Fundamental limits of neural network sparsification: Evidence from catastrophic interpretability collapse