There is no "classical implementation" of Mitchell-Netravali filters; the original paper describes the filter functions theoretically in the ]-2,+2[ real range.
The smoothness parameter is only used if you select the Interpolation_Auto interpolation mode for downsampling and the internal routines decide to use the Mitchell-Netravali cubic filter, based on the amount of downsampling on both axes. The smoothness parameter does not change the shape of the filter. It is just a magnification factor to define the dimensions of the rectangular interpolation neighborhood where the separable filter is evaluated.
If you select one of these filters directly (Interpolation_CatmullRomSplineFilter, Interpolation_CubicBSplineFilter or Interpolation_MitchellNetravaliFilter) the smoothness parameter will be ignored. Instead, the Image.interpolationXRadius and Image.interpolationYRadius properties will define the dimensions in pixels of the interpolation neighborhood, which will be:
interpolation width = 1 + 2*Image.interpolationXRadius
interpolation height = 1 + 2*Image.interpolationYRadius
The default radii are 2 pixels, which means that interpolation will take place on a 5x5 neighborhood by default. Note that these filters may require larger interpolation dimensions to be properly evaluated, depending on the application.