Main window: stop timers and undo polling on close, start them on Loaded (UI suite 2m20s → ~55s)

[niksedk]

What

Follow-up to #14591: hunting the thread growth in the UI test suite with per-test dotnet-stack dumps.

Findings

Threads	Where	Cause
318	UiTickPump.Loop	MainViewModel started its 50 ms position pump and 16 ms cursor pump in the constructor; a view model whose window closed without OnClosing (every test host, which detaches the save prompt) or that never had a window kept them forever. The 400 ms slow timer (a full GetFastHash pass) ran on the shared UI thread for every one of them too.
149	Dispatcher.Invoke in UndoRedoManager.CheckForChanges	The 333 ms change-detection poll of every closed editor window kept firing; its blocking Invoke parked one pool thread per window for the rest of the process. This is the starvation behind the OverlappingTick flake and the 15 s mpv test in #14591.

The same leak exists in the app for File > New window: closing the second editor window left its pumps and undo poll running until exit.

Change

• StartBackgroundWork() runs from OnLoaded instead of the constructor (the timers have nothing to drive before the window is up).
• StopBackgroundWork() (timers + StopChangeDetection) runs from CleanUp and from the host window's Closed event, so it also covers a close that bypasses OnClosing.
• The delayed StartChangeDetection in OnLoaded is skipped when the window is already gone.
• Test speed-ups on top: MultipleReplaceViewModel.PreviewIntervalMs (internal) lets the tests wait 25 ms per preview settle instead of 250 ms; the fix-common-errors "Analysing…" delays are internal statics the unfixable-errors tests zero.

Numbers (local, full UITests)

	Before	After
Wall time	2 m 20 s	54 s – 61 s
Sum of test durations	125 s	~60 s
ThreadPool threads at end	~200	9
Multiple-replace test family	18.5 s	6.5 s
SubtitleGridScrollPerformanceTests	11.6 s	2.2 s

Two full runs clean (4895 passed, 1 skipped).

🤖 Generated with Claude Code