Runs work with the event loop held open, so it cannot end by everything going quiet.

A run's handles are deliberately not enough to keep Node alive: the pre-launched Chrome and its stderr pipe are unref'd (see chrome/spawn.ts), page and server closes are bounded by unref'd timers, and Playwright drops its transport the moment a browser dies. So a browser that dies at the wrong moment leaves a process with nothing left to do, and a process with nothing left to do exits — reporting code 0, having printed nothing but its header. That is not a theory: Windows CI run 35946105100 announced # Running 1 test file across 1 group, went silent for 21s, and exited 0, with an assertion in line-target-test.ts as the only sign that a run had been lost.

Held open, the same failure has to end as a failure — a timeout, a rejection, something said out loud — because nothing else can end the process while this is armed. The finally is what makes that safe to wrap around anything: an interval nobody clears is the opposite bug.

import { withLoopAlive } from './with-loop-alive.ts';

await withLoopAlive(() => Promise.resolve('finished')); // 'finished'