[gijzelaerr]

Refs #972. The reporter tested #979 and the CP 343-1 still keeps a connection resource after each disconnect().

What the new capture shows. #979 works: the DR now goes to the CP's reference (06 80 4431 0001 …). What's still wrong is the close sequence:

10.110431  client → CP   DR  03 00 00 0c 06 80 4431 0001 00 00
10.110479  client → CP   FIN            (48 µs after the DR)
10.111586  CP → client   RST
10.112294  CP → client   RST
10.608075  CP → client   ACK            (the CP still thinks the connection is alive)
10.608114  client → CP   RST

The CP never gets the chance to send a DC. The first capture (before #979) has the same sequence.

Changes, in both clients:

• disconnect() waits up to DISCONNECT_CONFIRM_TIMEOUT (1 s, and at most the connection timeout) for a COTP DC, or for the peer to close or reset the connection, before closing the socket. Other frames still in flight are skipped. The emulator server already answers a DR with a DC, so the test suite isn't slower.
• The DR loses the zero byte it carried after its header. The length indicator is 6 (type, both references, reason), so that byte was DR user data, which class 0 doesn't allow. TPKT length goes from 12 to 11.

Whether either change frees the CP's resource needs the reporter's hardware. Their next capture will show whether the CP now sends a DC.

Tests: socket-pair tests check that no FIN reaches the peer before it sends its DC, that disconnect() returns at once when the peer closes, that a silent peer is bounded by the timeout, and that a late data frame before the DC is skipped. An asyncio-server test does the same for the async client. Full suite: 1657 passed, 62 skipped (Windows, Python 3.13). Pre-commit and mypy are clean.

🤖 Generated with Claude Code