The reason for stating iteration sweep was that I actually managed to get good constellations using random initial phases, although I don’t consider this a solution.
According to the bits per symbol parameter in modulation scheme and the bit rate, the symbol rate should be changed (my experience says as long as you haven’t used symbol rate in script mode, it has no effect on simulation, but you’d better do it anyway).
Since the constellation points are closer 16-QAM, decision region is smaller and therefore more power is needed, but none of this have to do with your problem (symbol rotation in the way you are facing) in my opinion.
Regards