LLM feature tests

Most defects in a model-backed feature live in ordinary code around the model call: the prompt it builds, the parser, the tool dispatcher, the retry policy, the output guard. Test that code deterministically with a mocked model, put a schema contract on every model output, and judge the model's behavior with an evaluation suite compared per case against a baseline. Success is a suite where a prompt or model change that breaks five cases fails the gate, even when five other cases improve and the aggregate score does not move.

Process

1. Split the feature at the model call. List the deterministic pieces and test each with a mocked model on every change, in seconds, for free: