[endel]

Summary

quic-zig builds, tests and runs natively on Windows again, and CI gets back its windows-latest job. Rebased onto main (Zig 0.17), which added the HTTP/1.1 and WebSocket stack after this PR was first opened, so that is ported too.

• Syscall seam. Windows lives in src/sys/windows.zig: Winsock for sockets, std.Io for files and clocks. The per-function OS switches in src/sys.zig became one supported-OS check.
• Helpers instead of repeated POSIX setup. sys.udpSocket, tcpSocket, setsockopt, setReuseAddr, setReusePort, setDualStack, send, sendVec, shutdown and argsIterator replace socket setup and raw std.c calls across src/ and the apps.
• Hosting the loop on Windows. xev_backend.PortWatch stands in for pollFd where a completion port has no fd to poll: a thread waits on the port, puts each packet back, and wakes the host.
• libxev. Pinned to quic-zig-2026-10-30, which carries mitchellh/libxev#240 (readiness through AFD), #259 (run(.no_wait) never asked the completion port) and #260 (a shut-down socket panicked in safe builds).

Also a fix on every platform: onReadable on Server and Client rearmed a cancelled watch, so on a backend that hands a cancelled watch its error they never finished stopping on a shared loop.

Windows behaviour worth knowing

In SPEC/STATUS.md#platforms:

• ECN is neither read nor set, and there is no send batching or UDP offload.
• SO_REUSEADDR isn't set, and a UDP socket no longer fails its next receive after an ICMP port-unreachable.
• No Unix sockets or SO_REUSEPORT: Http1Config.unix_path and reuse_port return an error.
• An HTTP/1.1 connection can't move between loops, as a socket stays on its first completion port.
• A nonblocking send can take all of a large write, so backpressure shows a write later.
• Timed waits round up to the ~15.6 ms system tick. sys.sleepNs under a millisecond yields in a loop instead.

Testing

• Windows (aarch64, under Wine 10): zig build test 1308 passed, 20 skipped; zig build fuzz 908 passed, 3 skipped. This PR's windows-latest job is the first run on real Windows.
• Linux (aarch64): zig build test 1328/1328, zig build fuzz 911/911.
• macOS (aarch64): zig build test 1323 passed, 5 skipped; zig build fuzz 909 passed, 2 skipped.
• routez tests/e2e/run.sh on this pin: macOS 475 passed, Linux 482 passed.
• netcode-io on this branch, both thread modes: 33/33 on Linux and macOS, and 29 passed, 4 skipped on Windows under Wine.

The skipped Windows tests need a feature above, or POSIX behaviour such as dup.

🤖 Generated with Claude Code