feat(voice): audible recipient ack on inbound /handoff in #general

[itskai-dev]

Summary

When an agent posts /handoff to=<name> in #general, the named
recipient emits a recipient-authored 1–4 word audible acknowledgment
through the existing /canvas/speak path (Kokoro/ElevenLabs →
voice_output SSE → activeSpeaker rail). Closes the joyless seam where
inbound handoffs were silent on canvas.

Locked behavior (per kai msg-1777921934933):

• real `/handoff to=` only (strict prefix; cloud SDK typo set not honored on this path)
• voiceCapable recipient only (`agent_config.settings.voice` non-empty)
• human present in room (≥1)
• recipient not already mid-TTS (gate on activeSpeaker)
• 1–4 word recipient-authored ack via Anthropic `claude-haiku-4-5`
• silence on any gate-fail or LLM failure (no canned chrome)
• no chat duplication, no extra canvas UI
• turn-local dedup (the inbound handoff turn gets ack-only audio; later turns are unaffected because voice-on-chat keys on message content)

Files

• `src/voice-ack-on-handoff.ts` (new): parse + decide + speak module, wired from `chatManager.sendMessage` right after `triggerVoiceOnChat`
• `src/active-speakers.ts` (new): `Set`-backed mirror of who is mid-TTS so server-side gates can ask the question without reaching into the `canvasStateMap` closure
• `src/canvas-interactive.ts` /canvas/speak: marks speaker active on queue, clears on `voice_output` (durationMs+500) or safety net (estimatedMs+8s). Covers chat-driven TTS lifecycle.
• `src/server.ts` `setActiveSpeaker` / `setActiveSpeakerAudio`: also write to active-speakers store so interactive voice sessions feed the same gate.
• `src/voice-ack-on-handoff.test.ts` (new): 24 unit tests — parse strictness, gate ordering (channel → sender → parse → self → voiceCapable → humans → activeSpeaker), clamp behavior, and wiring (fast-path skip, gate skip, happy path, silence on empty ackLine, silence when recipient mid-TTS).

Proof bar (canonical staging)

After merge + image roll on `rn-34faba44-wlgkeq`:

1. Happy path — kai posts `/handoff to=link` from chat with a human in the room → link speaks 1–4 word ack
2. No human — same handoff with no human in room → silence
3. Non-voiceCapable recipient — handoff to an agent without `settings.voice` → silence
4. Recipient mid-TTS — link is currently speaking → silence
5. Free-text @mention — `@link please look` (not a real handoff) → silence

Test plan

• `node --import tsx --test src/voice-ack-on-handoff.test.ts` — 24/24 pass locally
• `tsc --noEmit` clean
• CI green
• Roll canonical staging image
• 5-bar proof on canonical