feat(media): Vertex AI (Google) speech, transcription, music & video generation

[artickc]

Summary

The dedicated media-generation endpoints only recognized third-party providers in their
registries:

• /v1/videos/generations → kie, haiper, leonardo, pollinations, minimax, together, replicate, comfyui, sdwebui, veoaifree-web, runwayml
• /v1/music/generations → kie, suno, udio, minimax, comfyui
• /v1/audio/speech & /v1/audio/transcriptions → openai, deepgram, groq, elevenlabs, …

Google/Vertex was absent from all of them. So any vertex/* media model (Veo, Lyria,
Gemini TTS / native-audio) was rejected with Invalid <kind> model: vertex/... Use format: provider/model — even though the format was already correct. The provider simply wasn't
supported for those media types, while it is fully supported through Vertex AI.

Changes

• open-sse/executors/vertexMedia.ts (new): a shared Vertex media client that reuses the
  Vertex chat executor's auth (parseSAFromApiKey / getAccessToken for Service Account JSON →
  OAuth bearer, or Express API key via ?key=) and implements the verified contracts:
  • Speech — {model}:generateContent with responseModalities:["AUDIO"] + speechConfig,
    returning PCM L16 which is wrapped into a WAV container.
  • Transcription — {model}:generateContent with inline audio + a transcribe instruction.
  • Music (Lyria) — {model}:predict → predictions[0].bytesBase64Encoded (WAV).
  • Video (Veo) — {model}:predictLongRunning → poll {model}:fetchPredictOperation until
    done → response.videos[0].bytesBase64Encoded (MP4).
• Registries: added a vertex provider to videoRegistry (Veo 3.0/2.0), musicRegistry
  (Lyria 2), and audioRegistry speech (Gemini 2.5 TTS) + transcription (Gemini 2.5/2.0).
• Handlers: added vertex-veo, vertex-lyria, vertex-gemini-tts, and vertex-gemini
  format branches to the respective handlers, returning the standard OpenAI-like shapes
  ({created, data:[{b64_json, format}]} for video/music; audio stream for speech; {text}
  for transcription).

Because all four reuse getProviderCredentials("vertex"), an existing Vertex connection (SA
JSON or Express key) now serves media generation with no extra configuration.

Verification

All four contracts were validated live against Vertex AI with a real Service Account:

Endpoint	Result
Speech (gemini-2.5-flash-preview-tts)	PCM L16 24kHz → WAV ✓
Transcription (gemini-2.5-flash)	round-tripped "the quick brown fox…" ✓
Music (lyria-002:predict)	WAV base64 ✓
Video (veo-3.0-fast-generate-001:predictLongRunning)	operation → poll → MP4 ✓

Unit tests in tests/unit/vertex-media.test.ts (7 tests, all pass) cover WAV wrapping, the
SA-project vs Express ?key= URL building, request payload shapes for each modality, and the
Veo submit→poll→result flow (fetch mocked). The existing media handler/registry suites
(registry-utils, video-generation-handler, music-generation-handler,
audio-speech-handler, new-content-providers, registry-direct-exports) still pass (104/104).

Notes

• Express-mode keys use the project-less global publisher endpoint (?key=) — best-effort;
  the Service Account JSON path is fully verified.
• veo-3.1-fast-generate-preview is not GA in all projects; veo-3.0-* are used as the
  default registry entries.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Thanks, @artickc! 🙏 Solid feature — Vertex AI media across all 4 endpoints (Gemini 2.5 TTS → WAV, transcription, Lyria 2 music, Veo 3.x video via predictLongRunning poll), reusing the existing Vertex SA/Express auth. The shared vertexMedia.ts executor (under the 800 cap) + per-handler vertex/* branches follow the existing executor/registry/handler pattern cleanly, and the 7 unit tests cover WAV RIFF wrapping, SA-vs-Express URL building, and the Veo poll loop. Green locally (7/7). Small baseline bump for the two handler branches. Merging into release/v3.8.26. 🚀