Stage 4: cross-platform windows_threading::submit (Win32 fast path + std::thread::spawn fallback)

[Copilot]

Continues docs/cross-todo.md Stage 4. The threading_bench example added previously was run on Windows:

Workload	win32-pool	std::spawn	Ratio
single	15.67 µs · 63 813/s	67.20 µs · 14 880/s	~4.3×
burst	0.46 µs · 2 180 027/s	41.16 µs · 24 296/s	~90×
steady	0.46 µs · 2 162 686/s	42.15 µs · 23 724/s	~91×

Per-submit cost on the Win32 pool drops from 15.67 µs (cold) to 0.46 µs (warm) — clear worker-thread reuse — while std::thread::spawn stays flat at ~42 µs paying a full CreateThread per call. Well past the "≥ 5–10× in steady" threshold from the original decision rule, so: keep the windows-threading dependency in windows-future and make submit itself cross-platform.

Changes

• crates/libs/threading/Cargo.toml — new default-on std feature.

• crates/libs/threading/src/lib.rs — drop the crate-wide #![cfg(windows)]. submit gains three cfg-gated bodies:

  • Windows (any feature config): unchanged Win32 fast path through TrySubmitThreadpoolCallback.
  • Non-Windows + std (default): std::thread::spawn(f), fire-and-forget.
  • Non-Windows + no_std: unimplemented!() — links but signals at runtime that no portable thread primitive is configured.

  The Win32 surface (Pool / for_each / thread_id / sleep, the bindings module, check / try_submit helpers) stays Windows-only via #[cfg(windows)]. The crate-level expect(non_snake_case, ...) lint group is also gated to Windows since those idents only exist there.

• .github/workflows/linux.yml — windows-threading and windows-future added to the build and doc loops; windows-threading also added to the --no-default-features build loop so the no_std-without-std path stays green.

• docs/cross-todo.md — Stage 4 rewritten with the measured numbers, the decision, and the implementation. Marked ✅ done.

Runtime async tests under crates/tests/libs/future/ are intentionally untouched — they go through RoInitialize/COM and stay Windows-only; only the build/doc paths of windows-future are exercised on Linux. This unblocks Stage 5 (windows-collections on Linux).