fix(tray): stop auto-fit ResizeObserver feedback loop

[MiguelMachado-dev]

Summary

• Stop auto-fit measurement style changes from recursively scheduling another layout pass through the tray surface's ResizeObserver.
• Keep observer-driven sizing enabled until the initial layout converges, so a newly created flyout still reaches its normal size.
• Add a deterministic style-feedback regression test that fails on main before the hook fix and settles after it.

The auto-fit pass temporarily removes height and overflow constraints to measure the full surface. On the affected Windows/WebView2 setup, the observer saw those programmatic changes and re-entered auto-fit indefinitely, making the Pop Out Dashboard alternate between measured and committed layouts. This is a separate feedback path from the fractional-DPI native-size cycle handled in #272.

Related issue

Follow-up to #261 and #272.

Affected areas

• Tray panel
• Settings UI
• Config file / settings persistence
• CLI
• Provider-specific behavior
• Installer / release packaging
• Startup / background behavior
• Documentation
• Other: detached Pop Out Dashboard flyout

Validation

• powershell.exe -ExecutionPolicy Bypass -NoProfile -File scripts\local-check.ps1
  • Shared Rust: 1,293 tests passed
  • Tauri Rust: 353 tests passed
  • Frontend: 44 files / 269 tests passed
  • Frontend TypeScript/Vite build passed
  • Formatting and clippy gates passed
• Full pre-release validation: not applicable; no packaging or release changes
• Installer/release validation: not applicable
• Thermo-nuclear code quality review completed before submitting
• Focused red/green test: the new test fails before the hook change (3 reveal calls instead of settled 2) and all 3 sizing tests pass afterward
• Fresh Windows debug build: pnpm --dir apps\desktop-tauri run tauri:build:debug

UI / tray proof

• Not applicable
• CUA Driver visual proof attached
• CUA Driver could not be used; equivalent manual proof and explanation attached

CUA Driver is not installed on the test machine (%LOCALAPPDATA%\Programs\Cua\cua-driver\bin\cua-driver.exe is absent), so I used equivalent native Windows/WebView2 proof against a fresh debug build from this branch:

• 12-second CDP sample, 2,160 animation frames: 0 body, surface, max-height, or native inner-height transitions; 1 ResizeObserver callback
• 14-second desktop capture, 180 frames: 1 unique frame, 0 frame transitions, 0 single-frame flicker pulses, 1 native window geometry
• Flyout remained 328 x 920
• No bogus tiny remembered flyout size was persisted

For comparison, the affected build produced roughly 160 body/surface height transitions and about 240 observer callbacks in 12 seconds, plus 37 one-frame visual pulses in 12.6 seconds.

Notes for reviewers

The layoutReadyRef.current guard is intentional: observer callbacks remain active during first-open convergence, then callbacks caused by the hook's own measurement pass are ignored while programmaticInFlightRef is active. This preserves initial auto-fit and avoids misclassifying the hidden window's tiny bootstrap frame as a user-resized size.

The diff is limited to useTrayPanelLayout.ts and its sizing test; it contains no Rust/backend or local process-discovery changes.