Hybrid fleets: local models and frontier APIs under one orchestrator

The studio's local pipeline was updated to the new open-weight generation — Qwen 3.5, Gemma 4, DeepSeek V3.2 — served via vLLM with speculative decoding. On repetitive workloads, local models now reach practical parity with the frontier APIs of a year ago, and speculative decoding cuts latency by a further ~40%. The studio's router treats local and cloud as a single fleet: the boundary moves per task, not on principle.