Answers no longer stop abruptly at a fixed length. The model's full output limit is now respected, so long explanations and documents finish properly.