Changes
PR #2299 by ksylvan: fix: protect provider error details and bound response memory usage
- Redacted private provider error details in responses unless debug mode is enabled, preventing sensitive information leaks.
- Capped non-streaming model responses at 50 MiB and stopped retaining streaming response content in server memory to bound resource usage.
- Replaced the Ollama SDK with direct HTTP requests and removed the unused Ollama dependencies from the Go and Nix manifests.
- Added auditing for storage and configuration changes without logging sensitive request bodies.
- Documented authentication throttling, access scope, limits, and audit events, and added tests covering redaction, auditing, response limits, and compatibility.