[BlueX888]

Fixes #283

The else branch of _RPClient.__init__ creates its own Client but sets own_client = False, so close() skips self.__client.close() and the aiohttp session stays open until garbage collection warns about it. Set the flag to True in that branch; a client passed via client= remains borrowed and is still not closed.

Client.__setstate__ also restores _session = None, because __getstate__ removes it and unpickling left the attribute missing. Without this, the newly-enabled close() path on an unpickled wrapper raised AttributeError, and an unpickled Client also failed on its first session() call.

Regression tests:

• tests/aio/test_batched_client.py::test_own_client_reflects_client_ownership (parametrized over BatchedRPClient and ThreadedRPClient) asserts own_client is True for an internally-created client and False for an injected one. Both params fail on develop with assert False is True.
• test_batched_rp_client_pickling now checks the unpickled wrapper owns its client and that close() succeeds.

Full suite: 418 passed, 2 pre-existing network/port failures (tests/_internal/aio/test_aio_http.py::test_no_retry_on_ok_request — OSError: [Errno 48] Address already in use; tests/_internal/test_http.py::test_no_auth_request — requests.exceptions.ConnectionError). Both also fail on the unmodified base.

[HardNorth]

LGTM

[BlueX888]

The only failed job is build (3.9): actions/setup-python reports The version '3.9' with architecture 'x64' was not found for Ubuntu 26.04, and the other six jobs were cancelled by fail-fast. The change is plain pytest and does not depend on the interpreter version, so this looks like the runner image no longer carrying 3.9 rather than the patch. Thanks for the LGTM.