This is a deep-seated semantic confusion going back to at least 2006 [1] when the Protein Infobox had Entrez and OMIM gene IDs. Freebase naively adopted in its initial protein schema in 2007 when it was importing from those infoboxes. Although it made some progress in improving the schema later, anything not aligned with how Wikipedians want to do things is shoveling against the tide. It's also very difficult to manage equivalences when Wikipedia articles are about multiple things like the protein/gene articles.
If you look at the recent merge of Reelin [3] you can see that it was done by the same user who contributed substantially to the article back in 2006 [4], so clearly, as the "owner" of that article, they clearly know what's best. :-) It's going to be very difficult to get people to unlearn a decade of habits.
Another issue is that, as soon as you start trying to split things out into semantically clean pieces, you immediately run afoul of the notability restrictions. Because human (and mouse) genes don't have their own Wikipedia pages, they're clearly not notable, so they can't be added to Wikidata.
This problem of chunking by notability (or lack thereof), length of text article, relatedness, and other attributes rather than semantic individuality is much more widespread than just proteins/genes. It also effects things like pairs (or small sets) of people who aren't notable enough to have an article on their own, articles which contain infoboxes about people who aren't notable, so they got tacked onto related article to give them a how, etc.
The inverse problem exists as well where a single semantic topic is broken up into multiple articles purely for reasons of length. Other types of semantic mismatches include articles along precoordinated facets like Transportation in New York City (or even History of Transportation in New York City!), list articles (* Filmography, * Discography, * Videography, List of *). Of course, some lists, like the Fortune 500, make sense to talk about as entities, but most Wikipedia lists are just mechanically generated things for human browsing which don't really need a semantic identifier. Freebase deleted most of this Wikipedia cruft.
Going back to Ben's original problem, one tool that Freebase used to help manage the problem of incompatible type merges was a set of curated sets of incompatible types [5] which was used by the merge tools to warn users that the merge they were proposing probably wasn't a good idea. People could ignore the warning in the Freebase implementation, but Wikidata could make it a hard restriction or just a warning.
Tom