Guardrails can use typed judgments as their check engine

The llm_judge and moderation guardrails take an engine setting: utility_llm (the default, unchanged) or jev. With jev, every check on a stage is answered in one request, and the verdict is the configured threshold applied to a calibrated probability rather than a JSON verdict parsed back from text. A check can no longer pass because a response failed to parse.

Operators enable it with the UTILITY_TYPESAFE_API_KEY deployment credential. When it is unset, jev checks are skipped with a log line and the turn proceeds. An experimental jev capability with a jev_evaluate tool is available in development registries.

See Guardrails and TypeSafe. #3644