Seventeen UI tests found controls by reading screenshots with tesseract and
clicked hard-coded points, so they failed on slow CI runners (OCR timeouts,
"Allow" read before "Don't allow", dropped keystrokes) and read stale
layouts. They now use the Kryon test driver: they wait for widgets by
label, role or key, press them through the app's own pointer path, type
without losing keys, drag with a held pointer, and check what a screen
says in its widget tree. One pixel check per screen remains only where
pixels are what is tested (a scrollbar drawn, a piece's colour, bold text,
a scannable QR code). Tests that had gone stale against the app were
brought up to date. The Lumi selection, harness, guide and emphasis tests
and the navigation customizing test join the visual gate; the device link
test gets its own target. tests/rendered_word.py is gone.
Nineteen test files were run by no make target, workflow or script.
Five already ran through ziran-behavior-tests' loop; two of those and the
session results test had stopped linking (a node-app hook and curl), now
fixed. Eight more are wired into make test (frame pacing policy, Daochi
native adapter, device link, update policy, Android HTTP transport,
statistics dates, meditation audio archive, package receiver) after small
fixes to two of them. Hosted features and Lumi habit tools join the visual
gate; the Lumi performance and Sync Review in Harmony checks get their own
targets. Two tests of code that no longer exists are deleted.
make test-references-check (in make test) fails when a test entry point in
tests/ is run by no make target, workflow, hook, script or referenced test.