File watching on repo-root config files can hit EMFILE in large repos

[vkz]

0. Macos
1. large repo with a big cache tree like .clj-kondo (Clojure linter)
2. root-level watch entries like in pitchfork.toml daemons declarations: watch = ["shadow-cljs.edn", "deps.edn", "package.json"] intended to restart that daemon.
3. Pitchfork 2.2.0
4. supervisor hits Too many open files while setting up file watching, before a later daemon starts

Observed:

• starting the supervisor immediately drives its open fds very high
• in my case it then fails with:
  Too many open files (os error 24)
• Pitchfork log shows it failing while trying to watch the repo root, e.g. on a file under .clj-kondo/.cache/...

In my repo, removing those watch entries makes startup succeed immediately.

I'm not entirely sure how to report this any better tbh. I think I've heard of that file-watching limitation before where you can't just watch a single file so you gotta watch its dir and filter or something. Is pitchfork impl maybe exhibits just such problem? I think I encountered something similar with inotify in Erlang, but now I don't recall.

Am I holding it wrong?

[gaojunran]

this is a problem indeed and I will fix it.

[gaojunran]

fixed in #301.