You absolutely can predict someone's phenotype from her genotype, roughly speaking. In cases of recent admixture, the exact gene expression can be uncertain. Just look at Mendel and his peas! Even this variable heredity has limits. There was a zero percent chance that you'd end up looking like a typical Japanese person.
I don't understand your point. I'm not claiming that the classic races are fixed for all time. I'm not suggesting that we can't arrange genes in new combinations. I am claiming that people today cluster in certain historically-contingent ways and that these clusters reflect the everyday understanding of race.