fix twisted websocket close handshake

[alexrashed]

Motivation

The twisted serving layer never completes the WebSocket closing handshake, in either direction:

• When a client sends a close frame, WebSocketChannel.dataReceived only finishes the twisted request — no close frame is echoed back, and the TCP connection is left open.
• When the server closes (wsClose), the close frame is sent, but the TCP connection is again never terminated.
• On top of that, Request.finish() believes no response was ever started (the 101 upgrade is written raw to the transport, bypassing twisted's request), so it writes its default HTTP response into the upgraded websocket stream. The first byte — the H of HTTP/1.1 — decodes as a fragmented CLOSE control frame, so clients fail the connection with 1002 "Invalid attempt to fragment control frame".

Real-world impact: every well-behaved client burns its full close timeout waiting for a close reply that never comes (~10s with the websockets default). In LocalStack's integration test job this surfaced as a persistent failure on slow runners, where the timeout expiry additionally races the reader thread inside websockets' sync client and dies on assert self.protocol.state is CLOSED. Observed on the wire before this fix: client sends CLOSE 1000 → server answers with the mis-decoded 1002 garbage frame (or nothing at all) and the TCP connection stays open indefinitely.

Changes

WebSocketChannel now completes the closing handshake per RFC 6455 section 7:

• An incoming CloseConnection event is answered with event.response() (wsproto's close echo) before the channel shuts down.
• close() marks the request as written (startedWriting = 1) before calling Request.finish() when the connection was upgraded, so twisted emits no stray HTTP response — and then terminates the TCP connection via transport.loseConnection() (twisted flushes pending writes first, so the close frame still goes out).
• wsReject no longer attempts to write a response on an already-finished request.

After the fix, the wire shows a textbook close: CLOSE 1000 in → CLOSE 1000 echoed → TCP FIN.

Testing

• Two new regression tests in tests/websocket/test_websockets.py, run against both the twisted and asgi backends:

  • test_close_handshake_client_initiated — server must echo the close frame and terminate TCP;
  • test_close_handshake_server_initiated — client must receive a proper close frame followed by TCP termination.

  Both fail on the twisted backend without the fix; the asgi backend already conformed and passes unchanged.

• Full test suite passes (168 passed).

• Verified end-to-end against LocalStack with this rolo checkout overlaid: the previously failing test suite passes, and test execution time for affected tests dropped since clients no longer wait out their close timeout.

[bentsku]

LGTM, thanks a lot for the fix, I was also curious in the CI failures but didn't get to look at it yet, so thanks a lot!

This repo has the tag based release workflow so it should be easy to release 0.8.4 👍 let me know if you want me to do it