[thomasjm]

This is the fix for #638, as described in the issue and in the repro.

[kazu-yamamoto]

Thank you for the careful analysis and the repro. I checked it locally on Windows 11 with GHC 9.12.4.20260713 (__IO_MANAGER_WINIO__ = 2), building crash against network master (9ce1211) and against master with this PR cherry-picked:

• sync_completion.c: without the flag, 100/100 WSARecv and WSASend that finish at once each queue a packet; with it, none do.
• crash: master crashed within a second, three runs out of three (Access violation in generated code when reading 0x101010101010109, exit code 11); with this PR it survived 120s.
• The test suite passes with this PR under WinIO and MIO, and with GHC 9.8.2.

I also confirmed in GHC's withOverlappedEx that a CbPending start calls getOverlappedResult without waiting and completes the request inline when it has finished, so the packet left on the port is the only remaining reference.

Note that CI does not exercise this: its Windows jobs use GHC 9.4 and 9.14.1, which compile network's WinIO code out.

A few points:

1. When SetFileCompletionNotificationModes fails, associateSocket raises a bare userError. Could it raise the GetLastError code instead, like other Winsock errors (throwSocketErrorCode)? The socket is closed by bracketOnError either way.
2. The non-IFS LSP check is not needed, I think: LSPs have been deprecated since Windows 8.
3. This now conflicts with Give Winsock errors the IOErrorType of their errno #637 in Network/Socket/Cbits.hsc (both add definitions after wsaENOPROTOOPT); keeping both sides is all it takes. Could you rebase onto master?

@Mistuke In #602 you mentioned that FILE_SKIP_COMPLETION_PORT_ON_SUCCESS is not valid for sockets. Microsoft documents it for sockets except behind a non-IFS LSP, and libuv sets it on its TCP sockets in the same way (uv__tcp_set_socket). Are you OK with this change?

[kazu-yamamoto]

Merged into master as 0d8b844 after rebasing onto it locally (the conflict with #637 was only adjacent definitions in Cbits.hsc; your commit is kept as 5421eed). Thank you!