I am a bit puzzled as what you describe sounds exactly as simple sampling without replacement which was already suggested in the original thread (Constraining Output of Multinomial Distribution - #3 by LucC). Could you explain once again why that would not work for you?
Best of luck with your model