ci: cross-check opc-gui against the Windows target

[kyaky]

v0.2.0-alpha.21 was tagged, built artifacts on every Linux platform, and then died — because opc-gui failed to compile on both macOS x86_64 and Windows. create-release never ran, nothing was published, and the fix turned out to be eleven _f32 suffixes (#31).

That is a bad trade: a one-line-per-site lint cost a release cycle, because CI never builds opc-gui.

Why it is invisible today

bins/opc-gui sits outside the workspace with its own Cargo.lock. So:

• cargo build --workspace --all-targets does not reach it
• cargo clippy --workspace does not lint it
• cargo fmt --all -- --check does not format-check it

The GUI is compiled for the first time during a release. Every GUI regression is therefore discovered at tag time, on a runner, after everything else has already succeeded.

Why not just build it on Linux

I tried, with a full GTK stack. It cannot work:

error: The platform you're compiling for is not supported by winit
  --> winit-0.30.13/src/platform_impl/mod.rs:78:1

bins/opc-gui/Cargo.toml takes eframe with default-features = false and only default_fonts, glow, persistence — no x11, no wayland. That is consistent with the release artifacts, which ship the GUI for macOS and Windows only. Adding those features is a real decision about whether a Linux GUI is supported, not something to do incidentally for CI.

What this does instead

Cross-check against the Windows target from the existing ubuntu runner — no linker, no Windows runner minutes:

cargo check --manifest-path bins/opc-gui/Cargo.toml --target x86_64-pc-windows-msvc --locked

~26s for a cold dependency graph. The lints that bite here are type-inference lints, so they fire during check.

Verified, not assumed

Both directions, locally:

• With the fix in place, the check is clean.
• Reintroducing the alpha.21 bug on a single line reproduces the release failure exactly:

$ sed -i 's/Stroke::new(1.0_f32, TEXT_SECONDARY)/Stroke::new(1.0, TEXT_SECONDARY)/' \
    bins/opc-gui/src/theme.rs
$ RUSTFLAGS="-D warnings" cargo check --manifest-path bins/opc-gui/Cargo.toml \
    --target x86_64-pc-windows-msvc --locked
error: falling back to `f32` as the trait bound `f32: From<f64>` is not satisfied
  --> src/theme.rs:30:60
error: could not compile `opc-gui` (bin "opc-gui") due to 1 previous error

So this job would have caught the exact failure that blocked the release, on a PR, in under a minute.

-D warnings comes from setup-rust-toolchain's default, which CI keeps deliberately — #31 removed it from release.yml only.

Limits, stated plainly

• macOS is not covered. Cross-checking it needs the Apple SDK. A macOS-only break is still possible; the mitigation is that the code that changes is shared, and the alpha.21 failure hit Windows too.
• This is check, not build. It will not catch a link error or a Windows-only runtime problem — the release job still does that.
• views.rs has 21 rustfmt diffs on main, for the same out-of-workspace reason. Adding --manifest-path to the fmt job would turn that red immediately, so it wants its own reformat commit first and is deliberately not bundled here.