[agency] Create hybrid CLJS TTS planner with host-owned audio effects

[meekmachine]

Goal

Create the Polyester TTS agency as a hybrid ClojureScript planner plus host adapter: CLJS owns speech/timeline planning, while browser/backend/audio effects stay on the main thread or backend.

Current state

Polyester root still inherits the TypeScript TTS service from Latticework. There is no src-cljs/latticework/tts.cljs agency yet. The TypeScript src/tts/ttsService.ts currently owns Web Speech, SAPI/Azure backend calls, AudioContext playback, display-media playback reference, timeouts, and a VocalService instance.

This mirrors the Latticework research issue: meekmachine/Latticework#37.

Clojure-way direction

• CLJS owns text token parsing, emoji/emotion event extraction, provider response normalization, timeline planning, utterance state, cancellation tokens, and commands to Vocal/Prosodic/EmotionExpression.
• Main thread owns Web Speech, SpeechSynthesisUtterance, AudioContext, display-media capture, HTMLAudioElement, permission prompts, fetch calls, and playback start/end events.
• Backend owns Azure credentials and synthesis for the buffered Azure path. Polyester workers should consume provider timing data; they should not own secrets.
• Preserve current JS imports (createTTSService, TTSService) through an adapter during migration.

Azure/audio contract to support

• Buffered Azure: host calls LoomLarge POST /api/azure-tts/synthesize, receives audio_base64, visemes, word_boundaries, duration, decodes audio, starts playback, then sends a clock-anchored provider timeline into CLJS/Vocal.
• LiveKit drawer TTS: host calls POST /api/tts/synthesize, receives audio_base64 WAV and estimated word boundaries, plays audio, then sends text-predicted timeline plus word-boundary ticks.
• Web Speech: host listens for onboundary and start/end events, while CLJS plans fallback text visemes and prosodic/emotion events.

Acceptance criteria

• Existing TTS public API remains source-compatible.
• No CLJS worker code references Web Speech, navigator.mediaDevices, AudioContext, DOM, backend fetch APIs, LiveKit objects, or Loom3.
• TTS emits data commands for Vocal, Prosodic, and EmotionExpression instead of directly scheduling animation snippets.
• Stop/interruption reliably cancels pending timelines and active speech state.
• Current TTS tests continue to pass through the adapter.
• Add CLJS tests for token parsing, emoji timing, provider response normalization, utterance cancellation, and speech lifecycle state.