The commands live in
CONTRIBUTING.md and are run from there.
This file records what each flag is defending against, so nobody drops one that
looks redundant and reintroduces the failure it was added for.
Exposes gpu::test_utils, blocking_read(), and the engine's test_readback_*
accessors that integration tests rely on. The feature is the compile-time gate
enforcing No Blocking GPU
Readbacks: without it, production
code cannot even name those functions.
Mandatory, not a preference. GPU-touching integration tests (engine.rs,
blend_modes.rs, and friends) share a process-wide wgpu device and SIGSEGV when
run in parallel.
tsc --noEmit is the TypeScript gate for .ts files, but it cannot see inside
.svelte files: it does not parse the extension, and neither vite build nor
Vitest type-checks components. svelte-check (npm run check) is the only gate
that type-checks .svelte scripts and templates, via svelte2tsx plus the
TypeScript API, and it catches nonexistent engine methods, wrong props, and
null-safety in components. Both are required: tsc alone gives a false green on
component bugs.
Vitest runs in the node environment by default, so there is no DOM and globals
like KeyboardEvent, PointerEvent and window are undefined. Test against
plain object fakes ({ key, shiftKey } as KeyboardEvent), and for code that
touches window, stub it with vi.stubGlobal('window', …) and a fake node: see
src/lib/__tests__/dismiss.test.ts.
A file opting in with // @vitest-environment jsdom gets a DOM and can mount
a component, which is how behaviour that only exists in how a component
sequences DOM reads and writes is covered: see
addLayerModal.component.test.ts.
jsdom has no layout, so it computes no geometry, runs no animation frames and
implements neither showModal nor scrolling. A component test that needs any of
those stubs them and supplies its own canned rects, which is a decision about
what the test controls rather than a workaround: see
brushExplorer.open.component.test.ts.
Anything whose cause is real layout (scroll anchoring, percentage resolution,
display: none semantics) cannot be reproduced here at all.
It is also the staleness gate for catalog graphics: it re-renders each one and fails if the committed image no longer matches its component and stills. See generated-markdown.md.
Not a check, but the reason target/ balloons. Cargo orphans a ~300 MB static
test binary on every fingerprint change and never garbage-collects it.
cargo install cargo-sweep once, then periodically:
cargo sweep --time 7