fix(embedding): deterministic CLAP on full-length audio (random-truncation bug)

feat(embedding): expose CLAP similarity vector on the API as a stream-safe subprocess island

docs+copy: genre/mood/instrument are measured now, not LLM-inferred

docs(api): flag LLM-path latency — use submit() not predict()

feat(api): llm flag on audio_to_prompt — Pollinations-polished prompt

feat(api): headless audio_to_prompt endpoint — audio in, SA3 prompt out

docs: analysis cheat sheet — what measures what, what's LLM-inferred

refactor(transform): URL-pass pattern — bridge fetches source, not browser

feat(local-bridge): POST /transform — SA3 audio-to-audio via init-audio

Bridge: widen Access-Control-Allow-Headers to 'Content-Type, content-type'

Local bridge: handle Chrome PNA preflight; drop bad mixed-content guard

Local SA3 bridge: wire run_mlx_sa3 to ~/sa3_mlx/sa3 with verified CLI

Initial deploy