OK. I've looked at cnumber and though correctly executed, it is too limited in my opinion. It only produces the single largest condition number (no variance decomposition) for a linear model. Proper diagnostics for collinearity may involve the others as well.
My other complaint is trivial: the included example excludes the intercept as a regressor. As BKW argue, and Hill and Adkins (2001) agree (shameless plug), the intercept should ordinarily be included if the model is estimated with one . When the intercept is included in cnumber the computations are correct. So the function is doing what it is supposed to do,
The proposed bkw.gfn function produces the whole set of condition numbers and the BKW variance decompositions --without centering-- for any estimated model (linear or nonlinear) for which gretl returns a variance covariance.