Ah yes, that is precisely the problem. Back then, I think the errornorm() function automatically did some degree_rise=3 or whatever. If I now interpolate the exact solution onto a CG3 or CG4, I do in fact get the convergence rate I would expect.