libkernel: connect() on a listening TCP socket returns the host error instead of EOPNOTSUPP; fix branch ready #6923

connect() on a listening TCP socket returns the host error (EINVAL on Windows, EISCONN on Linux) instead of EOPNOTSUPP (45). FreeBSD 11.4 soconnectat fails any socket that has SO_ACCEPTCONN set with EOPNOTSUPP. A program that connects its own listener (a probe, or a retry loop that reuses the descriptor) tests for 45 on the console and sees a different number here.

Fix branch: https://github.com/husseinkad/AnyPS5/tree/fix/socket-connect-listener-eopnotsupp (two commits; it merges cleanly into current main; libkernel/Socket/src/Udp.cpp, the new test tests/GuestSocketConnectListening.cpp registered in the libkernel CMakeLists.txt). For stream sockets, connect now queries SO_ACCEPTCONN after the address has been read and validated (FreeBSD's kern_connect copies the address in before soconnect, so a null or short address still fails first) and returns EOPNOTSUPP for a listener, before the family check.

Checks:

• The new test fails on main (abort, errno 22) and passes with the change; the 18-test socket group passes.
• Probed against main on loopback: a valid address on a listener gives 45 for IPv4, IPv6 on an IPv6 listener, an IPv6 address on an IPv4 listener, a nonblocking listener, and after shutdown; connect(listener, NULL, 16) gives 14 and a one-byte address gives the same error as a plain socket, as in FreeBSD's order. Non-listening TCP is unchanged (NULL gives 14, a refused connect gives 61, a second connect gives 56, and a connect to the listener from another socket works), and UDP is unchanged. accept keeps working after every failed connect on the listener. In a 3,000-sequence fuzz, 351 connect results changed from main (156 from 22, 147 from 47, 48 from 14), all on listeners.
• Three independent reviews. The first found the check ran before address validation (null address gave 45), fixed in the second commit. check_conventions.py is clean and pr_overlap.py found no overlap; it conflicts only textually (adjacent insertions) with my listen auto-bind fix and the port-zero fix, and keeping both sides resolves it.

Not covered: addresses that Address() rejects with 47 (an AF_UNIX family byte, family 0 or 255, a 4-byte-long IPv4 address) still return 47 on a listener where FreeBSD would return 45; a length of 1 returns 14 for plain and listening sockets where FreeBSD gives EINVAL (22), which the sockaddr length checks in #4920 would change. The Linux error code was read from the kernel source, not run. Only Windows was built and run.

AI-assisted: yes