Harden ai-assistant scrolling tests against a scroll-settle race

[habdelra]

The Integration | ai-assistant-panel | scrolling suite intermittently fails on scrolling stays at the bottom if a message is streaming in with AI assistant is scrolled to bottom expected true, got false (failing CI job).

Root cause

The "open a room" assertions read the conversation's scroll position synchronously, immediately after the room renders. The panel auto-scrolls to the newest message when the last message registers its scroller, and re-scrolls whenever that message's subtree mutates (see the per-message MutationObserver in ai-assistant/message). Avatars, card pills, and markdown that finish rendering after the test runloop has otherwise settled shift the layout and briefly move the scroll position before the component re-scrolls to correct it. A single synchronous read races that correction and occasionally samples a position past the bottom threshold. The failing run shows this directly — the avatar set-background-image helper module 404s and retries at +100ms, i.e. late layout work landing after settled().

Only the first assertion in the test can fail this way: simulateRemoteMessage delivers events via setTimeout(0) and Glimmer batches renders, so the two assertions that follow a simulateRemoteMessage with no await read stale DOM (still at the bottom) and pass trivially.

Fix

Replace the point-in-time scroll reads in the "open a room" assertions with a waitUntil poll that tolerates the late re-scroll. The predicate is unchanged, so an already-settled panel resolves immediately; only the racing case now waits for the component's own correction to land. On timeout the helper reports the exact geometry (scrollHeight / clientHeight / scrollTop / distanceFromBottom / bottomThreshold) so a future failure is diagnosable instead of a bare expected true.

The same race exists in the sibling "open a room" checks (scroll-to-last, scroll-to-first-unread, and the unread-indicator click, which previously used a fixed 2s sleep), so they move to the same helper. The two checks that verify enqueuing a streaming event doesn't synchronously jank the scroll stay synchronous by design.

Verification

Test-only change; eslint and ember-tsc are clean for the file. The underlying flake is timing- and CI-load-dependent and doesn't reproduce reliably locally, so the diagnostics are included as insurance: if the poll ever times out, the next failure will print the exact scroll geometry.

[github-actions]

Preview deployments

• Host staging preview

• Host production preview

Host Test Results

    1 files      1 suites   2h 36m 2s ⏱️
3 376 tests 3 361 ✅ 15 💤 0 ❌
3 395 runs  3 380 ✅ 15 💤 0 ❌

Results for commit 70e06a6.

Realm Server Test Results

    1 files      1 suites   9m 7s ⏱️
1 674 tests 1 674 ✅ 0 💤 0 ❌
1 753 runs  1 753 ✅ 0 💤 0 ❌

Results for commit 70e06a6.