Gateway: OpenAI-shaped text-to-speech endpoint (phase 1: remote passthrough)

[wpak-ai]

The gateway routes chat, embeddings, rerank, and transcription models, but
has no speech synthesis kind. Clients wanting TTS through the gateway's
auth, queueing, and catalog must bypass it. Design:
design/report-gateway-tts-endpoint.md.

Scope (phase 1, remote passthrough only):

• Config: ModelKind::Speech plus a voices capability, content-validated
  at load (no empty entries, no duplicates, rejected on non-speech kinds);
  config UI gains the speech kind and a voices chips editor.
• Wire: SpeechRequest with validation (input cap 4096 chars, speed
  bounds, closed response_format enum defaulting to mp3 for wire
  fidelity against providers whose own default is wav).
• Upstream: Upstream::send_speech returning StreamedAudio, with a
  dedicated audio streaming client (no whole-request timeout).
• Routes: POST /v1/audio/speech (auth-before-extraction, voice checked
  before queue admission, permit held across the stream, mid-stream
  failure truncates) and GET /v1/audio/voices (id-first union across
  speech models).
• Errors: speech-only 429/503 envelopes; upstream bodies never leak.
• Verification: integration tests plus a credential-gated live parity
  script against Together AI, with the observed dialect recorded.

Acceptance: full CI gate suite green; live parity run passes; guide
chapter, README, and example config document the surface; as-built design
record committed.

Out of scope (phase 2): local Orpheus engine, SSE reframing, voice picker,
ElevenLabs/Baseten adapters.