Replace playwright-test/uvu with the libuild test runner

[brainkim]

Moves the core suite from playwright-test --runner uvu to libuild test, so crank consumes libuild's test stack alongside its build stack.

Verified on a clean install of libuild 0.2.17, no patching: 602 passed, 3 skipped, exit 0.

What changed

before	after
uvu suite()	describe()
test.before.each / test.after.each	beforeEach / afterEach
uvu/assert (Assert.is, …)	expect(...)
playwright-test --runner uvu	libuild test test/ --platform <browser>
playwright-test (bundled browsers)	playwright (explicit playwright install)

playwright-test.config.js is deleted; CI's browser matrix (chromium/firefox/webkit) is unchanged and now drives --platform.

Each file is wrapped in describe() rather than left flat. That is not cosmetic: uvu's suite() scoped hooks per file, but libuild bundles all 30 files into one root suite, so top-level beforeEach hooks would run before every test in the suite. Left flat, all 602 tests failed with Sinon's "Attempted to wrap error which is already wrapped" — 26 files' hooks each stubbing console.error on top of one another.

Three bugs the conversion shook out

test/jsx-tag.ts has never run. playwright-test.config.js declared input: ["test/*.tsx"], which does not match the one .ts test file. Its 25 tests were silently excluded from every CI run. Including them takes the suite from 577 to 602. They pass.

One assertion was never actually checked. uvu's throws(fn, expects, msg) treats a string second argument as the failure message, not a matcher — only a RegExp or function is compared. So:

Assert.throws(() => renderer.render(<Component />, document.body), "Context iterated twice");

asserted only that something threw. expect().toThrow(string) does substring-match, which surfaced that the real message is <Component> context iterated twice without a yield — lowercase, and differently worded. The expected string is corrected.

Two spots leaned on assertion signatures. uvu types ok(val): asserts val, so Assert.ok(x instanceof Element) narrowed x for later lines. expect().toBeTruthy() does not narrow, so mathml.tsx and svg.tsx now access those values explicitly.

Fidelity

• 602 test() declarations and 3 test.skips before and after — nothing added or lost. The run reports exactly 602 passed, 3 skipped.
• Assertion messages, which expect() has no parameter for, are kept as block comments instead of being dropped (expect(x).toBe(1) /* cleanup should be called */;).

libuild bugs this found, fixed in 0.2.17

This PR was blocked on two of them; both are now fixed upstream and the workarounds are gone.

1. The browser runner started before any test registered. @b9g/libuild/test picks its backend with top-level await, making it an async module, so every importing test file's body is deferred past the runner's queueMicrotask start. It ran with zero tests and reported a green empty run. Fixed by having the generated entry set a ready flag as its last statement, which ESM guarantees runs after every imported file has evaluated.

2. A zero-test run exited 0. Merging this against 0.2.16 would have silently disabled crank's entire suite with CI green. Now a hard failure — verified here: an empty selection exits 1 with "test files were bundled and loaded, but zero tests registered".

Also fixed in 0.2.17 and used here: test.skip exists in the browser runner (the three disabled hydration tests use it again instead of a local no-op, and are reported as skipped), and libuild test test/ discovers files correctly when the root is itself named test/ (so discovery is the plain directory form rather than explicit globs).

Verification

602 passed, 3 skipped, 0 failed on chromium against a clean install with no patching, exit 0. tsc --noEmit clean, eslint clean. A deliberately empty selection exits 1.

🤖 Generated with Claude Code

https://claude.ai/code/session_019pggktip8wsuxzy2VCY9p7

[brainkim]

@b9g/libuild@0.2.17 is published and should unblock this draft — both blockers are fixed, along with the minor note:

Blocker 1 (runner fires before registration): fixed, but not with the setTimeout patch — any scheduler-based start still loses to a test file whose top-level await crosses a macrotask (its later tests silently vanish from a green run). Instead the generated entry now sets a ready flag as its last statement — ESM guarantees that runs only after every imported file, TLA continuations included, has fully evaluated — and the runner waits on it. Your 602-passing conversion was produced under the equivalent patch, so it should pass unchanged; the local libuild patch can be dropped.

Blocker 2 (zero tests exits 0): a browser run that discovers files but registers zero tests now exits 1.

Minor (libuild test test/ finds nothing): fixed — when the discovery root itself is named test/tests/__tests__, files without the .test. infix (e.g. test/dom.tsx) are discovered. The explicit-glob invocation in this PR also still works.

Two things that may affect the diff here:

• The browser runner now has test.skip/test.todo/describe.skip (they register nothing and tally as skipped), so the local no-op skip() shim for the three hydration tests can go.
• Browser runs are stricter than what this PR was tested against: uncaught page errors now fail the run (previously they were log lines while the run stayed green), and a bundle that breaks before the runner starts is a reported failure. If anything in the suite was quietly relying on a swallowed page error, it will surface as red — that's the fix working, not a regression.

Full details in the release notes: https://github.com/bikeshaving/libuild/releases/tag/v0.2.17