[agency] Harden CLJS vocal agency as the active lip-sync timeline runtime

[meekmachine]

Goal

Finish Vocal as the active ClojureScript sentence-level lip-sync runtime for text and provider-timed speech timelines.

Current state

Polyester already has src-cljs/latticework/vocal.cljs, in-process and worker client APIs, text-to-viseme planning, sentence timelines, word-boundary drift detection, snippet building, jaw curves, coarticulation constraints, and cleanup planning. The TypeScript VocalService is currently the active lip-sync runtime used by TTS and LiveKit paths.

This mirrors the Latticework research issue: meekmachine/Latticework#34 and the stabilization discussion: meekmachine/LoomLarge#521.

Clojure-way direction

• Keep sentence-level timeline planning in pure CLJS data: viseme events, word timings, snippet curves, active sentence state, drift decisions.
• Keep all animation side effects at the host edge: schedule/remove/seek/pause/resume.
• Provider timelines should be pre-timed; snippet playback rate should remain neutral to avoid double-applying speech rate.
• Word boundaries should verify and correct drift by emitting seek requests instead of rebuilding clips on every word.

Acceptance criteria

• createVocalAgency and worker client can replace the TypeScript VocalService for text and provider timelines behind the existing JS API.
• Worker returns deterministic serializable snippets for text and provider viseme timelines.
• Word-boundary drift still emits seek requests when threshold is exceeded.
• Active sentence cleanup cannot remove unrelated snippets and does not leak timers.
• Add TS-vs-CLJS parity tests for representative phrases, jaw curves, coarticulation caps, sentence cleanup, pause/resume, and drift correction.
• Document that Vocal is the visible mouth runtime; legacy LipSync remains compatibility/research until folded into this path.