Reference Document: https://eprint.iacr.org/2016/027.pdf
According to appendix D section 4 definition 27
each vertex in V has predecessors sampled independently and uniformly at random from the set U
It appears from the pseudo-code in section 3.1 step 2b, and the reference code given on GitHub hash_state_mix(), that this has been implemented by the authors as some form of squeezing bits out of the hash function; however my question is could it just as easily have been done with Mersenne Twister or is there another reason why the hash is required?