Fix: close SQLite connections from terminated threads (#2192) - #2193

Summary

CoverageData keeps one SqliteDb per thread in self._dbs, keyed by threading.get_ident(). A connection is opened lazily in _open_db, but it was never closed when the thread that created it terminated. Entries were only reused if the same thread id recurred, and the whole dict was only closed at process end.

On workloads with many short-lived threads whose ids are not recycled, _dbs grows without bound and each dead thread leaks one open file descriptor to the data file, eventually hitting OSError: [Errno 24] Too many open files. This was observed on a multi-hour pytest run: ~6 live OS threads but 30k+ open fds to the .coverage file before failing.

Fixes #2192.

Changes

• Add CoverageData._reap_dead_thread_dbs(), which closes and drops any SqliteDb whose owning thread is no longer in threading.enumerate(). It runs under the existing self._lock, and closing is safe from another thread because connections are created with check_same_thread=False. Closing is best-effort so a close failure can't break collection.
• Call it from the cold path of _connect (only when the current thread has no connection yet), so the hot path on every add_lines/add_arcs is unchanged.
• Add a regression test in tests/test_data.py.
• Add a CHANGES.rst entry.

Test plan

• New test test_dead_thread_dbs_are_reaped uses a threading.Barrier to keep ~50 worker threads concurrent (so they get distinct thread ids, since a naive sequential start/join can let the OS recycle one id and hide the leak), joins them, asserts all 50 connections are retained, then records coverage from the main thread (the cold path) and asserts _dbs falls back to the live-thread count of 1.
• python -m tox -- tests/test_data.py passes (99 tests locally).

Notes

Thanks for maintaining coverage.py. Happy to adjust the approach — for example, reaping on a different trigger or under a separate config — if you'd prefer a different design.