I'm a bit skeptical of this overall, but it's obviously complicated to get data that filter out the effects of sociolectal solidarity. (Can you get people to speak "naturally" but without priming them to speak or to avoid speaking in a particular register or sociolect!?)
which can become conventionalized and eventually part of native speakers' intuitions even though they can be contrary to a more logically consistent or simpler understanding of agreement.
There are well-known phenomena like
https://en.wikipedia.org/wiki/Attraction_(grammar)
which can become conventionalized and eventually part of native speakers' intuitions even though they can be contrary to a more logically consistent or simpler understanding of agreement.