HandleWatcher replacement error handling logic is broken

[glasser]

7e10102 changed HandleWatcher.start to "Remove the old one if we found two same file handles" instead of throwing an exception.

Choosing to replace an existing HandleWatcher with a new one with just a logged error instead of throwing an exception might make sense. But the choice in that commit to call troubleWatcher.close() seems broken.

troubleWatcher.close() does two things: binding.unwatch, and handleWatchers.remove. The latter is reasonable: we're trying to replace the old bogus HandleWatcher with the new one, so we should remove it, sure.

But binding.unwatch here is problematic: the argument we're passing to it is just the fd that we got back from binding.watch, which we're going to pass to close (Mac) or inotify_rm_watch (Linux). But... that means we're actually closing the fd we just got back from binding.watch and which is saved in the new HandleWatcher!

ie, this error case means we've suddenly learned that a HandleWatcher in handleWatchers contains an invalid WatcherHandle (fd). Going and calling close on that invalid handle is just going to break the new thing we got.