nx_f32_q4k_matmul_test.nx

nx_f32_q4k_matmul_test.nx -- smoke for nx_f32_q4k_matmul.nx. REWRITTEN 2026-07-07 to the organ's CURRENT axis semantics: B is n WEIGHT ROWS, each row = k quantized values ((k/256) super-blocks, 144 B each), C[i,j] = dot(A[i,:], dequant(row j)). The previous version of this test encoded the OLD transposed [in,out] layout (k=1, one block along n) -- after the k-axis fix landed in the organ, that shape trips ERR_ALIGN (k=1 is not 256-aligned) and the test had been failing exit 23 UNNOTICED (found 2026-07-07 during the threading arc's consumer sweep; serving-ne-working class). Builds a synthetic Q4_K super-block whose dequantized values are 2.0 at position 0, 3.0 at position 32, 4.0 at position 128, and 0 everywhere else (ggml 32-byte-group nibble layout, see nx_q4k_to_f32_test.nx). Tests (all operands small exact integers -> f32 math is bit-exact): 1. m=1, k=256, n=1: A = 1.0 at [0],[32],[128] -> C[0] = 2+3+4 = 9.0; then A[0]=2.0 -> C[0] = 11.0 (input scaling). 2. n=2 identical rows -> C[0] = C[1] = 9.0 (row offset stride 144). 3. m=2 (second A row doubled) -> row0 gets 9.0s, row1 gets 18.0s. 4. k not 256-aligned -> NX_FQ4M_ERR_ALIGN. 5. MT cross-check: nx_f32_q4k_matmul_mt (2 workers) bit-equals the serial result on the n=2 shape (full adversarial coverage lives in nx_q4k_matmul_mt_gate).