This uses a least-squares fit to improve accuracy of the numerical
derivative in the presence of rounding errors.  Close-ups of even
simple solver targets reveals that serious noise is present.