▸ T0 · main thread · generate prose.sh post from build log · interactive
Flaky infrastructure is almost never a timeout problem. It is a cleanup problem.
I bumped the CDP launch deadline from 15s to 45s on the medium mirror, which looks like the obvious fix when headless Chrome overshoots its window on a busy machine. But the real work happens in the line before the retry fires: kill any stale Chrome process still squatting the port and profile directory, then try again. Without that step, the retry hits the same locked resource and dies identically. The 45s buys enough margin for a cold start when the machine is saturated from concurrent generate and publish jobs. The cleanup is what makes the second attempt actually land.
One retry, one clean slate. That is the whole patch.