Windows: main's host tests build again on Windows hosts

[jayleaton]

What this changes

On current main (76aafcf2), zig-windows fails 15 build steps on every pull request based on it. The errors are in main's own files.
This PR makes main's host tests build and pass on Windows again. It uses the sys layer (zig/src/platform/sys.zig) and the Windows gates main already has.
It unblocks the Windows job for every PR based on current main, and changes no POSIX behaviour; the touched lines are listed below.

zig-windows on main 76aafcf2: 15 failed steps, 51 error lines
  std.DynLib, std.c HANDLE types ... sys.DynLib, sys.c (the same calls on POSIX)
  Environ.getPosix ................. sys.getenv
  POSIX-only tests and hints ....... existing Windows gates / SkipZigTest
zig-windows on this branch: success

Error groups and the files fixed

Error on Windows	Fix	Files
std.DynLib: "unsupported platform", or not the sys.WinDynLib a Driver holds	sys.DynLib (which is std.DynLib off Windows); the native load fakes open kernel32.dll on Windows, as the copy-custody tests do	core/picture/webp.zig, cuda/roce.zig, families/drama3/cuda/blas.zig, cuda/copy_source_close_test.zig, cuda/copy_source_race_test.zig, native/cuda_load_gpu.zig
std.c types that are HANDLEs on Windows: clock_gettime's void parameter, pid_t (which JSON cannot carry), write	sys.c.clock_gettime, nanosleep, pid_t, getpid, getenv, write; the digest probe writes through std.Io on Windows, because it is built without sys	families/flashnext/cuda_rank_wire.zig, cuda/memory_receipt.zig, core/digest_cache_probe.zig
Environ.getPosix: no view on Windows' environment block	sys.getenv, as the JPEG case in the same file already does	core/picture/picture_test.zig
inode is i64 on Windows, but the field is u128	its bits read as u64	core/digest_key.zig (portable fallback only)
Socket handle is isize on Windows (std.c.setsockopt, getsockopt)	Windows sets TCP_NODELAY and the send stall through sys.posix; the test reads the option back only off Windows	server/speech.zig, server/listener.zig
ws2_32.MSG has no NOSIGNAL / DONTWAIT	send, receive and checkPeer refuse off Linux, as Link.open already does	families/flashnext/cuda_rank_wire.zig
Benchmark command (termios, SIG.HUP, Permissions.toMode, posix_openpt)	benchmark_config_test.zig joins the benchmark tests' existing non-Windows gate	server/root.zig
std.c.socketpair is void on Windows	Windows skips tests built on Unix socket pairs	server/root.zig (speech stream test), families/flashnext/cuda_rank_media.zig, families/flashnext/cuda_lanes_restore_test.zig
fork / waitpid death tests	SkipZigTest on Windows	vision/qwen_cuda_test.zig, families/nemotron/cuda_pages_safety_test.zig
mmap / munmap undefined at link	H3's host targets and its checkpoint test skip Windows, like the fake fabric's at build.zig:595	zig/build/cuda.zig
Digest reuse across runs, which digest_key.persistent keeps only on Linux and macOS	skip on !persistent, as every digest_cache_test does	core/checkpoint_identity_test.zig
std.posix.MADV is void on Windows	the madvise read hint skips Windows	families/flashnext/cuda_ple.zig

No POSIX behaviour change: the touched lines

Each line below falls into one of five groups. Lines in the first group take the same path on Linux and macOS, where sys.DynLib == std.DynLib, sys.c == std.c and sys.posix == std.posix. The other groups add a branch taken only on Windows, or are unreachable on POSIX.

• Same call on POSIX: webp.zig:69, roce.zig:4, 216, 288, drama3/cuda/blas.zig:3, 37, 53, copy_source_close_test.zig:87, copy_source_race_test.zig:100, cuda_load_gpu.zig:159, 281 (libc.so.6 / libSystem are unchanged), memory_receipt.zig:3, 10-11, 21-22, 26, 31, 36, cuda_rank_wire.zig:4, 47-48, 57-58, picture_test.zig:122, speech.zig:3, 282-283.
• Windows-only branches: speech.zig:280 (POSIX keeps its std.c.setsockopt), listener.zig:158-163, digest_cache_probe.zig:33-34 (POSIX keeps its std.c.write), cuda_ple.zig:391, server/root.zig:42, cuda_rank_media.zig:61, cuda_lanes_restore_test.zig:52, qwen_cuda_test.zig:248, nemotron/cuda_pages_safety_test.zig:162, zig/build/cuda.zig:228-229, 362-363.
• Off-Linux refusals: cuda_rank_wire.zig:135, 153, 185. On macOS these are unreachable, because Link.open already returns error.RankPlatform there.
• !persistent skips: checkpoint_identity_test.zig:127, 172. persistent is true on Linux and macOS.
• Moved inside an existing gate: server/root.zig:113. benchmark_config_test.zig still runs on Linux and macOS.
• Portable fallback only: digest_key.zig:34. Linux and macOS return from their own statx / fstat branches above it.

Receipt

• Environment: GitHub Actions windows-latest and ubuntu-latest, Zig 0.17.0, base 76aafcf2.
• Before: zig-windows on PRs based on 76aafcf2: "Build Summary: 124/156 steps succeeded (15 failed); 1243/1256 tests passed (13 skipped)", with the same 51 error lines in each run (for example run 39129528327).
• After: run 39133116461 on 023975f1, with zig-windows and zig both successful. The Windows job's zig build test, zig build test-golden ("303/315 equal to the frozen Python answers, 0 unexpected differences") and zig build fuzz-cortheon steps all pass. fmt, desktop and mobile pass too.
• Exactness, decode and prompt speed: n/a. No engine arithmetic or serving path changes on Linux or macOS.
• Not run: no local Zig build or tests; CI is the only test environment for this change. Lean rule: an emulation of tools/zig/lean_check.py's rules reports no new finding against 76aafcf2.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (No serving change.)
• No precision traded for speed. No bits change.
• No GPU kernel added or changed.
• Prompt processing is no slower than the last release. (No prompt path change.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not. (Windows and Linux host CI only; no device runs.)
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new tests; existing tests skip on Windows where they need POSIX.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Thanks @jayleaton. The native fakes now use kernel32.dll, and H3's host roots are excluded on Windows, so that part of the earlier ask is addressed. Your Windows and Linux CI receipts are useful.

The rank-wire issue remains: the restore and media tests construct links directly from socket pairs on macOS, bypassing Link.open, so send, receive and checkPeer still hit the new off-Linux refusal. Please keep those Mac controls working by limiting the new refusal to Windows or retaining their POSIX transport, and attach Mac and Linux host results. We'll carry the Windows fixes onto our line once those gates pass.

Posted by Ash's AI Assistant Synthena

[ashhart]

Thanks @jayleaton. The Windows-only rank-wire guards now preserve the macOS socket-pair restore and media controls, closing our remaining source finding. The native fake DLL fixes and Windows-only H3 exclusions remain intact. We'll carry this onto the 1.1.2 line, preserving the newer build registrations, optional-library errors and server recovery behavior. The next gate is the restored Mac controls plus Linux and Windows host checks on that integrated source. No new model or GPU run is needed for this portability change.

Posted by Ash's AI Assistant Synthena