[jayleaton]

What this changes

zig-windows fails on current main (d1018b31) for every pull request based on it, in two of main's files. This PR carries only those two files from #1034, unchanged.

It follows @akol1's proposed two-file split (#1034 (comment)), which also shows #1118 failing on the same lines. #1034's other 19 files remain there for the maintainers.

Error on Windows	Fix
core/digest_key.zig:34: inode is i64 on Windows, but the field is u128	the portable fallback reads its bits as u64
server/speech.zig:279: the socket handle is isize on Windows, so std.c.setsockopt rejects it	Windows sets TCP_NODELAY through sys.posix; the send stall uses sys.posix

No POSIX behaviour change:

• digest_key.zig:34 is reached only off Linux and macOS, which return from their statx / fstat branches above it.
• speech.zig:280 keeps POSIX's std.c.setsockopt call in its else branch.
• speech.zig:3, 283-284 use sys.posix, which is std.posix off Windows.

Receipt

• Environment: GitHub Actions on this PR (windows-latest zig-windows, ubuntu-latest zig), Zig 0.17.0, base d1018b31.
• Before: on d1018b31, zig-windows fails at digest_key.zig:34:42 and speech.zig:279:27 (for example test(deepseek): check CED replay against a linear window oracle #1118, job 118713181765).
• After (this PR, run 39487055416): the two errors are gone, but zig-windows still fails. It goes from 22 failed build steps to 18. With digest_key and speech compiling, the same steps reach further and expose digest_cache_probe.zig and two checkpoint_identity_test failures, which Windows: main's host tests build again on Windows hosts #1034 also fixes.
• Baseline on the same base (THIRD_PARTY_NOTICES: the tensor-parallel fabric's RoCE attribution beside the CUDA gather's (carried from #636) #1128, documentation only, run 39487050150): "165/211 steps succeeded (22 failed)", 80 error lines across main's files. These include the Windows errors Windows: main's host tests build again on Windows hosts #1034 fixes, and new ones from code landed since in fabric/tp/{bootstrap,sock,fate}.zig, native/*trainer*_test.zig, server/slide_*stream_test.zig and qwen35_cost_*.zig.
• So the two files alone do not clear the Windows job on d1018b31. Clearing it needs Windows: main's host tests build again on Windows hosts #1034's other files plus fixes for the newer ones. This stays a draft until the maintainers choose the route.
• Exactness, decode and prompt speed: n/a. No serving or arithmetic change on Linux or macOS.
• Not run: no local Zig build; no macOS host run (no workflow runs Zig on macOS).

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• No precision traded for speed. No bits change.
• No GPU kernel added or changed.
• New tests fail before the change, pass after it. (No new tests; this is a build fix for Windows.)
• 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]

Both fixes in this two-file split are included in 1.1.2 (https://github.com/ashhart/TensorFold/releases/tag/v1.1.2) through the Windows host work from #1034 and its digest-key follow-up, so we are closing this as superseded. Thank you.

The two files' changes landed in 61270ca and 476e16d.

- Synthema, Ash's AI Assistant.