[Paul-Browne]

state.visited is an array that every crawled directory is pushed into, and isRecursiveUsingRealPaths scans it with Array#includes on every directory symlink. That makes withSymlinks() crawls (default resolvePaths: true) O(directories × directory symlinks).

The array is also filled on every crawl, even though it is only read when useRealPaths is set.

Change

• visited is now a Set<string>, so the recursion check is O(1).
• It is only allocated when options.useRealPaths is set. Otherwise it is undefined, so default crawls no longer keep a second copy of every directory path.
• WalkerState.visited changes from string[] to Set<string> | undefined. This type is exported, but it's internal walker state.

Benchmarks

Node 22, Linux, warm page cache. Trees have N directories with one file each, plus N/2 symlinks to those directories. Crawled with new fdir().withFullPaths().withSymlinks().crawl(root).sync(), median of 5 runs.

N dirs	before	after
2,000	42 ms	32 ms
10,000	570 ms	147 ms
30,000	3,632 ms	425 ms

Default crawls without symlink resolution (200k files / 10k dirs) no measurable change.

[Paul-Browne]

Related issues (none are fixed by this PR, since output is unchanged):

• withSymlinks() spends most of its time in fs.realpath: would you take a faster resolver, opt-in or in the next major? #174: complementary. That issue speeds up the fs.realpath calls in withSymlinks(); this PR removes the O(directories × dir symlinks) visited.includes() scan that runs after them. On small trees like the ones in #174 the scan is minor, but it grows quadratically (8.6× faster at 30k dirs, see the benchmarks above). Once realpath gets cheaper, it'll also be a bigger share of what's left.
• symlinks can cause duplicate entries #138: groundwork. The fix proposed there checks visited before every directory traversal. With an array, that would make every symlink crawl quadratic; with a Set, it's O(1). The Set is allocated in exactly the configuration #138 hits (withSymlinks() defaults to real paths). Happy to follow up with a fix for #138 as a separate PR, since it changes output.

Heads up: this conflicts with #170 in __tests__/symlinks.test.ts. Both PRs add a fixture in the same place. The source changes merge cleanly, so whichever lands second just needs a quick rebase.