.2 s):  Ran 12 tests  OK
main, forced 0.0 s:          FAILED (failures=2)  x3, the two failures in the issue
fix,  as committed (0.2 s):  Ran 12 tests  OK
fix,  forced 0.0 s:          OK  x3

The file takes longer, about 41 s locally against 29 s on main. The cleanup now waits for sends that used to run on into the next test.

hermetic-env-structural passes, and so do the static checks (ruff, py3.9 compat, every lint-* with --diff, review-checks.sh, src-map, agents-md-sync, workspace-leak).

Author: discord username :: slovertop :: agent-assisted (Claude)

[sonichi-dot]

Reviewed exact head 7deb72ac8a0318b50d46d1d1a4690c1b8b07865c. No actionable findings; LGTM, non-blocking.

Checked the complete diff, test harness, and production delivery bookkeeping. The cleanup runs before patch teardown, and waiting for room B's post preserves the room-order assertions.

Validation:

• python3 tests/gateway-ordered-result-delivery.test.py -v: 12 tests passed, including the four hash-seed subprocesses.
• Forced DELIVERY_PASS_WAIT_S = 0.0 in memory and ran the two affected tests: the parent test reproduced both failures ([] instead of room B's post, then the leaked room-B post in the next test); HEAD passed both.
• review-checks.sh --diff: passed.

Worst-case disruption is additional test waiting or a cleanup failure for a stuck worker; production behavior is unchanged. This is focused Linux validation, not a full local suite or cross-platform run. CI was still running at review time, so merge readiness remains subject to completed required checks and the repository's approval/CLA gates.

[github-actions]

Coverage Gate

✅ Diff coverage PASSES the 95% bar. Whole-tree (informational): 93%.

Diff Coverage

Diff: origin/main...HEAD, staged and unstaged changes

No lines with coverage information in this diff.

[sonichi]

Read 7deb72a (test-only, +9 -0). The room B check now polls up to 10 s for calls == [b1] instead of asserting the instant the pass returns, so a late post on a loaded runner no longer fails it. The assertFalse(second.is_alive()) line still proves the pass never waits on room A. The new cleanup waits until _DELIVERING and _DELIVERING_ROOMS are empty, read under their mutex. addCleanup runs last-in first-out, so it runs before mock.patch.stopall, and a straggling send posts into its own test's fake rather than the next one's. The room A release is in a finally, so the cleanup's 30 s wait only costs time on a real leak. No findings.

[liususan091219]

Approving 7deb72a (test-only, +9 -0, fixes #6279).

• Approve bar: @sonichi (owner) approved this exact head. No change request stands.
• What I checked: the room B assertion now polls _until(... == [b1], 10) before the exact-equality check, and assertFalse(second.is_alive()) still proves the pass does not wait on room A. The cleanup is registered after mock.patch.stopall, so it runs first and waits on _DELIVERING/_DELIVERING_ROOMS under _DELIVERING_MUTEX. A late send therefore posts into its own test's fake, not the next one's. The test still imports and drives the real gateway, with no copy of the logic.
• Evidence: the body's forced DELIVERY_PASS_WAIT_S=0.0 run reproduces both CI failures on main and passes with the fix. @sonichi-dot reproduced that independently.
• CI and CLA: 43 checks pass, 1 skipped. license/cla success. No unresolved threads.

— Sutando (susans-MacBook-Pro)