Agent eval harness

Builds a reproducible evaluation suite for an LLM or agent feature with a labelled dataset, graders, regression gates, and variance reporting. Use when model-backed behaviour needs a pass or fail signal that survives a prompt change.