ci: build and run Windows tests on Windows

[wan9chi]

Motivation

Windows was the only platform with its own test pipeline. A Linux job cross-compiled the tests with cargo-xwin, which needed a clang/MSVC STL version workaround and a cached MSVC CRT/Windows SDK. It packed them into a nextest archive and uploaded it as an artifact, and three Windows jobs then downloaded and ran it. This adds Windows to the test matrix instead, so it builds and runs with the same steps as Linux and macOS.

🤖 Generated with Claude Code

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[github-actions]

fspy benchmark

linux

dynamic/launch             change  +0.25%  [ -5.34% ..  +5.12%]  overhead  +299.10%
dynamic/access             change  +0.15%  [ -1.64% ..  +1.79%]  overhead   +13.64%
dynamic/access-relative    change  +0.22%  [ -1.10% ..  +2.13%]  overhead   +60.41%
dynamic/access-contended   change  +1.50%  [ -0.80% ..  +4.59%]  overhead   +16.67%
static/launch              change  -0.20%  [ -4.64% ..  +4.92%]  overhead  +765.12%
static/access              change  -0.08%  [ -1.47% ..  +1.05%]  overhead  +803.51%
static/access-relative     change  -0.02%  [ -2.07% ..  +1.14%]  overhead +1368.26%
static/access-contended    change  +0.09%  [ -0.46% ..  +0.75%]  overhead +3158.39%

macos

dynamic/launch             change  -0.27%  [ -5.55% ..  +5.04%]  overhead  +241.25%
dynamic/access             change  +0.35%  [ -9.27% .. +14.69%]  overhead    +4.84%
dynamic/access-relative    change  +1.11%  [ -6.04% .. +109.82%]  overhead  +246.99%
dynamic/access-contended   change  -0.79%  [ -9.90% ..  +8.10%]  overhead    +4.22%

windows

dynamic/launch             change  +0.58%  [ -5.63% ..  +7.36%]  overhead   +25.04%
dynamic/access             change  -0.53%  [ -9.09% ..  +1.92%]  overhead    +1.27%
dynamic/access-relative    change  +0.11%  [ -4.05% .. +26.28%]  overhead    +1.07%
dynamic/access-contended   change  -0.11%  [ -3.46% ..  +5.39%]  overhead    +2.49%

[wan9chi]

Closing: building and running the tests natively on Windows made PR CI slower.

Wall time, measured until the last job other than macOS x64 and musl finishes:

	Runs	Wall time
Linux build → Windows run (current)	8	145–202s, median ~163s
Native Windows, warm cache	2	177s, 194s
Native Windows, cold cache	2	271s, 199s

The single Windows job takes 153–160s even with a warm cache: 85–93s of build on the 4-core runner plus ~27s of Node setup. That makes it the slowest job in every run, while the rest finish by ~120s. The Linux cross-build keeps compilation off the Windows runner, so it stays.