FileSystemRouter: take each route's path from Resolver::real_path_of, not from Entry.abs_path - #44612

Decisions for a maintainer

1. This goes one step past Remove get_fd_path: derive paths from cwd and what was opened, not from fds #38365, which applies the same path rule but still reads and writes Entry.abs_path.
2. A mis-spelled dir on a case-insensitive file system: filePath follows dir, as Bun.resolveSync does. The old value had the spelling on disk. If the same file is also imported under the spelling on disk, import(filePath) loads a second copy of the module.
3. A new router reads the directory listing that the resolver cached and can be older than the disk. This was so before, but a deleted route file then made the constructor throw. reload() reads the disk.

Settled on main: a dangling symlink

Main skips a dangling symlink in the router and in the resolver since #44968 (EntryKind::Dangling, from #38756), and the base of this PR includes that commit. The router skips the entry before Route::parse runs, so this PR does not change what happens to the link. One new test pins that the router lists the link only if the resolver resolves it. It passes on the base too.

#34614 has a case that expects ENOENT opening route for a dangling symlink. Main no longer reports that error for a dangling link, and this PR removes it for every file.

What still depends on order

filePath is now independent of Entry.abs_path. It is the resolver's real path, so it follows the resolver in two cases:

• After a case-mismatched import, the resolver loads the file under the import's spelling, and the router reports the name on disk (related to Error importing same name .ts and .tsx modules with different case #22686 and resolver: reject case-mismatched extension probes that do not exist on disk #36054).
• Under --preserve-symlinks, the resolver returns the link path before a Bun.build and the real path after one, because the bundler's resolver fills DirInfo.abs_real_path in the shared cache (Make --preserve-symlinks and --preserve-symlinks-main work in the runtime and bundler #36757). The router reports what the resolver returns in both states.

A directory symlink that is created or retargeted after its parent was listed gives the old answer until the cache is busted. The router stays a reader of the resolver's cache here.

Reproductions on released 1.4.3 (Linux x64)

// <root>/link -> <root>/real, pages = <root>/link/pages
Bun.resolveSync(pages + "/b.tsx", root); // or Bun.build({ entrypoints: [pages + "/b.tsx"] })
const router = new Bun.FileSystemRouter({ dir: pages, style: "nextjs" });
router.match("/a").filePath; // <root>/real/pages/a.tsx
router.match("/b").filePath; // <root>/link/pages/b.tsx, src "../../link/pages/b.tsx"

• Plain dir with pages/c.tsx linked to ../shared/e.tsx: after a resolve the router reports <root>/pages/c.tsx, and reload() changes it to <root>/shared/e.tsx.
• No symlink: Bun.resolveSync(pages + "/B", root) (wrong letter case) makes the router report <root>/pages/B.tsx, a file that does not exist on Linux.
• Reverse: proj/tsconfig.json sets jsxImportSource: "custom-src" and proj/link points outside proj. Bun.build of proj/link/pages/b.tsx imports custom-src/jsx-dev-runtime. With a router over proj/link/pages built first it imports react/jsx-dev-runtime.
• --preserve-symlinks: the router reports the real path with src ../../real/pages/a.tsx, and Bun.resolveSync of the symlinked path returns the real path after the router ran.

No user has reported these. The bug was found by a read of the code.

Measurements (debug builds of the base and of this PR, 100 and 1100 route files, per-route difference)

The three syscalls are a few microseconds per route. They are not the reason for the design. DirnameStore is never freed and has a fixed number of slots, so one slot per route per load doubles the number of reloads before it is full. A route below a symlinked subdirectory, and a symlinked route file, still take a slot for the name.

strace, perf, valgrind and bloaty are not installed here. The syscall count comes from a small ptrace counter and the append count from gdb breakpoints on BSSStringList<4096, 129>::do_append. Binary size is not measured: the change removes code and adds no host function.

Other behavior changes

• Under --preserve-symlinks, filePath is the path the resolver returns in that mode. src is then a.tsx instead of ../../real/pages/a.tsx.
• A file name with a backslash on POSIX made the constructor throw ENOENT opening route: <root>/pages/a/b.tsx, because the old join treated the backslash as a separator. It is now listed as <root>/pages/a\b.tsx.
• The router no longer needs /proc to load routes that are not symlinks.

Related open PRs

• Decide a directory watch event from the watchlist, not from the resolver's cached listing #41617 is the base of this PR. Without it, --hot stops reloading a route file saved by rename after router.reload(). The last new test covers that.
• bake: serve routes behind symlinks, reload every route that shares a file #42813 adds Resolver::real_path_of for the bake route scan. Its resolver hunk is included here unchanged except for one doc line. Whichever PR lands second drops it.
• resolver: a busted directory listing keeps its slot and is refilled in place #36675 builds Entry::is_named() on Entry::abs_path(), which this PR deletes together with set_abs_path() because the router was their only caller. If resolver: a busted directory listing keeps its slot and is refilled in place #36675 lands second it adds the accessor again.
• resolver: publish Entry.abs_path through the per-entry mutex #40258 publishes Entry.abs_path under the per-entry mutex. With this PR the router no longer reads or writes that field.
• FileSystemRouter: derive route names from the directory walk #41166 fixes reload() over a symlinked directory, which mangles route names today. The new tests do not call reload() on a symlinked directory, except under --preserve-symlinks, where the directory keeps its link path.

Not changed

bun test's scanner still uses Entry.abs_path as its "already seen" mark, and the six resolver sites still fill it. Entry::symlink() still reads the cache without the mutex while finalize_result stores cache.symlink under it. The router now reads through the same accessor as the resolver, so a fix there covers both. scopeguard stays in src/router/Cargo.toml although the crate no longer uses it, because dropping it also changes Cargo.lock.

Not run locally

Windows and macOS. cargo check passes for x86_64-pc-windows-msvc and aarch64-apple-darwin on the touched crates. One new test needs a file system that ignores letter case and is skipped on Linux.