Evaluation harness with LangGraph and CrewAI adapters (8 stubs remaining),
4 user stories (basic→advanced), structlog + Langfuse observability,
baseline isolation, metrics collection, and CLI scaffold.