[rchl]

Fixes the flaky TextDocumentTestCase.setUpClass timeouts on macOS CI (for example https://github.com/sublimelsp/LSP/actions/runs/36761225504 and https://github.com/sublimelsp/LSP/actions/runs/36917769576).

Cause

A trace from a failing macOS run showed this sequence:

1. The view of the previous test is still closing when the next test calls add_config.
2. The queued restart_sessions_async registers all known listeners again, including the listener of the closing view. That view is still valid, so a session starts for it. The new test view waits in the queue.
3. When the session is ready, the old view is closed. _dequeue_listener_async adds the session to _sessions, finds that it has no views and ends it at once. The test can find the session in that short time.
4. Then LSP starts a second server for the new test view. The test keeps the stopping session and times out. The second session is never ended, so the next test finds a leftover file watcher (1 != 0).

This is also a real issue outside the tests: when the view that started a session closes during initialize, LSP starts a second server for nothing.

Fix

When a new session has no views, do not end it at once. End it only if no view uses it after a short delay (Session.end_async_if_unused). This is the same idle shutdown that unregister_session_view_async already uses. The pending listener then uses the same session.

Testing

• macOS CI, 10 parallel runs before the fix: 2 failed.
• macOS CI, 10 parallel runs with the fix: all passed. The trace showed the race 7 times in 4 runs, and each time the new view used the same session.
• Full suite in Docker (Linux): passed.

The branch has temporary commits (macOS-only matrix, 10 attempts, trace prints) and a final cleanup commit that removes them. The net change is in plugin/core/sessions.py and plugin/core/windows.py only. Squash on merge.

🤖 Generated with Claude Code