Add voice dictation to mecatui

Problem

Long prompts are cumbersome to type in mecatui, and the TUI has no microphone-to-text workflow. Users have to leave the terminal, dictate elsewhere, and paste the result back into the composer.

Claude Code provides useful prior art: opt-in voice dictation with hold-to-record and tap-to-record modes, visible recording state, and editable transcription inserted into the prompt.

Proposed experience

Add opt-in voice dictation to mecatui:

• Add /voice controls to enable, disable, and choose hold or tap mode.
• Support a configurable push-to-talk key. A single press must retain its normal typing behavior when dictation does not activate.
• Show clear textual states such as warmup, listening, transcribing, and failure.
• Insert finalized transcription at the current cursor position without losing existing typed text.
• Keep the transcript editable before submission; do not auto-submit by default.
• Gracefully explain unavailable microphones, denied permissions, unsupported/headless or remote environments, missing recorder dependencies, and transcription failures.
• Persist the enabled mode and key binding in user settings.

Privacy and architecture

Voice mode must clearly disclose where audio is transcribed and require explicit opt-in before audio leaves the machine. Prefer reusing a daemon-side transcription capability so TUI and WebUI behavior can converge rather than coupling mecatui directly to one hosted speech service.

This overlaps with the backend transcription gap tracked in #725, but this issue covers the mecatui interaction, terminal audio capture, settings, and error states.

Acceptance criteria

• A user can enable and disable dictation from mecatui.
• Hold and tap recording modes work without preventing ordinary space entry.
• Recording/transcription state is visible in text, not conveyed by animation or color alone.
• Transcribed text is inserted at the cursor and can be edited before sending.
• Existing composer content is preserved across recording, success, cancellation, and errors.
• The push-to-talk binding and mode are configurable and persist across sessions.
• Unsupported environments and permission/device/service failures produce actionable errors.
• The UI discloses whether transcription is local or remote before first use.
• Tests cover key handling, composer insertion, cancellation, and failure states.
• Help and configuration documentation describe setup, privacy behavior, shortcuts, and platform limitations.

Prior art

• Claude Code voice dictation: https://code.claude.com/docs/en/voice-dictation
• Related Mecatl transcription service gap: #725

Start by examining the existing mecatui interaction and settings paths, then read the related transcription gap in #725. Map how key handling, composer editing, terminal audio capture, and daemon-side transcription can meet the acceptance criteria. Done means opt-in recording works in hold and tap modes, preserves editable text, persists settings, reports failures clearly, and includes tests and documentation.