Do Sparse Autoencoders Generalize? A Case Study of Answerability