Desktop Smoke E2E shards 3 and 4 were red on this branch's CI while the
branch touches neither spec. Both failures are pre-existing races in the
specs themselves, so fix them where they are.
Shard 3 — "profile hover uses the channel hover surface" sampled the
channel row's background straight after hover(), which lands mid
`transition-colors` fade (rgba(0, 0, 0, 0.027) on the failing run). The
profile card then settles on the real hover colour, rgba(0, 0, 0, 0.04),
so toHaveCSS could never match and all three attempts failed. Poll the
reference element until two consecutive reads agree on a value that has
moved off its resting colour. Reproduced locally at 1-in-5 before the
change; 12 consecutive runs green after.
Shard 4 — "round-trips and reopens structured message-text conditions"
failed two ways. The CI page snapshot shows the reopened dialog in YAML
mode holding `filter: str_ends_with(trigger_text, "Workflow
notification")` and a textless send_message step: addMessageStep filled
"Message text" before the inspector handed over from the trigger to the
new step, so the step body overwrote the trigger condition. That
definition no longer parses as a form, getInitialEditorMode falls back
to YAML, and the trigger controls the test waits on never render — a
30s test timeout. Wait for the step inspector's own Action control
before filling, and pin the saved YAML right after create so any
recurrence fails there instead of stalling.
The remaining attempt measured the three Match operator buttons with
sequential boundingBox() calls while the inspector was still animating
in (Framer Motion, width 0 → 26rem and x 24 → 0 over 240ms), reporting
the second button 44px left of the first. That animation is driven off
requestAnimationFrame, so document.getAnimations() is empty and
waitForAnimations cannot see it; read the boxes from one frame via
evaluateAll and poll until the geometry stops moving.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Matt Toohey <contact@matttoohey.com>
01ea7e5824