It looks like PTC could do a better job on matching at least the first derivative.
It isn't clear whether follow-on geometry is adversely affected to a noticable extent. It's possible that the curvature analysis sampling is based on the generated mesh / line segment approximation of the curve . A follow-on feature should use the underlying feature definition. I would hope the internal definition would be driven by a piecewise cubic spline or other curve that enforces first and second derivative and passing though the points.