This is more of a question/discussion - I'm trying to understand if there's a reason for the rather complex "session" fixture.
If you have an actual PG database already setup, all tables were recreated by setup_test_db (that said, this is happening per session and not per test), why do you need a special flow to create the session and override the fast api dependency?

Some observations I made -

1. The engine is returned before the transaction is completed in setup_test_db - is that how you are making sure the db remains clean?
2. That said, you begin a new transaction in session fixture. At the end, you rollback the connection (but assuming some "commits" might have happened).
3. end_savepoint - what is the point of it? if I understand correctly, it will automatically re-start a nested transaction when one is over in a sync session. But why?

Appreciate your clarifications!