[robhogan]

Without Watchman, a watch-mode start has to be watching before it crawls, so that a change made during the crawl isn't missed. NativeWatcher (macOS) isn't live as soon as it starts, so Watcher also confirms each root's watch with a health check before the crawl begins:

metro/packages/metro-file-map/src/Watcher.js

Lines 375 to 382 in ef50b94

	// Confirmed only once every backend has started, because starting one may
	// restart the others' watches (NativeWatchers share one FSEvents stream).
	this.#confirmWatches = !WatcherImpl.isLiveOnStart();
	if (this.#confirmWatches) {
	await Promise.all(
	this.#options.roots.map(root => this.#waitUntilLive(root)),
	);
	}

That takes about 13ms on macOS, and build() only starts it once the cache has been read, although the watch doesn't need anything from the cache.

This starts the watch, and the Watchman capability check (which spawns watchman), while the cache is read. The crawl still waits for the confirmed watch, so nothing can be missed that wasn't before. To allow this, Watcher takes the cached state in crawl() rather than at construction.

Benchmark

This takes 13% off a watch-mode start on macOS (102ms -> 88ms). Linux and Windows watch with FallbackWatcher, which is live as soon as it starts, so they're unchanged, as are starts outside watch mode. With a cold cache there's no read to overlap with.

	main	#2220	This PR	vs main	vs #2220
macOS: Watch-mode start	130ms	102ms	88ms	-32.1% (-32.8% to -31.2%)	-13.0% (-14.0% to -11.3%)
Linux: Watch-mode start	1,889ms	493ms	491ms	-74.0% (-74.2% to -73.6%)	-0.5% (-1.4% to +0.5%)
Windows: Watch-mode start	2,569ms	524ms	524ms	-79.6% (-79.7% to -79.5%)	+0.1% (-0.5% to +0.6%)
macOS: Warm start	118ms	88ms	88ms	-25.1% (-26.4% to -23.6%)	+0.1% (-1.8% to +1.9%)
Linux: Warm start	636ms	353ms	353ms	-44.5% (-45.3% to -44.1%)	-0.1% (-1.8% to +1.1%)
Windows: Warm start	355ms	201ms	201ms	-43.4% (-43.7% to -42.7%)	+0.0% (-0.9% to +0.7%)

Changelog

 - **[Performance]**: metro-file-map: Start watching while the cache is read

Test plan

New unit tests. Benchmark as in #2104, with 22 rounds on macOS.