Four layers under app/test/:
* kernel/, extension/ — unit tests for every kernel service and
the extension contract. i18n fallback chain gets the full
matrix (exact/lang/default/placeholder, namespace isolation,
unknown namespace, interpolation). Theme resolver and
extension-manager topo sort covered.
* widgets/, builtin/ — widget tests for every primitive and each
Tier 0 built-in. Assertions reach into the Semantics node so
a missing label fails structurally, not visually.
* goldens/ — Alchemist + Ahem font for cross-platform pixel
stability; primitives only (button, tab bar, icon set). No
goldens for compositions — they'd churn through every tier.
* a11y/ — contract-level gate: semantic coverage walks the
built-in catalogue, contrast walks every token pair in every
bundled theme against WCAG-AA thresholds (catches real
regressions: the first run caught `tab.inactive_text` at 2.81
in summer-night), i18n coverage asserts every referenced key
exists in its catalog (exists-in-map, not "key == value" — too
ambiguous for keys that happen to equal their translation).
Helpers under helpers/: `KernelFixture.create()` boots a
KernelServices with in-memory stores + a FakeDaemonClient (no
socket, drivable connected-state), `widget_harness` wraps a
widget in the minimum tree that resolves theme + i18n.
Co-Authored-By: Claude <noreply@anthropic.com>
jpmschweitzerandClaude
96b26e29f0