So with the OPs script, ... that removes 18 of the 90 values,

And do you think that is his intent, or just clumsy coding? Has he said?

Additionally, the remaining digits aren't equally distributed: 2 appears about half as often as 3 and about a third as often as 9.

Again, is that a carefully designed intent, or just a badly coded demo thrown together for posting? Is true regardless of whether he uses the same PRNG as you tested?

Didn't he say preserving order was a goal, even if not a requirement? With does he mean the order before or after that sort in the demo code?

Even if the reduction to 72 values, is intended, you still get ( 1 - 6.16992500144231 / 8 ) *100 = 22.875937481971125% best possible saving.

If you take his description as the 'spec', as I did, -- rather then the output of a few trials of his code on a given machine/PRNG as you have -- the answer to his title question: "Data compression by 50% + : is it possible?", is: no.