fix(libSceNet): report EAGAIN when a stream send or receive times out on Windows - #4758

What

SO_RCVTIMEO and SO_SNDTIMEO on a libSceNet socket are set as the host socket's timeouts. On Windows a blocking receive or send on a stream socket that times out then fails with ETIMEDOUT (0x8041013c), where FreeBSD 11.4 returns EAGAIN:

• sbwait() sleeps for the buffer's sb_timeo and returns EWOULDBLOCK when it runs out;
• soreceive_generic() and sosend_generic() return that error;
• kern_sendit() returns the partial count instead when part of the data was sent.

Winsock reports its own timeout as WSAETIMEDOUT with no partial count, and documents that the connection is then in an indeterminate state (SOL_SOCKET options). #4092 fixed the datagram case, where the host timeout is safe. The stream case was left in TechnicalDebt for this reason.

On Windows, a blocking sceNetRecv, sceNetRecvfrom, sceNetRecvmsg, sceNetSend or sceNetSendto (and so sceNetSendmsg) on a stream socket with a timeout now waits for the socket with WSAPoll for that timeout first. It fails with EAGAIN when the wait ends, so the host timeout does not fire and the connection stays usable. A zero-length transfer does not wait, as in soreceive_generic().

A socket the host reports as unconnected (getpeername fails) does not wait either, because WSAPoll reports no event for it until the timeout. This covers a never-connected socket and a listening one, so their receive still fails at once with ENOTCONN.

The host option stays set for one case: a receive whose data another thread takes between the wait and the receive. That timeout now also maps to EAGAIN for stream sockets.

Linux and macOS are unchanged; their host timeouts already behave as FreeBSD's. The TechnicalDebt entry is rewritten. What remains is a send that starts while the socket is writable and still blocks past the timeout: it fails with ETIMEDOUT, where FreeBSD returns the partial count. No title is claimed.

The dependencies on #4669 and #4673 are only about merge order. They change the lines next to this change: the libSceNet TechnicalDebt entry above it, and sceNetSend's send call. This branch follows them once they merge.

Tested

Windows 11 (10.0.26300), WinLibs MinGW-w64 GCC 15.2.0 r7 (UCRT), Release, -DBUILD_TESTING=ON, on main at c952c2b:

• New CheckStreamTimeouts in guest_sce_net, on a connected loopback TCP pair:
  • Receives: with a 50 ms SO_RCVTIMEO, sceNetRecv (without and with MSG_PEEK), sceNetRecvfrom and sceNetRecvmsg fail with EAGAIN after at least 40 ms. Data sent afterwards still arrives.
  • Unconnected sockets: sceNetRecv on a never-connected socket and on the listening socket, both with the same timeout, fails with ENOTCONN.
  • Sends: with a 4096-byte SO_SNDBUF and a 50 ms SO_SNDTIMEO, 64 KiB sends to a peer that does not read succeed until one fails with EAGAIN after at least 40 ms.
  • Afterwards: the peer reads exactly the bytes the successful sends reported, so the failed send sent nothing. A further receive times out, and the connection still carries data.
• Before and after:
  • With main's Export.cpp, the test aborts at the first timed-out sceNetRecv, which fails with 0x8041013c (ETIMEDOUT) after 50 ms.
  • With only the receive-side wait, the receives return EAGAIN, but the send that fills the peer fails with ETIMEDOUT after 49 ms, 393216 bytes in.
  • Without the getpeername check, the receive on the unconnected socket waits and fails with EAGAIN.
  • With the change, guest_sce_net passes 20 runs in a row.
• 147 tests pass: guest_sce_net, guest_memory_pool_stats, guest_resolver, the guest_datagram_* and guest_socket_host_write_* tests, and the tests registered in core/libs/prx/libc/CMakeLists.txt and core/libs/prx/libkernel/CMakeLists.txt that build on Windows.
• tools/check_conventions.py --base origin/main passes.
• Linux: Ubuntu 24.04 (WSL2), GCC 13.3.0, Release with CI's configure flags (-DANYPS5_ENABLE_SPIRV_TOOLS=ON), on the same commit. guest_sce_net, including CheckStreamTimeouts, passes 20 runs in a row. The code change does not apply there; the test checks that Linux's own timeouts already give EAGAIN, the sent count and ENOTCONN.
• Not tested: macOS, any title.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: fix(libSceNet): dissolve a datagram association on connect to AF_UNSPEC #4669, fix(libSceNet): accept MSG_DONTWAIT in the send and receive calls #4673
• Unblocks: none
• AI-assisted: yes