Yes, you're right. Restricting the GA to the match sets (in effect, the niches) also eliminate the take-over problem.

One does have to be careful about the rate of invocation of the GA in the different match sets, and deletion, as discussed later in the paper. You have to do both in such a way that the niche sizes stay more or less in balance.

Won't this niching solution also address the first problem (preventing strong-niche classifiers from taking over the population), in addition to the last problem?