Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders