Hey folks, hitting multiple blockers building a production voice agent on Vertex AI Live API and would love any workaround or guidance from the team. Setup: Twilio Media Streams → Cloud Run (europe-west1) → Vertex AI Live API. Python google-genai SDK. The problems compound: 1) gemini-live-2.5-flash-native-audio in europe-west1 — first turn works perfectly. After the first turn_complete, model goes silent. Audio keeps reaching Gemini (verified RMS up to 2143), no input_transcription, no model_turn, no error. Tried auto VAD HIGH sensitivity, manual VAD with activity_start/end, no kickstart, continuous audio + markers. Same outcome every time. 2) gemini-live-2.5-flash cascade — REST publishers endpoint says it exists in europe-west1 and us-central1 (HTTP 200). Live API WebSocket connect returns 1008 policy violation “Publisher Model not found” in every region tried. 3) gemini-3.1-flash-live-preview — not reachable via Live API in europe-west1, us-central1, or global. 4) gemini-2.0-flash-live-preview-04-09 (the model name in the official Vertex AI Live quickstart) — 404 in all regions. Net: only working option in EU Vertex AI Live is native-audio with Bug 1. Multi-turn agents are non-functional today. Two questions: a) Is there ANY model that streams reliably multi-turn via Vertex AI Live API right now? In any region? b) Should we migrate to AI Studio direct (genai) for cascade access? If yes, the Vertex AI docs should say so. Production agent for a Spanish SEO agency. Want to stay on Google. Happy to share full repro + logs.