TopK sparse autoencoders trained at matched sparsity and dictionary size on Llama-3.2-3B, Mistral-7B-v0.3 and Qwen2.5-3B. Reports abrupt dictionary collapse followed by partial recovery, dense-feature degeneracy severe enough to invalidate frequency-based feature selection, and a null cross-architecture matching result whose detection floor is measured directly by planted-signal power analysis at correlation 0.95 to 1.00. A positive control bounds the method: two SAEs differing only in training seed agree on just 8% of their matchable features, recovered at 108x the permutation null, so the near-zero cross-model result is not instrument failure and 8% is the ceiling. The three seed runs are indistinguishable by loss, FVE, L0 and dead-feature count while disagreeing about 92% of their dictionaries.