Runs trace-grounded error analysis on a Databricks AI agent: samples MLflow traces, open-codes failures, axial-codes them into a small taxonomy, and writes labels to a UC Delta table the next skills consume. Use when the surface is known but failure modes are not yet named, after a model or prompt change, or before judges or release gates are written. Do NOT use when failure modes are already labeled, the surface is unknown, or no traces and no synthesis path exist.
