Even ignoring vector magnitudes, wouldn't cosine distance as a measure of similarity only make sense if you're working with a convex set? That seems like it's far from a guarantee working in a high-dimensional space.