Stream Speech (WebSocket)

init   { type, voice_id, output_format?, consistency?,
         expressiveness?, auto_flush?, auto_flush_char_threshold? }
speak  { type, text, flush? }
flush  { type }
close  { type }

ready  { type, session_id, output_format }
audio  { type, data (base64), is_final }
done   { type, session_id, characters_billed, cost_usd, latency_ms }
error  { type, code, message }

Buffer cap:    4000 chars (per accumulation before flush)
Speak rate:    500 msgs / 10s window
Idle timeout:  60s (no client messages → close)
Billing:       once at "done", partial disconnect not charged