There is a bug in the block-diagonal schur matrix computation in line 473. It should be tmodal = tmodal @ rperm.T. Otherwise tmodal @ amodal @ np.linalg.inv(tmodal) $\neq$ a where a is the input to the function.