A test that went offline too early
The first red Verify on main since the site opened came from the test of a stalled worker update. The update was meant to give up after 45 seconds and failed after 122 milliseconds. The test was at fault, not the worker. It cut the network while the new worker was still fetching files of the shell, and the first such file failed. The test now waits for those files first. The failure reproduces on the old test and is gone from the new one.
What failed
Verify on 3230ae6 went red, though the commit only changed names in the data. One spec in pwa.spec.ts failed: an update whose shell file never answers must give up by SHELL_WAIT, 45 seconds. The test requires at least half of that to pass, and 122 milliseconds passed. The next run, on 9acd8f5, was green, and so were the four before. The spec is not broken, it is flaky.
What the run’s trace showed
The trace shows the worker’s own requests. The new build fetches the shell of the Ukrainian axis. The test holds the Timeline file on purpose, and it hung as meant. But the worker asked for react-dom and relations only after the test had cut the network to open the axis offline in another tab. Both failed with ERR_INTERNET_DISCONNECTED, so the install ended by the network, not by the limit. The worker did what it should: a file that does not come fails the install. The fault was in the test, which cut the network before the worker had asked for everything else.
The test used to pass because the worker usually managed it within the 80 milliseconds the test spent opening the second tab. On the runner of 28.09 it did not.
The fix
The test now records every request of the worker that finishes. It cuts the network only once every file of the shell but the held one has arrived and the held one has reached the gate. The worker itself is unchanged, and the limit was not raised.
First I reproduced the failure. A probe held react-dom for up to three seconds and failed it as soon as the test went offline, as Chromium failed it in CI. The old test failed as it had in CI: 78 milliseconds instead of more than 22,500. The new one waited for the file and passed. The probe did not go into the repository.
What was checked
Locally in Chromium the spec passed three times in a row. The rest of pwa.spec.ts passed too, but for the kill-switch test: it builds the site into a temporary folder, which on this computer sits on another drive than the repository, and Windows will not move files there. That has nothing to do with the fix, and the test passes in CI. CI sees the fix with this push. I did not re-run the red run on 3230ae6: that happens only on the 6-in-1’s command.
Entry written September 28, 2026
Commits this entry accounts for
2b047a5