nx_f32_matmul.nx

nx_f32_matmul.nx -- bits-up f32 dot product + GEMM kernel. L7 composition brick. Last load-bearing math primitive for an end-to-end f32-mode transformer block. Composes nx_f32_mul + nx_f32_add over rows / columns. API: nx_f32_dot(a, b, n) -- scalar dot product (sum of a[i]*b[i]) nx_f32_matmul(a, b, c, m, k, n) -- C[m,n] = A[m,k] @ B[k,n] row-major storage Bit-exact when all operands and intermediate sums are representable f32 (e.g., small ints). Otherwise round-to- nearest-even accumulates per the L4 f32 ops. No libm. Inner MAC uses nx_f32_hw SSE intrinsics (__f32_add/__f32_mul -> real addss/mulss on the FPU), which are BIT-EXACT vs the bits-up software f32 (proven by nx_f32_hw_matmul_gate's hw==sw differential) but replace dozens of integer ops per multiply with one instruction -- the hardware rung the LLM forward stands on.