When we originally updated the query for the tag cloud in 2020 it was under the assumption exact numbers weren't too important because we didn't show them (https://github.com/metabrainz/musicbrainz-server/commit/ace41c641667b1c9d91872996f73cea959773f32). That was incorrect on two counts, we already showed them on hover back then, and the result accuracy is skewed significantly. Removing the limits on each subquery and just limiting the result brings a bunch of entirely new tags to the list, since the current query severely undercounts tags that are used a bit on many entity types in comparison with tags used a lot on one entity type only.

The original worry why these limits were added (a slow query that got a second or so faster with them) still applies, of course. That said, this query runs once, and then is cached for 24h. It seems acceptable to have one person a day wait a bit longer for their results in exchange for actually correct results.