v0.6.5

Bug fixes

• Image-based reranking now transports Python SDK image inputs correctly and includes document images in the Qwen3-VL reranker's prompt.
• Structured generation correctly resolves JSON Schema references without discarding constraints beside a reference.
• SGLang model loading no longer blocks the server event loop, allowing health checks and other requests to remain responsive during startup.
• Configuration updates correctly account for pool ownership, replace stale snapshots, and stop advertising removed model configurations as ready.

Breaking changes

• Provisioning responses now use HTTP 503 with retry information instead of HTTP 202. The Python and TypeScript SDKs understand the updated response; direct API clients should handle the new status when waiting for capacity.