[codex] Trigger prosodic nods from vocal speech timing

[meekmachine]

Summary

Refs meekmachine/LoomLarge#697.

This keeps speech-time head nods owned by Polyester's prosodic expression agency. The speech/vocal agency now triggers prosodic start, word-boundary pulses, and stop; the prosodic agency remains responsible for loading, scheduling, fading, and playing the head-nod snippets.

What changed

• VocalService accepts a prosodicService dependency and calls:
  • startTalking() when a sentence/timeline is scheduled
  • pulse(wordIndex) on word boundaries
  • stopTalking() when the speech snippet stops or cleans up
• TTSService passes its optional prosodicService config through to VocalService, so Web Speech, SAPI, Azure, and external word-boundary paths can use the same agency bridge.
• Prosodic snippet normalization preserves inherit flags instead of dropping them.
• Bundled headNodSmall and headNodBig speaking snippets now use inherited AU53/AU54 anchors.
• AGENTS.md now documents that speech-time prosody belongs in Polyester, not LoomLarge call sites.

Why

The previous LoomLarge-side nod scheduling was the wrong boundary. LoomLarge should configure/pass agencies. Polyester owns speech timing, prosodic pulse handling, and animation scheduling. This PR fixes the inherited-keyframe issue inside the owning agency path and wires speech timing to that agency.

Validation

• npm exec vitest run src/vocal/__tests__/service.test.ts src/animation/__tests__/snippetPreloader.test.ts src/prosodic/__tests__/prosodicService.test.ts
• npm run typecheck

[meekmachine]

Closing this draft for now. The current goal is to avoid changing Polyester and first verify whether the prosodic/head-nod animation data should be updated in stored animation data rather than package code.