As explained in logs 7. v4 and 8. 3:2 compression here is the current result:

So it's a bit simplified, with a newcomer: the boolean mixer in front of the write ports.

v4.3 should provide a much more decent entropy source for the 14 bits consumed during each cycle, such as LFSR32. Maybe consume one word from D alternately ? Or enlarge D to 32 bits ?

Source code: v4.tgz (not fully tested though)