ci(standalone): run the crate's tests, and compile the platform-gated Rust

[dormouse-bot]

CI's only Rust step is cargo check on ubuntu-latest, which means two things nobody would guess from a green PR: the crate's 23 unit tests have never been run anywhere, and none of its #[cfg(windows)] / #[cfg(target_os = "macos")] code has ever been compiled by a check. This makes the Linux job run the tests, and adds a Windows/macOS job so the platform-gated halves compile on a PR instead of for the first time on a v* tag.

Surfaced by the nightly survey, which drew standalone/src-tauri/src/main.rs today and could not verify a one-line change to its Windows-only sibling.

What is currently invisible, precisely

The tests never run. cargo check type-checks #[cfg(test)] bodies only incidentally and executes nothing. 19 tests in lib.rs (sidecar resolution, dor CLI resolution, verbatim-path stripping, per-window session round-trips) and 4 in clipboard_win.rs (BITMAPFILEHEADER construction) are dead weight today — a broken assertion in any of them merges green. All 23 are pure path/filesystem tests with no platform assumptions beyond the two explicitly #[cfg(windows)]-gated ones, so they are safe to run on all three runners.

The platform-gated code never compiles. #[cfg]-excluded code is not type-checked at all — it is dropped before name resolution. On Linux that silently excludes:

• clipboard_win.rs (~330 lines) and pe_subsystem.rs, both #[cfg(windows)] mod
• ~25 #[cfg(windows)] / #[cfg(target_os = "macos")] / #[cfg(unix)] arms inside lib.rs
• the target.contains("windows") branch of build.rs, which is what stamps the bundled Node binary's PE subsystem

release.yml does build all three platforms — but only on: push: tags: v*. So the failure mode this closes is: a typo in Windows-only Rust merges green, and the next release tag fails its Windows leg mid-pipeline.

Cost, stated plainly. Windows and macOS runners bill at a multiple of Linux, so this is not free — it is two new runners per PR that touches anything. swatinem/rust-cache (the same pinned action release.yml already uses) keeps the steady-state to an incremental compile, the TypeScript check is not repeated on the new runners since it is platform-independent, and fail-fast: false means a Windows failure still lets macOS report. If the trade still isn't worth it, dropping macos-latest from the matrix keeps most of the value — clipboard_win.rs and pe_subsystem.rs are the two whole modules at stake, and both are Windows.

The new job is deliberately separate from standalone-smoketest rather than a matrix over it, so that job keeps its registered check name and branch protection is unaffected.

Verification. I could not run any of this locally: this checkout has no GTK/WebKit dev libraries and its Node is v22 against the crate's pinned build, so build.rs refuses before cargo gets anywhere. The YAML parses and the job graph is build-and-test, standalone-smoketest, standalone-platform-check; everything else is for CI on this PR to answer, and the two new legs are themselves the evidence. If they come back red for an environmental reason rather than a real compile error, that is worth knowing here rather than at a release.

Two notes for review: this edits .github/workflows/, so tonight's workflow-audit run will report the commit as an unexplained workflow change — that is the audit working, not a second problem. And it is the check that would have verified #427, a one-line fix to clipboard_win.rs that no job on that PR can compile.

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	ce8a2e2
Status:	✅  Deploy successful!
Preview URL:	https://37fcafbe.mouseterm.pages.dev
Branch Preview URL:	https://ci-platform-cargo-check.mouseterm.pages.dev

View logs

[dormouse-bot]

The macOS leg came back green in 2:13 from a cold cache and the Linux cargo test passed all 23, so the approach is answered — Windows is still compiling. Three efficiency notes, all against the premium-runner cost the PR body flags as the trade.

The pnpm install in standalone-platform-check (and the pnpm/action-setup that feeds it) do no work. The job runs no npm script — the TypeScript check is deliberately skipped — and the Rust build's only Node dependency is node on PATH, which build.rs resolves via resolve_node_binary → node -p process.execPath, plus the root package.json pin it reads off disk. Nothing under standalone/src-tauri/ references node_modules. On this run those two steps cost 16s on macOS and 48s on Windows.

Two follow-ups outside the diff, which I can push if you want them:

• The smoketest went from cargo check to a full cargo test — 1:46 uncached on this run — but only the new job got swatinem/rust-cache. The same step on the Linux job would recover most of that, though restoring and saving a Tauri target/ isn't free either, so it's worth measuring rather than assuming.
• ci.yml has no concurrency block, so stacked pushes to a PR now start two premium jobs each with nothing superseded. A cancel-in-progress scoped to pull_request is the cheapest cap on the cost this PR adds; the tradeoff is that cancelled runs leave superseded red check entries on the commit that never clear.

[dormouse-bot]

All three inline suggestions are on the branch as ce8a2e2 — pnpm install and pnpm/action-setup dropped from standalone-platform-check, timeout-minutes: 20 added — and the run on that head is green on every gating check, including both premium legs: Windows 1:39, macOS 0:40. That answers the only thing worth verifying about the removal, which is that neither leg needs node_modules; the wall-clock numbers themselves aren't a like-for-like against the 2:13 / 5:14 above, since that run was cold-cache and this one restored swatinem/rust-cache.

The two follow-ups outside the diff are deliberately not in that commit, since both are judgment calls rather than corrections: adding swatinem/rust-cache to the Linux smoketest trades a target/ restore+save against the 1:46 cargo test, which is worth measuring rather than assuming, and a cancel-in-progress concurrency block on ci.yml caps the cost this PR adds but leaves superseded red check entries on the commit that never clear. Both are one-liners whenever you want either.