Your commit has erased all the other tests. Suggest you copy and paste your test somewhere, do a hard rebase on upstream/main and try again.

Further, this shows that violinplot does not participate in the units system currently (for the main axis, I think it would be fine with doing positions in dates, which perhaps we should test regardless).

And arguably, it shouldn't?

We have a similar problem in #27038 (boxplots) because these statistical measures don't really translate well to dates.

OTOH, I think it could translate well to physical units, so perhaps we should make it so that at least we try to adhere to the units behavior...

I could make an argument for doing the unit conversion and doing the KDE/stats on the resultant floats (not sure I love that, though for most things I think it should be fine... its really nonlinear units that we'd have to worry about there...)

All that said, I am now curious if at least the positioning values (position and width) work as desired (with the latter being a timedelta). Because while doing statistics on dates can get strange fast (not impossible, just have to be careful about knowing what you are doing), saying "I want a violin plot comparing measurements taken on different dates" seems completely reasonable.

I think my leaning is to deprioritize fixing of units along the stats axis of these, but ask that this test case is written to focus on positions instead for the time being.