[agency] Create hybrid CLJS transcription analyzer

[meekmachine]

Goal

Create the Polyester transcription agency as a hybrid ClojureScript analyzer/reducer plus host adapter for browser speech and microphone effects.

Current state

Polyester root still inherits the TypeScript transcription service from Latticework. There is no src-cljs/latticework/transcription.cljs agency yet. The current TypeScript service owns SpeechRecognition, mic permission, AudioContext analyzers, reference-track echo filtering, restart timers, and interruption state.

This mirrors the Latticework research issue: meekmachine/Latticework#38.

Clojure-way direction

• Main thread owns SpeechRecognition, getUserMedia, AudioContext, MediaStreamTrack, analyzer nodes, and sampling loops.
• CLJS owns transcript normalization, interim/final transcript state, echo-filter policy, interruption threshold/latch state, and serializable transcript/interruption decisions.
• Host sends plain audio-level samples, reference-level samples, and speech-recognition events into the planner.
• Planner emits data events: transcript interim/final, interruption detected/cleared, restart/stop recommendations, and state snapshots.

Acceptance criteria

• Existing transcription public API remains source-compatible through an adapter.
• Current interruption behavior remains equivalent: audio interruption stops TTS, final interrupted transcript processes once, transcript-only fallback stays disabled by default unless configured.
• No CLJS worker code references SpeechRecognition, navigator, AudioContext, MediaStream, DOM, or requestAnimationFrame.
• Add tests for transcript normalization, echo-filter thresholds, interruption latch/release, final transcript handling, restart policy, and cleanup.
• Conversation agency can consume transcription events as plain data, independent of browser API objects.