[nhobin219]

Fixes #106 and #154.

A publisher's contract is to reconnect on ConnectionClosed. A send on a connection that is closing or closed raised something else in two ways, and a publisher that catches only ConnectionClosed died on it.

#106: AttributeError from websockets' abort

On a connection that is no longer open, websockets' send_context calls transport.abort() before raising ConnectionClosed. After a close with data still buffered, asyncio's selector transport has already run connection_lost without counting it, so its _loop is None. abort() then raises AttributeError: 'NoneType' object has no attribute 'call_soon', and the send raised that. websockets 17.1 still has the unguarded abort.

Fix (_publish.py, submit_many, which all four send methods go through): handles an Exception from connection.send, other than ConnectionClosed, when the protocol state is no longer OPEN.

• It raises the connection's own close_exc, chained from the original, or the refusal its close code carries (_refusal, as the ConnectionClosed branch already did).
• On an open connection the exception propagates unchanged.
• BaseException (cancellation) is untouched.

#154: the reader drains the pending replies first

The IndexError('pop from an empty deque') itself is already fixed on main. _unowed removes by identity; it shipped in 0.23.0 with #167, untested. Two things were still wrong:

• The dropped reply's failure was never retrieved. The reader had set an exception on the send's future, and the future is never handed out. asyncio logged "Future exception was never retrieved" for each occurrence. _unowed now retrieves it.
• publish: a send racing the connection's close raises IndexError('pop from an empty deque') instead of ConnectionClosed #154's point about a lost refusal (e.g. 4410 StreamRetired) is covered on main. The send's own ConnectionClosed carries the same received close frame, and _refusal maps it.

Tests: tests/test_publish_closed.py

These run on a real loopback server and publisher, with only the faulty piece injected.

• TestTheAbortItself::test_a_failing_abort_raises_connection_closed[send|submit|send_many|submit_many] closes the connection and makes transport.abort raise, as asyncio's does. It asserts ConnectionClosed, caused by the AttributeError.
• TestTheReaderGotThereFirst::test_the_send_raises_connection_closed forces publish: a send racing the connection's close raises IndexError('pop from an empty deque') instead of ConnectionClosed #154's interleaving from a hook in connection.send: close, wait for the reader to drain, then send. It asserts ConnectionClosed, and that asyncio reports nothing never retrieved.

On unfixed main (26e903d) all 5 fail:

E       AttributeError: 'NoneType' object has no attribute 'call_soon'      (x4)
E       AssertionError: the dropped reply's failure is retrieved
E         Left contains one more item: {'message': 'Future exception was never retrieved', 'exception': ConnectionClosedOK(...)}

Falsified. Each break below made tests fail; each was restored and passes:

what I broke	result
no retrieval in _unowed	the race test fails
_unowed back to self._pending.pop(), the 0.22 code	the race test fails with IndexError: pop from an empty deque, #154's exact error
the closed-state branch removed	all 4 abort tests fail with AttributeError

The new tests take about 2 s in all.

Full suite, locally: STREAMCAST_REQUIRE_S3=1 STREAMCAST_REQUIRE_INTEROP=1 STREAMCAST_REQUIRE_POSTGRES=1 pytest gave 959 passed, 2 failed, 2 skipped. Neither failure is in the publish path.

• test_maintain.py::TestFlushing::test_a_served_row_reaches_the_published_table fails the same way on unmodified main on this box.
• test_snapshot_readers.py::TestAReader::test_the_event_loop_runs_while_it_reads is a tick-rate timing check (212 ticks in 4.78s). Run alone here it passed 2 of 3 times. The box has 2 vCPUs.
• Node is not installed here, so test_node.py is left to CI.

No critical review was run. This is routine error mapping on the client side. It touches no durable state, no frames and no ordering, and the set of futures owed a reply is unchanged: a send that raised was already removed.

Downstream

Not breaking, so it should ship as a patch. No consumer pins move. A publisher that also caught AttributeError or IndexError as a workaround can drop that.

🤖 Generated with Claude Code