CI runs the unit suite on ubuntu only; macOS and Windows run a single test #373

All tests must run on every supported platform. Today the unit suite runs on ubuntu-latest only, and macOS and Windows run exactly one test.

Current state

make test-scenario is go test -run TestWritDeployScenario -v -count=1 -timeout 600s ./cmd/writ (Makefile). It is the writ deploy end-to-end scenario, deliberately scoped, and it is the only thing macOS and Windows execute.

So a green scenario (windows-latest) says one writ deploy scenario passed on Windows. It says nothing about the other 120 packages.

Consequence: ten platform-gated files are unit-tested nowhere

Build constraints exclude these from the ubuntu run, and the platforms where they compile only run the scenario test:

• pkg/op/default_funcs_windows.go
• pkg/op/provider/file/helpers_windows.go
• pkg/platform/detect_windows.go
• pkg/platform/windows_managers_windows.go
• pkg/platform/detect_darwin.go
• pkg/platform/darwin_managers_darwin.go
• cmd/writ/writ/snapshot/protect_darwin.go
• pkg/platform/detect_linux.go
• pkg/platform/linux_managers_linux.go
• pkg/platform/cross_distro_managers_linux.go

The _windows.go and _darwin.go files cannot even be compiled by quality-gate — go build on ubuntu excludes them. Their only compilation in CI is make dist, which cross-compiles via GOOS=$$os GOARCH=$$arch (Makefile:195-196) and runs on release tags only. Compiling is not testing, and it does not happen per pull request.

This is the test-side form of a blind spot this repository has already been bitten by: the lint recount had to add an explicit GOOS=linux pass because CI is ubuntu and platform-gated code was escaping the gate.

How it surfaced

PR #372 added characterization tests under cmd/lore/lore and cmd/lore/lore/onboard. All five checks went green, including scenario (windows-latest), and that was reported as the tests passing on Windows. They did not run on Windows — -run TestWritDeployScenario ./cmd/writ never reaches those packages. They ran on ubuntu only, via test-race in quality-gate.

Those tests assert exact byte sequences including newlines, which is precisely the class of assertion that can differ across platforms.

Required change

make test-race runs on every leg of the matrix, not just ubuntu.

Known risks, to be handled by the plan rather than assumed away

1. Turning the suite on for Windows and macOS will likely surface pre-existing failures. Those are real defects being revealed, not regressions introduced by this change, and each needs triage rather than a skip.
2. Some existing tests may carry Unix assumptions — path separators, file permission bits (0o600 behaves differently on Windows), symlink creation privileges, line endings.
3. CI cost rises: the full race-enabled suite on three runners instead of one.
4. quality-gate currently mixes concerns (build, tidy, frontmatter, vet, lint, test, shell lint, metrics). Only the test step needs to fan out across platforms; lint and shell lint arguably should not.