Testing

Four layers, from fastest to slowest.

Unit / component (npm test)

Vitest, with resolve.conditions: ['browser'] under Vitest so Svelte 5's real browser build resolves under jsdom — a .test.ts can mount() a component directly rather than regex-scanning source. Controllers (reviewController, clusterController, slotGalleryController, ingestRunController) are tested directly, not by mounting the whole page.

Stubbed end-to-end (npm run test:e2e)

e2e/stubbed/ — Playwright + pytest against a vite preview build, with every {API_PREFIX} request intercepted by a fail-closed stub (an unregistered request gets a 501 and fails the test, rather than a silent 200 {}). Never touches a live backend.

Live read-only tier (npm run test:live)

e2e/live/ — drives the real deployed build against a real OpenProcessor backend, structurally read-only (every non-GET/HEAD request is aborted and any attempt fails the test). Skipped entirely unless OP_CROPWRIGHT_LIVE_URL is set — never runs in CI. Also saves full-page screenshots per route at two viewports; a human must actually look at them after a run — the assertions catch what's mechanically checkable, not a layout regression.

Mutation testing (npm run test:mutation)

Stryker, against the ~10 highest-value pure modules (wire mapping, undo store, dataset stats, review tabs, curationSettings.ts, strategies.ts, classPicker.ts, readSlot.ts). Answers "if I break this line, does a test notice" rather than "does every assertion pass." Runs on a schedule, not per-push — too slow for that.

New test discipline

Verify a new test actually fails against a mutated copy of the code it covers before trusting it (edit a scratch copy, confirm red, restore byte-for-byte).