fix: reject empty recordings on retry, restore push-to-talk under compositor shortcuts

[xAlcahest]

Summary

The double-trigger behind #864 left a lasting scar: header-only recordings (~110 bytes) saved to history by older builds fail every Retry/Recover with the same confusing "Audio file might be corrupted or unsupported", because the retry-transcription handler in ipcHandlers.js ships the stored buffer to a backend with no validation. The handler now rejects buffers smaller than MIN_AUDIO_BYTES (reusing isEmptyRecording from recordingGuard.js) and returns a clear "Recording contains no audio" instead.

The reconcileNativeKeyListeners gate from #1001 also turned out to kill push-to-talk on KDE/GNOME/Hyprland: it blanks the native key listener whenever the compositor owns the hotkeys, but in push mode the compositor callback intentionally early-returns to defer to that same listener, so holding the hotkey did nothing at all. getNativeListenerKeys now takes a compositorOwnsHotkeys flag and keeps just the dictation key watched in push mode; tap slots stay compositor-delivered, so the original double-trigger cannot come back. Verified on KDE Plasma Wayland (Fedora 44): tap unchanged, hold works again.

Fixes #864 #847

Changes

• src/helpers/ipcHandlers.js: reject stored audio buffers smaller than MIN_AUDIO_BYTES in the retry-transcription handler instead of sending them to a backend
• src/helpers/hotkeyManager.js: getNativeListenerKeys takes a compositorOwnsHotkeys flag and keeps only the push-to-talk dictation key when a compositor delivers hotkeys
• src/helpers/windowManager.js: reconcileNativeKeyListeners passes isUsingNativeShortcut() instead of blanking all keys
• test/helpers/recordingGuard.test.js: add a require() interop test for the main-process guard
• test/helpers/nativeListenerKeys.test.js: add compositor carve-out cases

Supersedes #872.