macos: fix Option+Arrow chords + modifier-state corruption in repeat-task cleanup

[tyvsmith]

Fixes #440.

Two related bugs in the macOS input-emulation backend, each addressed in its own commit so they can be reviewed and bisected independently.

Commit 1 — macos: post NumericPad and SecondaryFn flags for synthesized arrow keys

Hardware-generated arrow key CGEvents on macOS carry kCGEventFlagMaskNumericPad and kCGEventFlagMaskSecondaryFn in addition to any user-pressed modifiers. CGEventTap-based hotkey matchers (tiling window managers, accessibility tools) often check for those flags to disambiguate navigation keys from generic chords. Synthesized arrow chords without them fall through to the focused app instead of being captured by the WM.

Fix: set both flags on arrow key events (Mac key codes 0x7B–0x7E).

if is_arrow_key(key) {
    flags |= CGEventFlags::CGEventFlagNumericPad | CGEventFlags::CGEventFlagSecondaryFn;
}

Commit 2 — macos: stop corrupting modifier state in repeat-task cleanup

spawn_repeat_task() takes a Mac CGKeyCode, but the cleanup block was passing that value to update_modifiers(), which expects a Linux evdev scancode (it calls scancode::Linux::try_from(key)). The two codespaces collide on several values, silently clearing still-held modifiers when the repeat task is cancelled by a follow-up key:

Mac code	Mac key	Linux scancode	Cleanup effect
56	LeftShift	KeyLeftAlt (56)	clears Mod1Mask
125	Down arrow	KeyLeftMeta (125)	clears Mod4Mask
126	Up arrow	KeyRightMeta (126)	clears Mod4Mask

This made Shift+Option+X arrive as Shift+X and Cmd+Up/Cmd+Down arrive without Cmd — see issue #440 for the full trace.

Fix: remove the buggy update_modifiers call. The main consume() loop already updates modifier state with the correct Linux scancode on the real release event.

Verification

• Built with cargo build --workspace and cargo bundle --release on macOS 26 (Sequoia, arm64)
• cargo clippy --workspace --all-targets --all-features: clean
• cargo fmt --check: clean
• cargo test --workspace: passes (no unit tests for the macOS backend; AGENTS.md says "OS-specific backends: test via GTK/CLI on target OS or document manual verification")
• Manual verification on a Mac receiver with OmniWM and a Linux sender:
  • ✅ Option+Arrow → WM focus moves (was: focused app got the arrow)
  • ✅ Shift+Option+Arrow → both modifiers in flags
  • ✅ Option+1/2/3/4 → WM workspace switch (regression check, still works)
  • ✅ Shift+Option+1/2/3/4 → both modifiers in flags (was: only Shift)
  • ✅ Cmd+Up/Cmd+Down → Cmd in flags (was: missing)
  • ✅ Plain typing and Down-arrow auto-repeat behave normally (regression checks)

Notes / follow-ups deliberately left out of this PR

• Other navigation keys (Home/End/PgUp/PgDn/ForwardDelete/Help/F-row) also carry SecondaryFn on hardware. They aren't part of the reported symptom, but a follow-up could extend the same pattern.
• Numeric-keypad number keys also carry NumericPad. Same situation.
• The cleanup still posts a KeyUp for the cancelled key even when the user hasn't released it (because a different key was pressed). This is independently questionable behavior but doesn't visibly misbehave once the codespace bug is fixed, so left alone.

[Copilot]

Pull request overview

Fixes two macOS input-emulation issues that affected modifier correctness and global hotkey recognition in CGEventTap-based tools (issue #440), by aligning synthesized events more closely with hardware-generated events and preventing accidental modifier-state mutation during repeat-task cleanup.

Changes:

• Stop calling update_modifiers() in repeat-task cleanup (it expects Linux evdev scancodes, but the repeat task is keyed by macOS CGKeyCode).
• Add detection for arrow-key macOS virtual keycodes and attach NumericPad + SecondaryFn flags to synthesized arrow-key events.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[feschber]

Thanks for catching this! I've been wondering whats up with the arrow keys ...
We should probably still release the key repeat task but with the correct code.