Push-to-talk: hold the mic / floating button to record, release to send

[DevEmperor]

Idea

Today the mic button (keyboard) and the floating button are tap-to-toggle: tap once to start recording, tap again to stop + send. Add an opt-in "push-to-talk" (hold-to-record) mode: press and hold to record, release to stop and send — the walkie-talkie feel.

Inspired by the desktop app Handy, which offers both push-to-talk and toggle.

Why

• Faster for short, one-off dictations — no second tap to stop.
• Harder to accidentally leave recording running (ties into the "mic kept open" reports, Microphone kept open for a long period of time. #147): you physically hold to record, so letting go always stops it.
• Familiar interaction (voice-message style).

Proposed behavior

A setting (default off) — e.g. "Hold to record" — that, when on, switches the primary dictation trigger to hold-to-record:

• Keyboard mic key: hold to record, release → stop + transcribe.
• Floating button: press and hold to record, release → stop + transcribe (inject).
• A very short hold (below the touch-slop / long-press threshold) could still fall back to a normal tap-toggle so a quick tap isn't lost — TBD.

Key design consideration: long-press already does things

The mic already has long-press gestures, so "hold" is overloaded and needs a clear resolution when push-to-talk is on:

• Keyboard mic, idle: long-press → pick an audio/video file to transcribe (Feature Request: Floating overlay/bubble to trigger dictation (inspired by Wispr Flow) #88).
• Keyboard mic, recording: long-press → send with the on-device model (opt-in, Feature request: Long press record button to toggle local mode. #228).
• Floating button: long-press → prompt menu (rewording presets + Live Prompt, [Feature Request] Freeform AI Voice Command execution mode via Floating Button #230).

Options to resolve the conflict (pick one in implementation):

1. When push-to-talk is on, hold = record, and the above long-press shortcuts are disabled or moved to another gesture (e.g. double-tap) while the mode is active.
2. Keep push-to-talk as a separate gesture/affordance and leave tap-toggle + long-press as they are.

Option 1 is probably the cleaner mental model (one primary gesture), but it removes the existing hold-shortcuts in that mode — worth deciding deliberately.

Scope / notes

• Reuse the existing recording pipeline (onMicClick / startRecording / stopAndTranscribe); this is purely an input-gesture change.
• Not for realtime/streaming or long-form segmented, where "hold" doesn't map cleanly — those keep their current behavior (mirrors how the send-with-local shortcut is gated).
• Applies to the classic/legacy dictation layout too.