Alex Clark brought up a point about how to understand the implications of Chomsky's cute example. I was resistant (to say the least) to his point. However, he and Thomas Graf have convinced me that there is an interesting problem afoot. I confess as much in the comment section (here), but thought that honesty (something I generally admire more in the breech) demanded that I say so prominently. So, Alex, thx for stimulating a useful discussion.
Let me add one more point: the problem Alex points to is that in Stablerian formalisms of MG lots of info gets packed into the features associated with a lexical item. What Thomas Graf points out is that if we are not careful in what features we allow here, we can bypass the locality restrictions that we have taken to be part of UG (e.g. Islands, Minimality). In some sense we have known this for a while (certainly in the case of minimality for once "relativized" all that is required to finesse it is a novel feature and see Berwick and Weinberg's discussion of a similar problem in chapter 3/4 for analogous concerns), as we have all known that profligate use of features (something not unknown in minimalist analyses, IMO) can allow one to derive pretty much anything. However, the problem of what constitutes "allowable" features and how we ought to be coding dependencies in terms of features (and subcat and criterial features are part of virtually every system) is even more critical than I had considered. The Stabler formalization of MG uses such features extensively precisely because our minimalist theories rely on notions like sub cat, theta marking, wh-agreement etc extensively. A terrific theoretical question is how to circumscribe these features so as toe exclude the "crazy" ones, ones that allow one to simply bypass the UG restrictions. Thx to Alex and Thomas for making this clear, at least to me.