test(chat): guard session flyout identity

[shanselman]

Summary

Adds real WinUI regression coverage for #970. Current main already contains the production fix from #991 and #984: the composer uses a declarative ContentFlyout, and FunctionalUI preserves the native flyout by render path.

The new test exercises the actual UiRenderer, Button, and WinUI Flyout lifecycle. It proves repeated Agent thinking-style status renders keep the picker open, session rows and selection still reconcile, the latest callback runs once, and mark-and-sweep pruning closes and releases a removed picker.

CI hang root cause and hardening

The first PR run (29777215014, job 88469448341) reached xUnit startup but emitted no completed test or TRX before the one-hour cancellation. The test had an unbounded _ui.RunOnUIAsync around Flyout.ShowAt; if WinUI stalled opening the popup against UIThreadFixture's default off-screen 1x1 host, the later five-second Opened timeout was never reached.

The test now owns its host geometry before ShowAt, bounds every UI-dispatch await to ten seconds, awaits popup closure during cleanup, detaches the flyout/content, restores the shared host geometry, and avoids creating an extra post-prune flyout. A future UI stall fails with a bounded diagnostic instead of consuming the job timeout.

Validation

Current head: 8c3562d9a73cbe0a6aecd383c827f85a0e1117f2

• .\build.ps1 — passed; all five source projects built.
• dotnet test .\tests\OpenClaw.Shared.Tests\OpenClaw.Shared.Tests.csproj --no-restore — 2,927 passed, 31 skipped, 0 failed.
• dotnet test .\tests\OpenClaw.Tray.Tests\OpenClaw.Tray.Tests.csproj --no-restore — 1,834 passed, 0 failed.
• dotnet test .\tests\OpenClawTray.FunctionalUI.Tests\OpenClawTray.FunctionalUI.Tests.csproj -r win-x64 — 19 passed, 0 failed.
• Focused FunctionalUiFlyoutIdentityTests — 1 passed, 0 failed.
• Exact CI command, complete non-accessibility Tray UI suite under dotnet-coverage — run 1: 89 passed in 6.56s; run 2: 89 passed in 6.71s.
• Hosted Build and Test run 29782799246, Run Tray UI Tests — passed in 19s on the updated head.
• Red/green mutation proof: temporarily replaced CreateContentFlyout path caching with new Flyout() on each render; the focused test failed at native flyout identity. Production was restored with no production diff.
• Final rubber-duck review — clean, no actionable findings.
• Structured Codex autoreview — blocked before review by the helper's isolated runtime refusing its temporary CODEX_HOME; no source finding was produced.

Exact repeated command:

dotnet-coverage collect --output TestResults\TrayUI\coverage.cobertura.xml --output-format cobertura "dotnet test tests/OpenClaw.Tray.UITests --no-build -c Debug -r win-x64 --verbosity normal --results-directory TestResults\TrayUI --logger trx;LogFileName=OpenClaw.Tray.UITests.trx --filter Category!=Accessibility"

Real behavior proof

• Ran the focused test with SLOW_UI_TESTS=1 in the real WinUI test host.
• Computer Use observed the live window titled Session picker still open after thinking renders.
• The accessibility tree simultaneously showed Agent thinking 3 and an open native popup containing Session Alpha and Session Beta; the run then passed row update, selection callback, and prune cleanup assertions.
• An lvt capture independently identified WinUI 3 PopupRoot / FlyoutPresenter and the session rows while the picker was open.
• Not verified / blocked: a gateway-mediated live chat recording was not collected because the isolated deterministic test host has no paired gateway. The current-head real-WinUI proof directly drives the renderer behavior responsible for Windows Companion: "Agent thinking" animation closes the chat/session selectbox before a different chat can be chosen #970 without depending on gateway timing.

Closes #970.