fix(poweraccent): Add injection hygiene and focus-loss cleanup

[crutkas]

Summary

Add injection hygiene and focus-loss cleanup to PowerAccent (Quick Accent) to prevent cross-module stuck keys and stuck accent mode.

Note: PR microsoft#46593 on microsoft/PowerToys is also touching PowerAccent with shift key state fixes. Coordinate before merging to avoid conflicts.

Changes

D1: Add LLKHF_INJECTED filter to LL keyboard hook

Skip injected events in LowLevelKeyboardProc. Prevents Keyboard Manager remapped keys from triggering accent mode — e.g., if KBM remaps Q to A, pressing Q should not activate the accent toolbar for 'A'.

D2: Tag all SendInput output with dwExtraInfo

• Added POWERTOYS_INJECTED_TAG = 0x110 constant
• Set dwExtraInfo on all KEYBDINPUT structs (backspace, character insertion)
• Added SendInput return value checking with error logging
• Other hooks (centralized, KBM) can now identify PowerAccent output and skip it

D3: Add ForceReset() for focus-loss cleanup

• New ForceReset() method on KeyboardListener clears all state: letterPressed, m_toolbarVisible, trigger flags, shift flags
• Exposed through PowerAccent.Core.ForceResetKeyboardState()
• Wired into Selector.OnDeactivated — accent mode now cleans up when the window loses focus (Alt+Tab, click away, etc.)

D4: Replace SendKeys.SendWait with tagged SendInput

• Replaced SendKeys.SendWait("{RIGHT}") / "{LEFT}" with WindowsFunctions.SendArrowKey()
• Uses tagged SendInput instead of blocking SendKeys.SendWait
• Arrow key events now carry dwExtraInfo so other hooks identify them as synthetic

How to Reproduce

D1 — Cross-module trigger:

1. In KBM, remap Q to A
2. Press and hold Q, then press Space
3. Before fix: Accent toolbar appears for 'A' (triggered by KBM-injected input)
4. After fix: Accent toolbar does NOT appear (injected input filtered)

D3 — Stuck accent mode:

1. Trigger accent toolbar (hold a letter with accents, press Space)
2. While toolbar is visible, press Alt+Tab to switch apps
3. Before fix: Accent state stays active, Space/arrows swallowed globally
4. After fix: Toolbar hides, state resets on deactivation

Coordination

• PR [QuickAccent] Fix UI glitches, DPI-related issues, selection bugs, and add hardware shift key state fallback microsoft/PowerToys#46593 on microsoft/PowerToys is also modifying PowerAccent shift state handling. Review both PRs together to avoid conflicts.

PR Checklist

• Builds clean on x64 Release
• Manual verification per repro steps
• No interference with KBM when both active

[yeelam-gordon]

Incomplete locking around the new m_state_mutex.

m_state_mutex (KeyboardListener.h L69) is acquired in ForceReset() (KeyboardListener.cpp L54-68), but not in OnKeyDown / OnKeyUp, which mutate the same fields letterPressed, m_toolbarVisible, m_triggeredWithSpace, m_triggeredWithLeftArrow, m_triggeredWithRightArrow, m_leftShiftPressed, m_rightShiftPressed.

OnKeyDown / OnKeyUp run on the LL keyboard hook thread. ForceReset() is called from Selector_Deactivated (Selector.xaml.cs L60), which is a WPF window event runs on the Dispatcher (UI) thread.

So this is real cross-thread mutation: the hook-thread writers race the UI-thread reset. Without a guard on OnKeyDown/OnKeyUp the reset can observe partial state, miss writes, or race a flag flip and on the hook thread we can lose a write that ForceReset just did.

Two safe fixes:

1. Add std::lock_guard<std::recursive_mutex> lock(m_state_mutex); at the top of OnKeyDown / OnKeyUp. The mutex is already recursive_mutex, so any callback into PowerAccent.cs that ends up calling back is safe.
2. Or, since these are tight LL-hook paths, convert the flags to std::atomic<bool> and std::atomic<LetterKey> and drop the mutex on the reset side. Lock-free is friendlier for the hot path.

Either way, the current "lock only one writer" pattern is the canonical incomplete-locking shape and undoes most of the safety the mutex was meant to add.

[crutkas]

Went with the atomics option and dropped the mutex.

Locking OnKeyDown/OnKeyUp would deadlock: their PowerAccent.cs callbacks Dispatcher.Invoke synchronously onto the UI thread, and ForceReset (now wired to Selector_Deactivated) runs on that same UI thread — so holding a lock across the hook callbacks would have the hook thread waiting on the UI thread while the UI thread waits on the lock. The shared flags and letterPressed are now std::atomic instead, which is race-free and lock-free on the hot hook path.

I also removed a broad LLKHF_INJECTED filter that had been added here — same accessibility regression as in the centralized-hook PR (it blocked OSK/AT/macro injection). The dwExtraInfo tagging on PowerAccent's own injected keys stays, so the centralized keyboard hook still ignores them.

[crutkas]

Superseded by the upstream PR against microsoft/PowerToys: microsoft#48572 (rebased onto current main, builds/tests verified). Closing this intra-fork review PR.