Right. If you make them super-racist, you get a chaotic state because the polygons move around constantly. I don't see it as a problem with the model. I think that it makes sense, because people who are extremely racist are (in addition to being morally in the wrong) impractical and chronically pissed off. However, you can get a very-segregated state by gradually turning up the racism (from 33% to 50%, then to 60%, then 70%). If you do it too fast (i.e. change the racist temperature too quickly) you get the chaotic state you're talking about.
In the real world, this chaotic effect might be a negative (i.e. violent) interaction. Or, more hopefully, people might recognize such extreme bias as untenable and gradually find it undesirable to live in monoculture.
By the way, I was able to get 99% (ETA: 100%) segregation (because engineering racist dystopias is the point of this exercise? :) ) by altering the racism level to move around the patches and "attack" small enclaves of blue in the yellow area. If you think of it as abstractly playing with optimization parameters with a hand-controllable Lagrange multiplier, it feels less evil.