perf: light background mode, idle input cadence and performance measurement

[nextestudios]

Summary

Makes ControlFS light enough to sit next to a game, measured before and after.

• Measurement (build/Measure-Performance.ps1, new step in the manual Smoke workflow, full mode only; no new workflows or triggers): launches the portable exe (first run with an empty extraction dir + 3 warm runs) and the installed exe, and records ms to "Janela ativada" / "Primeiro quadro" (new precise log lines), CPU %, working set, private bytes and threads over 10 s idle in the foreground and 10 s minimized (ShowWindow(SW_MINIMIZE)), plus package sizes. Table goes to the job summary and the performance artifact; a run whose app exits mid-measurement fails the step.
• Bug found by the measurement: minimizing the window crashed the app on main (fail-fast E_INVALIDARG from a WinUI AppWindow.Changed handler reading the minimized title bar; dump analysis in the smoke log). Fixed: title-bar updates skip the minimized state and never let an exception escape that event; minimizing no longer relayouts.
• Idle cadence (InputCadence): the 8 ms ticker + 1 ms Windows timer now run only with the window in front and a controller connected; with no controller the UI timer polls at 250 ms just to notice one arriving (keyboard/mouse come through XAML events).
• Light in the background (BackgroundModePolicy, new setting Leve em segundo plano, on by default): when the window is inactive or minimized (a minimized sole window stays "active", so the presenter state counts too): BelowNormal priority + EcoQoS, controllers only watched for connections every 1 s with SDL input events disabled (a game using the same pad no longer floods our queue; stick/trigger state is resynced silently on return), drive polling paused (checked immediately on return), and after 5 s a single aggressive GC + LOH compaction + working-set trim. Activation restores at once; media playing keeps normal priority.
• Startup/memory: automatic update check deferred until after the first frame; icon caches capped by size (details panel 32, cards/nav 128, tiles 256, rows 512); logo decoded at display size. Already lazy and verified: LibGit2 native (only with Git status on), Media Foundation/PDF (first use). SDL init measured 16–32 ms, kept on the critical path.
• Packaging decisions (measured on experiment branches, not merged): ReadyToRun off (+21 MB portable / +9 MB installer, no startup gain); single-file compression kept (171 MB vs 70.5 MB uncompressed; only the first launch pays ~0.25 s).

Before / after (GitHub windows-latest, WARP, no controller; median of 3 Smoke runs, each the median of 3 warm launches; installed in parentheses)

Metric	Before (main + crash fix)	After
To window (ms)	1031 (933)	901 (851)
To first frame (ms)	1274 (1169)	1099 (1058)
Idle CPU, foreground (% of one core)	1.72 (1.41)	0.47 (0.62)
Working set, foreground (MB)	155.3 (155.1)	152.8 (152.4)
Private bytes, foreground (MB)	61.7 (62.2)	60.5 (60.6)
Threads, foreground	35 (36)	34 (35)
CPU minimized (% of one core)	2.03 (1.41)	0.47 (0.47)
Working set minimized (MB)	153.1 (152.9)	17.9 (17.8)
Private bytes minimized (MB)	61.7 (62.3)	55.7 (55.7)
Threads minimized	34 (35)	32 (33)
Portable / installer / installed folder (MB)	70.5 / 46.6 / 182.4	70.5 / 46.6 / 182.4

Pure main could not be measured minimized (it crashed). Runner-to-runner noise is large (the same build measured 543–1100 ms to window), so the startup difference is within noise. Details and methodology in docs/performance.md.

Tests

• CI: InputCadenceTests (2), BackgroundModePolicyTests (2), UpdateFlowTests::Automatic_check_waits_for_the_first_frame_when_deferred.
• Full Smoke (launch, installer, UIA, dump analysis, measurement) on this branch: green in runs 36376541012, 36378316173, 36378320665. Run 36377487211 failed only in the dev-only glyph gallery of --render-screens (UIElementCollection.Add E_ABORT at the very end), a flake also seen on the measurement-only commit (36375709199) and not touching code changed here.
• Manual (no controller/GPU on the runner): gaming with the same controller while ControlFS is behind, EcoQoS leaf in Task Manager, no ghost action on return, controller arrival within 250 ms, long extraction and music while minimized — added to docs/TESTING.md ("Leve em segundo plano e repouso").

Docs: CHANGELOG pair (Desempenho/Performance + fix), GUIDE pair (new setting), docs/ui-redesign.md row, docs/performance.md rewritten.

🤖 Generated with Claude Code