Add ST_M stage: t_hat[i] = e_hat[i] + sum_j A_hat[i][j] o s_hat[j] via poly_mul_sync + inline mod-add accumulation. Per (i,j): stream 256 (A,shat) pairs into poly_mul, then accumulate 256 products into T_i (seeded from E_i when j==0, else running T_i). m_pending waits for poly_mul IDLE between terms. Verified vs ml-kem-r golden: 512/512 t_hat coeffs exact (19885 cyc).