* Add Langfuse-based LLM observability * Document Langfuse configuration * Don't hardcode model in use