One of the points about my claim that HPSG does a better job with this data that I should have spelled out more but didn’t, though, was exactly the point you raise. The advantage to HPSG here is that you DON’T need any ad hoc mechanisms to explain this seemingly incongruous data. In the standard theory - where all syntactic rules are universal - you would. But in HPSG, where all mechanisms are based on feature unification, you don’t. The “language universals” (UG) are just the fact that words are represented as bundles of features and, for UG purists, also, one presumes, the inventory of features. All languages have a NEG feature on all their verbs, and all languages have a rule that requires a [NEG +] input. Since NEG is not specified on the lexical item, then applying the rule (which in HPSG involves unifying the rule item with the argument lexical items) renders the lexical item [NEG +]. It started out as [NEG boolean] - i.e. unspecified - but when it unifies with the negation rule it becomes [NEG +]. Polish, by this story, just so happens to be one of those rare languages that has some items that are prespecified for [NEG -] in the lexicon. These words obviously can’t unify with the rule, and so can never be made negative. Polish also has some words that are [NEG +] in the lexicon, and you would have to invoke some other mechanism (also involving feature unification) to account for any uniqueness in the way they pattern, but being [NEG +] they can presumably be arguments to the rule.

So again, I’m not an HPSG expert, but the way I understand it all HPSG claims about UG revolve around the idea that humans can perform unification operations, and that they all universally store words as feature bundles, and furthermore, for some people, I presume the inventory of features is genetically determined (however that’s meant to work). So I probably should’ve listed this as another advantage to HPSG: it has a much more plausible implementation of UG.