feat: native voice replies and first-party TTS

[furukama]

Summary

HybridClaw has inbound audio transcription and can send generated audio files back through some channels, but it does not ship a first-party TTS runtime config or built-in speech synthesis provider.

Why

Voice is already partially present in the product surface. A native TTS path would make voice interactions feel complete instead of requiring custom scripts or external MCP wrappers.

Proposed scope

• Add first-party tts.* runtime configuration.
• Add at least one built-in speech synthesis provider abstraction with pluggable backends.
• Support generating outbound audio replies directly from agent text.
• Add per-channel delivery rules for supported platforms.
• Allow voice preferences per agent/session where practical.

Candidate UX

• /tts on|off
• /tts voice <name>
• gateway config for provider, voice, format, and max duration
• optional "reply in voice" channel/session setting

Implementation notes

• Start with one reliable provider/backend and a clean abstraction for future backends.
• Reuse the existing media delivery path instead of inventing a separate outbound transport.
• Ensure generated files are treated as sensitive transient artifacts and cleaned up correctly.
• Consider Discord file delivery and WhatsApp voice-note/PTT support separately so the first version can ship incrementally.

Acceptance criteria

• A user can enable TTS and receive spoken replies without custom tooling.
• At least one built-in provider/backend is documented and tested.
• Generated audio is delivered through supported channels using the existing media pipeline.
• Config and transcripts make it clear when a text reply was synthesized to audio.
• Tests cover provider selection, generation failure paths, and channel delivery behavior.