Eval dataset

The cases an agent is tested against. Each row is two lines, what the agent is asked and what a good answer says, and any text is edited where it stands. Open a row for its context, tags, name and why it failed last time. Type in the pinned last row to add a case, accept a real run's answer in one press, filter by tag or by what failed, import JSONL or CSV by button, drop or paste, and run exactly what is in view. Ships helpers that turn AI SDK messages and Review Queue rows into cases and build the state for experimental_evaluate.