Thanks for the feedback. We know that matplotlib unfortunately has some inconsistencies in its API. Matplotlib has been developed by different people over a very long time, and API consistency and design has not always been watched for as strongly as we do nowadays.
Today we have to make a trade off between improving the existing API and breaking changes we would introduce with that. Given that there are millions of users, we have to be very conservative with breaking changes.

Just weighing in here with my opinion (Instead of opening a new issue: https://twitter.com/sjvrijn/status/1395746154223181828 ) The ncol in legend() versus by now pretty consistently used ncols elsewhere is, I feel, an example that might be worth changing by now, or at least start accepting ncols as well...
Maybe as a "good first issue"?
Or is it worthwhile to start collecting a list of such minor inconsistencies to target for a next release (more explicitly than the label does)?

For the specific case of here I would agree with making ncols a synonym for ncol (or rather having ncols be the canonical name). This can be done using _api.rename_parameter with a small modification to allow having just the renaming effect, but never emitting a deprecation.

This was fixed in 3.6: https://matplotlib.org/stable/users/prev_whats_new/whats_new_3.6.0.html#ncol-keyword-argument-to-legend-renamed-to-ncols

PR: #23198