Sparse Autoencoders Enable Scalable and Reliable Circuit Identification in Language Models | Synapse