nx_sparse_tensor.nx

nx_sparse_tensor.nx -- compressed sparse row tensor + sparse matmul. FIRST algo-led-not-brute-force kernel per the min-hardware-floor cardinal. Post-softmax attention is mostly zero; LLM weights after pruning are mostly zero; world-model frame deltas are mostly zero. Skipping those zeros is 10-100x speedup on the substrate AND 10-100x memory reduction -- both essential for the "$50 SBC beats Genie 3 in coherence" trophy. CSR (Compressed Sparse Row) format: row_ptr[m+1] -- row_ptr[i+1] - row_ptr[i] = nnz in row i col_idx[nnz] -- column index of each non-zero values[nnz] -- the non-zero values themselves For an M x N dense matrix with NNZ non-zeros: dense storage = M * N * 8 bytes CSR storage = (M+1) * 8 + NNZ * 16 bytes crossover = NNZ < M * N / 2 (sparsity > 50%) typical attention = NNZ ~ M * N / 100 (>99% sparse) Operations shipped: dense_to_csr -- materialise a dense tensor as CSR csr_to_dense -- back-convert (audit + numeric_oracle gate) csr_matmul_dense -- C[m, n] = sum over non-zero entries of row m times dense column n. Skips ZERO ROWS entirely. O(NNZ * N) instead of O(M * K * N). Future variants (gated on oracle): AlphaTensor 2022 / AlphaEvolve 2025 sparse-matmul algorithms register here as alternative dispatch entries. genealogy_id: tinkham_1969_sparse_matrices + bell_garland_2009_csr + pissanetzky_1984 + alphatensor_2022 lineage_id: substrate_sparse_tensor_v1