This seems to be an appeal to generative power, in a loose sense. Throughout the history of generative grammar, there has been a concern with / interest in the class of "languages" that can be generated using a given theory. The underlying premise is that the best theory is one that allows the description of all attested languages, and nothing else (often justified in terms of Popperian falsifiability and the quasi-Popperian claim that a theory should not predict the existence of a thing not known to exist).
What is not expressly stated is the alternative theory: that non-head material can be anything. The theory in question says that all non-head material must be phrasal, therefore non-head material cannot be non-phrasal (obviously). The alternative theory allows both outcomes. The implication is that therefore the class of languages that can be generated by the theory in question is a proper subset of the class of languages that can be generated by the (unnamed, implicit) alternative theory. Given the aforementioned premise about "exactly the attested languages", the proposed theory is "mathematically nicer".
There are a few problems with this line of reasoning, investigation of which would take us too far afield. Briefly, though, (1) the "smallest set of languages" premise is quite questionable as a logical principle, (2) it remains to be proven that there is any actual difference in the set of languages that the two theories allow (the proposed theory probably can describe any imaginable language, it just does it a different way). A competing meta-principle is that the best theory is the one that posits the fewest theoretical entities (classical Occam's Razor), in which case adding a stipulation "must be phrasal" complicates the theory, with no concomittant benefit.