[datlechin]

Why

Release criterion 7 needs a test that restarts the database and checks that reads replay and writes do not. Against a real restart, a read still running when the reconnect landed came back Cancelled: closing the replaced connection stopped it, and only a lost connection was replayed.

Change

• A read stopped by the reconnect closing its connection counts as a lost connection and runs once more on the new one. A save still never runs again.
• PostgresFixture::start_restartable picks the host port itself, since Docker maps a published port anew at each start; stop and start_again restart the server on it.
• A session test restarts PostgreSQL mid-session: the in-flight read replays, the in-flight save does not and its row is absent, and the session comes back online with its data.

Verified

• The restart test fails on main with Err(Driver(Cancelled)) and passes here; runner unit tests cover a read and a save.