[rajivpant]

Every CLI command that touched the database (ragbot db status, ragbot index, ...) ended with this on Python 3.14:

Exception ignored while calling deallocator <function ConnectionPool.__del__ ...>:
  ...
PythonFinalizationError: cannot join thread at interpreter shutdown

The cached backend kept its psycopg_pool.ConnectionPool alive until interpreter finalization. The pool's __del__ then tried to join its worker threads, which Python 3.14 refuses at that point. Ragbot moved to 3.14 in #158.

Change

• PgvectorBackend._ensure_pool registers an atexit hook that closes the pool. Exit hooks run before finalization, while threads can still be joined. The hook holds a weak reference, so it never keeps a discarded backend's pool, and its connections, open.
• PgvectorBackend.close() closes the pool and is idempotent. VectorStore.close() is a no-op default for backends that hold no connections.
• reset_vector_store() closes the cached backend before dropping it, so tests that reset no longer leave pools open.

Verification

• ragbot db status against a live database: the traceback appears on main and not with this change.
• New unit tests cover the exit hook registration and its weak reference, the hook on a collected or already-closed pool, close() idempotence, and reset_vector_store() closing the backend.
• A new live test runs a real interpreter exit in a subprocess. The backend stays referenced until shutdown, as get_vector_store()'s cache keeps it in the CLI, and the test asserts stderr is clean. It fails on main with PythonFinalizationError and passes here.
• make test-fast is green. With a throwaway pgvector 18 database (app and migration roles split, as in compose), the full suite passes, including the live tests.

🤖 Generated with Claude Code