fix(lsp): rescan when a watched directory is replaced while its watch attaches

[mikn]

On macOS the tsserver hook worker can miss a directory replaced while its watch attaches, so an editor keeps a stale resolution map and //tests/lsp:test_worker_map hangs until its 60 s timeout, failing macOS CI.

Problem

The worker learns that a watched ancestor was replaced only from the parent watcher's event. On macOS a directory fs.watch is an FSEvents stream that starts asynchronously, and a change made while it starts sends no event. Under load, that window covers the test's replacement, so the worker never rebuilds. Linux inotify is live when fs.watch returns, which is why Linux passes.

Observed in production

Not observed in production; the worker is editor tooling. It failed rules_typescript#267's macOS unit job twice, on a runner busy compiling oj from scratch. An editor on a loaded Mac can hit the same gap.

Gain

The worker no longer depends on an event FSEvents may drop, so replacements reach the map on every platform, and #267's macOS job stops timing out.

What changes

• After attaching a watch, the worker re-stats the directory and schedules a rebuild when its dev/inode changed during the attach.
• rebuildTimer is declared before the initial build, which can now schedule a rebuild during its own watch attachment. Previously that call hit the uninitialized let and was swallowed by watchDirectory's error handler.
• worker_map_test.mjs drops the parent's event for the replaced directory, as FSEvents can, so the test fails without the re-check on Linux too.

🤖 Generated with Claude Code