Attempt to fix ENAMETOOLONG error by filtering and deduping open_basedir

[fredrikekelund]

Related issues

• Fixes STU-2038

How AI was used in this PR

I used Claude to iterate in relatively small chunks towards a result I liked. I used Codex to review the result.

Proposed Changes

In #4173, @Tropicalista reported an issue where native PHP sites with "site directory" file access would fail to start and report an ENAMETOOLONG error.

ENAMETOOLONG errors indicate that the CLI command was too long for the OS to handle (there is such a limit). We don't know what exactly made the command so long, but our best theory is that it's the open_basedir directive, which is passed as a CLI option. It could have happened if there was a pnpm-style node_modules directory anywhere in the site directory, for example (since this means there can be hundreds of symlink paths in the open_basedir directive).

This PR fixes the problem by applying the following changes:

1. Ignore node_modules, .git, and .DS_Store paths in findSymlinksInDir(). Previously, we only ignored those paths when watching for new symlinks. If a pnpm-style node_modules dir is the source of the problem, then this is the key fix.
2. If the open_basedir list contains /lorem/ipsum and /lorem/ipsum/dolor, /lorem/ipsum/dolor is now dropped before passing the list to PHP (because it is redundant). This change also ensures that newly added symlinks pointing to an already allowed directory don't trigger a process restart.

The next logical step would be to put the open_basedir directive in a site-specific php.ini file. I'll leave that for another PR, though.

Testing Instructions

1. Download a WordPress plugin like performance-lab and unzip it
2. Have a Studio site
3. Start that site
4. cd into wp-content/plugins for that site
5. ln -s PATH_TO_DOWNLOADED_EXTRACTED_PLUGIN_DIR
6. Open wp-admin
7. Click Plugins in the sidebar
8. Ensure that the plugin (Performance Lab in my example) is available to activate

This routine ensures that basic symlink watcher functionality and open_basedir compilation still work as expected. It does not test the ENAMETOOLONG error specifically, as we don't know exactly how to reproduce that.

Pre-merge Checklist

• Have you checked for TypeScript, React or other console errors?

[wpmobilebot]

📊 Performance Test Results

Comparing dc8b98c vs trunk

app-size

Metric	trunk	dc8b98c	Diff	Change
App Size (Mac)	1402.88 MB	1402.78 MB	0.10 MB	⚪ 0.0%

site-editor

Metric	trunk	dc8b98c	Diff	Change
load	1062 ms	1036 ms	26 ms	⚪ 0.0%

site-startup

Metric	trunk	dc8b98c	Diff	Change
siteCreation	7532 ms	7536 ms	+4 ms	⚪ 0.0%
siteStartup	2872 ms	2865 ms	7 ms	⚪ 0.0%

---

Results are median values from multiple test runs.

Legend: 🟢 Improvement (faster) | 🔴 Regression (slower) | ⚪ No change (<50ms diff)

[bcotrim]

Logic makes sense, and the testing instructions work as described.

I added 2 issues flagged by Claude that I couldn't quite verify, take it as non-blockers.

[bcotrim]

split( path.sep ) never matches on Windows — chokidar normalizes to forward slashes before calling ignored (chokidar/index.js:44-48). Verified on Win11: it receives node_modules and returns false, where trunk's regex returned true. So the watcher descends into the link farms it means to skip.

Suggested change

	return entryPath
	.split( path.sep )
	.some( ( segment ) => IGNORED_SCAN_DIRECTORY_NAMES.has( segment ) );
	return entryPath
	.split( /[\\/]/ )
	.some( ( segment ) => IGNORED_SCAN_DIRECTORY_NAMES.has( segment ) );

[fredrikekelund]

Good catch 👍

[bcotrim]

Runs the full O(N²) dropCoveredPaths() per symlink event, and arePathsEqual does 2 statSync per comparison — 50 entries ≈ 19ms / 22k sync stat calls, on the same loop as the proxy (:460). Trunk was an O(1) Set.has.

The dedup is redundant here: if a dropped entry covers the target, whatever displaced it does too.

Suggested change

	return getEffectiveOpenBasedirAllowlist().some( ( entry ) => containsPath( entry, target ) );
	return [ ...staticOpenBasedirAllowlist, ...symlinkOpenBasedirAllowlist ].some( ( entry ) =>
	containsPath( entry, target )
	);

[fredrikekelund]

Good feedback 👍