Every model call now has an output cap

Chemical classification, safety data sheet reading, chat and feedback no longer fail with a credits error when the key's limit runs low. Each cap is set from what that call really produces, so the longest safety data sheets still come back whole. A test now catches any new call without a cap.