[tomassrnka]

What

On Windows, Close can block forever once the error consumer has gone away and the backend then reports an error.

Close puts a reply channel on done, wakes the reader, and waits on that channel. sendError takes the channel from done when the watcher is closing and drops it, where sendEvent puts it back. Its callers ignore the false return, so the reader handles the wakeup without seeing a shutdown request. It then skips the cleanup of its watches and of the completion port, and Close never returns.

This happens whenever nobody reads Errors any more and an error arrives during shutdown: an overflow (ErrEventOverflow), a failed CancelIo or CloseHandle in remWatch, or a GetQueuedCompletionStatus failure. Remove reaches the same sendError calls.

Fix

sendError puts the reply channel back on done, exactly as sendEvent does, and still returns false. The reader takes the channel on its next wakeup and runs its normal shutdown.

Tests (Windows)

• TestSendErrorKeepsCloseReply: with a reply queued and nobody reading Errors, sendError returns false and leaves the same channel on done.
• TestCloseAfterUnreadError: a real watcher. Close's reply is queued, an error meets an unread Errors channel, and the reader is woken. The reply must arrive within 10 s.

The first commit adds only the tests, and its CI run on windows-latest is the failing baseline. The second commit is the fix.

On Linux, the existing TestAdd/permission_denied fails when the tests run as root, before and after this change. It is unrelated.

CI

• windows-latest (Go 1.19 and 1.24): at the test-only head c6f96512c9 (run 39993141422) only the two new tests fail. At the fix head 73a141ee4f everything passes.
• Linux, macOS, OpenBSD, staticcheck: pass.
• dragonflybsd and illumos: fail on both heads, so they fail with the tests alone, before the fix. The change touches only backend_windows*.go.
• netbsd: passed at c6f96512c9 and was cancelled at 73a141ee4f.

After merge, please tag v0.0.2. belt's envd pins this commit's pseudo-version until then.

[cursor]

PR Summary

Medium Risk
Changes Windows watcher shutdown and error delivery on a one-slot coordination channel; wrong behavior could still hang Close or mishandle concurrent close.

Overview
On Windows, Close could block forever if shutdown was in progress and sendError could not write to Errors because nothing was reading it. sendError took Close's reply off done and discarded it unlike sendEvent, so the reader never saw the shutdown reply and never finished cleanup.

sendError now re-queues that reply on done when delivery fails. Close sets closed and returns early for duplicate calls under one lock so concurrent Close does not stack multiple replies on the single-slot done channel. Windows tests cover undelivered errors during shutdown and concurrent Close.

^{Reviewed by Cursor Bugbot for commit 17a6ca9. Bugbot is set up for automated code reviews on this repo. Configure here.}