[kolkov]

Moves ui's CI to the pattern gogpu now runs on main (ci.yml at ef8260b).

The main gap it closes: tests did not gate the merge on macOS and Windows. Only the ubuntu job ran the plain tests, and the race step on every OS ran kolkov/racedetector with continue-on-error: true, so neither a race nor a macOS/Windows test failure turned CI red.

Changes in .github/workflows/ci.yml:

• go test on all three OS, plus go test -race with CGO_ENABLED=1 on that step only. Everything else stays at CGO_ENABLED=0, the build users ship.
• golangci-lint v2.15.0 comes from the shared gogpu/.github setup action (cached, sha256-checked) instead of downloading latest on every run. Earlier today that download returned 504 for over an hour.
• New Lint - windows/amd64 and Lint - darwin/arm64 legs. ui has no per-OS files of its own, but desktop/ and the examples compile against gogpu's per-OS API, which the Linux lint never sees.
• Every action is pinned by commit SHA with a version comment, and .github/dependabot.yml keeps them current. The root module is left out of Dependabot, since ecosystem deps move only to released versions in cascade order.
• go mod tidy must leave go.mod/go.sum unchanged. The separate examples build step is gone, because the examples are part of ./... since 0.1.55.
• A Required checks job depends on every other job and fails unless all of them succeeded. The branch ruleset will require only this job; that switch is a separate settings change.

A second commit corrects the 0.1.55 CHANGELOG note about the readback tests. They ran on the 0.1.54 dependencies; wgpu v0.34 would have made them skip, and the row padding prevents that. The release notes already say this.

The third commit fixes a data race in a test double that the old setup never reported. go test -race on macOS caught it on the first run: the window's animation pumper calls RequestRedraw from its own goroutine, and the mock provider counted those calls in a plain int while the test read it. The hosts allow that call from any goroutine (gogpu documents it), so production code is unchanged and the counter is now atomic. The Test matrix also gets fail-fast: false: on that run the macOS failure cancelled the Linux and Windows jobs, which hid their results.

Proving the new checks can fail. Two runs had a temporary probe commit, since dropped (runs 42469289570 and 42469586731). It added an unused function that only windows compiles, another that only darwin compiles, and a test with a data race. On the second run:

• Lint - windows/amd64 and Lint - darwin/arm64 each flagged their probe, and Linux Lint stayed green;
• the plain tests passed on all three OS, then -race failed on all three, on the probe test only;
• Required checks failed.

Local, with GOWORK=off: golangci-lint for linux/amd64, windows/amd64 and darwin/arm64 reports 0 issues without the probe commit. With it, windows and darwin each report their probe and linux reports 0. go mod tidy leaves no diff.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!