guess I'm not satisfied with these answers.
The creases are visible in low roughness areas so adding roughness isn't a solution, by definition. They seem to be an artifact of the noise variation engine, however that works. My guess is that its taking two low-frequency (unrelated) components and (accidently) creating higher-frequency components by attempting to morph between them. I would suggest addition of a higher-order derivative matching (thus morphing smoothly wrt surface curvature as well.)
btw, why isn't voronoi one of the main noise options, more than just a function node? Voronoi is so useful for anything with a more cellular structure.
oh, and I could really use a function node called 'get curvature', prolly have a 'patch size' parameter. A complete result would have to have two parameters to account for a 2D surface, ie a saddleback, or maybe just one result that measures along the fall line.