[Bug]: add_patch incorrectly decides whether to autoscale axes for non-separable transforms #26796

Bug summary

Axes.add_patch(patch) is intended to update the axes limits based on the patch limits only if that can help making the patch fit within the view; in other words, if the patch uses e.g. Axes.get_xaxis_transform() (i.e. is an axvspan), then autoscaling the y (data) limits doesn't do anything to help making the patch fit within the view as the patch's y coordinates are in axes relative units.

However, this patch is incorrectly implemented for non-separable axes.

Code for reproduction

from matplotlib.patches import Rectangle
from matplotlib import pyplot as plt

fig = plt.figure()
axs = [fig.add_subplot(121), fig.add_subplot(122, projection="polar")]
for ax in axs:
    theta0 = 0; dtheta = 1
    r0 = 0; dr = .25
    ax.add_patch(Rectangle((theta0, r0), dtheta, dr, transform=ax.get_xaxis_transform()))
    ax.autoscale_view()
plt.show()

Actual outcome

On the rectilinear axes, add_patch correctly understands that the rectangle uses relative y limits (0, 0.25) and therefore keeps the default y limits (-0.05, 0.05)).
On the polar axes, it doesn't (even though the patch indeed goes to 0.25 the max radius), and thus tries to autoscale the radial axis to (0, 0.25) (+ margin).

Expected outcome

No autoscaling of the y limits on the second plot.

Additional information

I suspect this comes down to understanding/writing down what would be the really useful semantics for contains_branch_separately (upon which add_patch relies) in that case.

(Noted while working on #26788, where I just manually undo the "wrong" axes limits update.)

I suspect that a proper generic solution for #15518 (which was fixed by special-casing in #25324) would require something similar too.

Operating system

macOS

Matplotlib Version

3.8

Matplotlib Backend

qtagg

Python version

3.11

Jupyter version

none

Installation

git checkout