forum.alglib.net

In a generic random forest model, if the number of available features used ("mtry") is small, then it is likely that few terminal nodes will be constructed for which the class membership of objects at the daughter nodes is pure. In fact, it is more likely that the entire tree may involve only one or two splits of objects, only because there are no more than 1 or 2 features available.

Has this been discussed in a papers (books) in terms of importance scores -- and the overall issues that the trees are not really trees but rather one-step splits?

I think that it is in the very spirit of random decision forest - to build a lot of biased "simple" and random classifiers using subset of full sample. So although individual trees can be weak classifiers, forest as whole is stronger classifier.