#220: fixed another SVD NaN problem, this time from rounding errors in calculation Cproto_diag See merge request !315