ci: verify behaviour on every platform and halve the macOS queue

[PathGao]

Summary

test.yml runs npm run check, npm test and cargo test on a single ubuntu-22.04 runner. test_build.yml has a platform matrix but only builds.

The matrix therefore proves the tree compiles everywhere and nothing more. A Windows- or macOS-only behavioural regression in the cfg-gated paths — which is exactly where this project's platform-parity bugs keep landing — reaches a release build unchallenged.

This runs the suites inside the matrix, and pays for the added time twice over.

Why the matrix shrinks

The matrix built macOS twice, once per architecture. Measured across the currently open pull requests:

job	runs in	waits for a runner
linux	9m10s	starts immediately
windows	10m19s	starts immediately
macos-aarch64	6m06s	19–30 min
macos-x64	8m10s	19–30 min, several still queued

macOS jobs are the fastest here and the slowest to start: capacity, not duration, is the constraint, and every pull request was taking two of those slots.

One entry targeting universal-apple-darwin takes one slot and covers strictly more. It compiles both architectures, so an Intel-only break still fails the check — and it fails on the same artifact the release produces, rather than on a pair of single-arch binaries nothing ships.

Why the concurrency group

Each push to a pull request queued another full matrix while the previous one was still waiting for a runner, and both then competed for the same macOS capacity. Several open pull requests are currently showing duplicated test, linux and windows checks from exactly this. Superseded runs are now cancelled.

Notes

• npm test is pinned to shell: bash — its script globs scripts/*.test.ts, and the Windows runner's default pwsh neither expands the glob nor strips the quotes.
• cargo test uses working-directory rather than cd … && …, so it does not depend on the shell's operator support.
• The Rust toolchain step already installs both Apple targets for macos-latest, and the artifact upload already globs src-tauri/target/*/release/bundle/…, so both keep working unchanged.
• Also ignores dist/, .cache/ and .worktrees/, none of which have tracked files today.

Validation

• Workflow parses; step order confirmed as install → npm test → cargo test → build across all three matrix entries.
• Both suites were run on a Windows runner ahead of this PR, on a branch carrying the tab-menu and PDF fixes: npm test and cargo test both pass there. That was their first execution on Windows and it found no platform-specific failures — the coverage this PR establishes starts from a green baseline.