• ResultSet.close() (result_set.py:184-187) wraps the same call in try/except RequestError, swallowing CursorAlreadyClosedError. The new branch has neither the guard nor the except — only a finally that clears active_command_id but does not contain the raise.
• Connection._close() (client.py:547-550) runs for cursor in self._cursors: cursor.close() with the try/except covering only the subsequent session.close(). A raise from any cursor aborts the loop → remaining cursors never closed, session.close() never runs.
• Blast radius: conn.close() / with teardown when an async cursor's handle is already gone server-side (post-cancel, expired session) or a transient network blip → whole-session leak + raw stack trace out of close(). This is the exact scenario the PR targets, so it's a live regression, not hypothetical.
• Fix: mirror ResultSet.close() — swallow RequestError/CursorAlreadyClosedError (+ a catch-all) and log. This is what issue Cursor.close() leaks server-side handle for async commands that were never fetched #791's suggested fix specified and the PR dropped.

ResultSet.close() gates the RPC on status != CLOSED and not has_been_closed_server_side and connection.open (result_set.py:180-182). The new branch fires unconditionally, sending a pointless (and, per F1, unguarded) close_command RPC even when the session is already closed server-side. Gating on self.connection.open would avoid the round-trip and much of F1's trigger surface.

Both added tests use bare Mock() backends whose close_command never raises, so they only prove the happy case. The one genuinely new invariant — active_command_id still cleared (via finally) and close() staying best-effort when close_command raises — is entirely unverified. Fix: add a test with mock_backend.close_command.side_effect = RequestError(...), asserting close() does not raise and active_command_id is None.