This does a great job of demonstrating some of the curious effects of premature rounding. The non-transitivity is only possible because after generating a pair of random numbers, the outcome is rounded to one of two values ("Win for Player A" or "Win for Player B"). If there was a notion of "Player A won by two", then the dice could be ranked by expected value.

This is one reason why (for example) counting the number of studies that confirm vs. reject a hypothesis is not going to give the best indication of its validity. It's also why we don't track changes in market indices by counting how many items increased against how many decreased.