Rework swipe-gesture recognition and speed up HID++ device detection

[laziukdavid]

Summary

This PR adds two core reliability/performance reworks, plus a few smaller
improvements, all developed and tested against an MX Master 4 (Logi Bolt
receiver, receiver PID 0xC548).

Core changes

Swipe-gesture recognizer rewrite

The old gesture-button swipe logic summed every movement delta into one running
total, fired when it crossed gesture_threshold, then blocked input for a
fixed cooldown. That dropped inputs (the cooldown swallowed the start of the
next swipe), produced false positives (slow drift summed past threshold; the
return stroke between two swipes fired the opposite direction), and made
back-to-back repeats impossible.

• New core/gesture_recognizer.py — a pure, thread-safe GestureRecognizer
  that segments motion into strokes instead of integrating one global sum.
  A stroke commits a swipe on threshold travel along one axis within a commit
  window and within an off-axis ratio; a peak detector fires every fresh flick
  while return strokes only re-arm it; a pause clears the lock. No cooldown.
• Shared by the macOS, Windows and Linux hooks — removes the triplicated
  _accumulate_gesture_delta / _detect_gesture_event code.
• Click-jolt rejection (minimum movement-report count) and quick-flick capture
  fix phantom swipes on a tap and missed swipes during a fast flick.
• Config schema v9 → v10: retires gesture_deadzone / gesture_timeout_ms
  / gesture_cooldown_ms; adds gesture_commit_window_ms, gesture_settle_ms,
  gesture_cross_ratio. Obsolete keys in existing configs are left ignored.
• The raw gesture_threshold px slider is replaced with a discrete 5-level
  Sensitivity dropdown (stored key/units unchanged).

HID++ detection speed

Cold-start connection on a setup with several Logitech receivers was ~46s,
almost entirely spent probing the wrong devices before reaching the mouse.

• Receiver-aware device-index scan order
• Persistent device cache (device_cache.json) so a relaunch / reconnect goes
  straight to the last-good device
• Dead-slot early-out — stop probing a slot once its control-count read fails
• 600ms discovery probe timeout applied across the whole discovery phase
  (operational requests keep the 2s timeout)
• Candidate dedup — a multi-interface device (e.g. a webcam exposing vendor
  usage pages) is opened once, not once per interface

Additional changes

• Screenshot action — new screenshot action: Cmd+Shift+4 on macOS,
  Win+Shift+S on Windows.
• Application Log panel — core/log_setup.py ring-buffer log handler with
  listener fan-out; appLog property on the QML backend; a new
  "Application Log" section in the Debug Events panel showing captured console
  output alongside the existing mouse-events log.
• macOS Accessibility auto-recovery — first launch after granting
  Accessibility used to leave the app dead until a manual quit/reopen. Now: a
  non-modal prompt with a deep link to the Accessibility pane, a 1s grant poll,
  and a clean re-exec once the grant lands.
• macOS build — build_macos_app.sh redirects build artifacts off network
  volumes to local disk (codesign rejects the resource-fork metadata SMB/AFP
  shares attach); auto-detects a non-local repo volume, overridable with
  MOUSER_BUILD_DIR. New install_macos_app.sh, a documented build +
  install-to-/Applications wrapper.

Testing

• Full suite green: python3 -m pytest tests/ → 555 passed, 1 skipped
• New tests: test_gesture_recognizer.py (21), plus additions to
  test_hid_gesture.py, test_config.py, test_log_setup.py,
  test_backend.py, test_macos_app_shell.py, test_key_simulator.py
• Several features are verified by tests but not yet hardware-verified across
  every platform — feedback welcome.

🤖 Generated with Claude Code