fix: coalesce concurrent watch-mode restarts, deflake the pool stderr fixture - #10822

Analysis of the recent CI failure increase across main and PR branches (runs from Jul 21-23). Two of the recurring clusters are root-caused and fixed here; the rest are catalogued below with evidence.

Fixed

1. Windows/macOS e2e: unhandled TypeError: Cannot read properties of undefined (reading 'logger') (main runs 30009378686, 29992865327, 29926056935; attributed to file-watching.test.ts)

The stack is _restart → report('onServerRestart') → VerboseReporter.log → this.ctx.logger. _start re-creates the reporters (this.reporters = await createReporters(...)) but ctx is only assigned later by the onInit report. Chokidar regularly fires several change events for one config edit; a second _restart beginning inside that window reports onServerRestart to constructed-but-uninitialized reporters. Fix: requests arriving during an in-flight restart coalesce into one trailing restart (which re-resolves the config from scratch, so it picks up whatever changed meanwhile). Regression test drives three concurrent _restart calls and verifies the instance still reruns.

2. e2e pool.test.ts > can capture worker's stdout and stderr (main runs 29805385576, 29923860728; also on unrelated PR branches)

The fixture triggers a MaxListenersExceededWarning and the test asserts it in captured stderr, but process.emitWarning prints on a later tick and the test body was synchronous — a fast worker teardown exits before the warning reaches stderr. The fixture now awaits the warning event plus one tick.

Catalogued, not fixed here

• Browser CDP/mock race family — cdp.test.ts > cdp sends events correctly, mock-importActual (browser+coverage), mocking.test.ts > mocking dependency correctly invalidates it on rerun, 1_mocked-on-watch-change, and runner.test.ts > tests are actually running (fails on a stderr-must-be-empty assertion polluted by the cdp failure). One family: module mocks and CDP events race the module fetch because delivery is CDP-only. Needs dedicated work; a naive interceptor-fallback was already tried and hangs.
• pool-worker-exit.test.ts > worker death on a shared runner does not skip coverage finalization (macOS, main 30010040100) — after two worker crashes, 3-crash.test.ts/4-third.test.ts are entirely absent from the reported tree, i.e. the queued files were never scheduled onto a replacement worker before the run finalized. This is a crash-recovery scheduling race in the pool, not test timing — needs its own investigation.
• Browser hooks-timeout.test.ts — hook-timeout assertions sporadically observe Playwright's own locator.click: Timeout 249ms exceeded instead of the expected hook timeout; timing-margin dependent.
• import-durations.test.ts > should handle tests with no imports gracefully (main 29825620887) — exit code 1 instead of 0; the captured log doesn't include the cause.

None of the failures on the three split PRs (#10820, #10821, #10710) are introduced by their changes — each maps to one of the clusters above, all of which reproduce on main or on unrelated branches.