[robhogan]

Without Watchman, the node crawler lstats every file on every start, so that getDifference can tell what changed while Metro wasn't running:

metro/packages/metro-file-map/src/crawlers/node/index.js

Lines 83 to 96 in 56efab3

	activeCalls++;
	fs.lstat(file, (err, stat) => {
	activeCalls--;
	if (!err && stat) {
	result.set(childNormal, [
	stat.mtime.getTime(),
	stat.size,
	0,
	null,
	isSymbolicLink ? 1 : 0,
	]);
	}
	resolveIfDone();

Most of those files are never read, and an mtime only matters for a file with something derived from it (a SHA-1 or plugin data) - that's all a change can invalidate.

This takes the idea in @kitten's #1686 and makes it hold across restarts. A regular file is crawled without a stat, and its mtime is recorded when data is first derived from it, from a stat taken before it's read. A file with no mtime then has nothing that could be stale, so a start only stats the files that had an mtime before.

• A recrawl while watching still stats every file, because consumers like PackageCache hold contents read outside the file map.
• Symlinks are still read eagerly - every lookup through a link needs its target.
• linkStats() gives a null modifiedTime and size for a file that hasn't been stat'd.
• CACHE_BREAKER is bumped, or an existing cache would keep every file stat'd.

Benchmark

Without Watchman, this takes 40-45% off a cold or warm start on Linux and Windows (Linux warm: 649ms -> 371ms), 9-15% off on macOS, and 4-7% off RSS. Watch mode gains less on Linux and Windows, where setting up the watcher dominates - the rest of this stack deals with that. The cost is a stat when a file is first read: 10-20% more on a first SHA-1 pass over the bundle's files (8-47ms here), once per file per cache.

	main	This PR	vs main
macOS: Cold start	156ms	132ms	-15.4% (-23.3% to -8.7%)
Linux: Cold start	644ms	360ms	-44.0% (-45.7% to -43.3%)
Windows: Cold start	369ms	208ms	-43.7% (-46.6% to -42.7%)
macOS: Warm start	142ms	129ms	-9.4% (-10.7% to -8.3%)
Linux: Warm start	649ms	371ms	-42.8% (-43.3% to -42.5%)
Windows: Warm start	357ms	225ms	-37.2% (-37.7% to -36.7%)
macOS: Watch-mode start, warm cache	143ms	130ms	-9.6% (-10.6% to -8.7%)
Linux: Watch-mode start, warm cache	1,896ms	1,617ms	-14.7% (-15.9% to -14.0%)
Windows: Watch-mode start, warm cache	2,554ms	2,429ms	-4.9% (-5.4% to -4.3%)
macOS: First SHA-1 of the bundle's 3,125 files	75ms	83ms	+10.4% (+8.6% to +11.8%)
Linux: First SHA-1 of the bundle's 3,125 files	110ms	131ms	+19.5% (+17.2% to +20.5%)
Windows: First SHA-1 of the bundle's 3,125 files	270ms	317ms	+17.2% (+13.7% to +19.5%)
macOS: RSS after a warm start	142.6MB	132.1MB	-7.4% (-7.9% to -6.0%)
Linux: RSS after a warm start	175.1MB	164.1MB	-6.3% (-6.9% to -6.0%)
Windows: RSS after a warm start	149.6MB	144.0MB	-3.8% (-4.1% to -3.3%)

Changelog

 - **[Performance]**: metro-file-map: Start faster without Watchman, by not stat-ing files whose contents Metro hasn't read

Test plan

New unit tests, and a stats-between-starts-test.js that edits files between two runs of FileMap over a real directory and cache. SHA-1s, eager and lazy plugin data and a retargeted symlink are recomputed, and an unread file is stat'd only once it's read. These include the cases #1686 gets wrong.

Benchmark app (AI-driven harness, used throughout this stack): a React Native 0.87.1 app with 23 common dependencies, 34k files in the file map and 3.1k modules in its iOS dev bundle, started through Metro's DependencyGraph with resolver.useWatchman: false. Each round runs, per variant: 2 cold starts, a watch-mode session that computes the SHA-1s of the bundle's files (as transforming them does) and saves them to the cache, 3 warm starts and 3 watch-mode starts from that cache, and 2 watch-mode cold starts. 10 interleaved rounds per OS, alternating the variant order, then medians and bootstrap 95% CIs. Heap is measured after a forced GC. Machines: Mac mini M4 Pro (macOS, APFS, Node 26.11.1), Mac mini i5-4278U (Linux Mint 22.3, ext4, Node 24.21.0) and a Ryzen 7 7730U laptop (Windows 11, NTFS, Defender on, Node 24.21.0).