I think most of these of these are the result of the “hops and jumps” methodology that the Census Bureau used to grow an urban area beyond its initial urban core (which is based on a group of census blocks), specifically what’s known as a road connection. On the other end of the spectrum, apparently it wasn’t aggressive enough to link Concord–Walnut Creek to San Francisco–Oakland through the Caldecott Tunnel across a mountain range, which was a point of contention in Slack.
This is reminiscent of something @ZeLonewolf tried recently:
Putting aside the taste-testing, one problem with this approach is that it prioritizes state boundaries that are more or less artificial. Is Texarkana half a city, half a town, because it straddles the line between the big state of Texas and the small state of Arkansas?
If we want to scale place classification to achieve an even density, we would instead use something spatially uniform like quadtiles or hexbinning. But data consumers like OpenMapTiles and external datasets like Natural Earth can already implement scale ranking more effectively, even varying it by zoom level and projection. It goes back to the question of why uniformity would matter for a key like place=*. As I demonstrated earlier, a map can still achieve a usable information hierarchy without uniformity.