nx_matmul_tile_kat.nx

nx_matmul_tile_kat.nx -- proves the m>1 COLUMN-TILED matmul is BIT-EXACT against the existing kernel, and measures what hoisting the column actually buys. WHY: nx_batchscale_kat measured the matmul path FLAT across m (us_per_row 4705/4056/4295/4191 at m=1/2/4/8, speedup 1.12x) -- the signature of each row re-streaming the weight column. mmt_range walks OUTPUT CELLS, so column B[*,j] is read once PER ROW. mmt_tile_cols hoists that load: read the column once, accumulate against all m rows. THE EQUALITY CLAIM IS NOT A TOLERANCE CLAIM. For any (i,j) both kernels start at f32 zero and add over kk ASCENDING -- identical order, so identical rounding. Only the interleaving across i differs. Anything other than BIT-EQUALITY here means the tiled kernel changed the arithmetic, and that is a REJECT, not a rounding difference to be waved through. T4 is the load-bearing negative control: a DELIBERATELY WRONG tiled result must be caught by the same comparison, so a passing T1-T3 cannot be an artefact of comparing a buffer with itself. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0