Fix `geom_violin(position = "dodge2")`

[teunbrand]

Absolutely horrifying, perhaps even more so after the quantiles were reworked:

library(ggplot2)

ggplot(mpg, aes(drv, displ, fill = factor(year))) +
  geom_violin(position = position_dodge2())

^{Created on 2025-09-29 with reprex v2.1.1}

@karawoo had described this many moons ago https://youtu.be/UGAuXyxdrvA?si=q9GH8ety6Xj1ak5t&t=986, but I couldn't find a related issue that kept this on the radar. This issue is that blip.

I think fundamentally position_dodge2() doesn't know the relation between rows and groups. For many geoms, 1 row is one group, like geom_boxplot(), but in others there is a many-to-one relationship, like geom_violin() or perhaps geom_dotplot(). I think the meat of the issue is how we can communicate this information between geom and position. The implementation at the position level worries me less.

[karawoo]

This would be soo nice to fix. I remember feeling like it was intractable back in 2017, but with everything you've been doing lately that may be less the case today.

[karawoo]

Absolutely horrifying

also yes 😬

[clauswilke]

I think the meat of the issue is how we can communicate this information between geom and position.

I've always felt the fundamental problem is that positions don't know anything about aesthetic mappings. Not sure if that can be remedied somehow. I vaguely remember that I thought about this quite a bit at some point but then gave up. There may be some issues still floating around where I made comments along those lines.