I haven't looked in depth into this, but why did #16832 (which fixed #7184) not fix this? attn @brunobeltran perhaps?

It looks like #16832 fixed Path.get_extents, but Path.get_extents doesn't seem to be used in Axes._update_patch_limits, which controls the autoscaling, and which I've tried to patch in this PR. Thanks for pointing out #16832, I can probably replace my horrendously large new function with something smaller from that.

So you're trying to replace individual points by the extent corners only? How does this affect log scales? Because the fix in #18642 was to not do that as much.

Log scaling seems fine; the failing examples seem to be mainly pie charts. Most of the changes look reasonable, but there's a couple that definitely were better before this patch, so I need to do a bit more investigating.

😆 it turns out that log scaling wasn't quite fine. I think I've fixed it this time, but will see how the tests go.

This is now good to go, and was a pretty small change (in terms of lines of code) in the end.

ping for a second review on this!

@dstansby popping to draft until you can address review comments, but feel free to pop back on the queue!

The change in lib/matplotlib/tests/baseline_images/test_axes/pie_frame_grid.png does not seem better?

I think this is because the (0, 0) point is sticky, and is now being snapped into view.

I don't understand how that happens?

My suspicion is that this PR has uncovered another bug somewhere in (auto)scaling - I'll try and work out what's going on and report back.

I'm going to leave a trail of crumbs as I investigate the funny axes scaling.

The issue only seems to appear with plt.style.use('classic') (which the tests use by default), where the pie text doesn't seem to be taken into account in the scaling:

import matplotlib.pyplot as plt
plt.style.use('classic')

# The slices will be ordered and plotted counter-clockwise.
labels = 'Frogs', 'Hogs', 'Dogs', 'Logs'
sizes = [15, 30, 45, 10]
colors = ['yellowgreen', 'gold', 'lightskyblue', 'lightcoral']
# only "explode" the 2nd slice (i.e. 'Hogs')
explode = (0, 0.1, 0, 0)

plt.pie(sizes, explode=explode, labels=labels, colors=colors,
        autopct='%1.1f%%', startangle=90,
        wedgeprops={'linewidth': 0},
        frame=True, center=(2, 2))
# Set aspect ratio to be equal so that pie is drawn as a circle.
plt.axis('equal')
plt.show()

It seems that autoscaling doesn't ever take into account the labels added by pie() (I've opened #22057 to track this). By coincidence the older test images never encountered this issue, but one of the new test images does because of this fix.

As such, I'd advocate for merging this to solve the Bezier scaling issue, and track the issues with pie() autoscaling separately.

That sounds reasonable. Can you fix the test so that is not classic and we can track pie issues separately. Definitely ping if you need a review.

Can you fix the test so that is not classic and we can track pie issues separately.

👍 - I updated all the test images that are changing anyway to mpl20 style.

To check my understanding the pie images are regenerated because:

• previously the auto scaling was wrong, but worked out because the limits were too big because the control points were significantly away from the
• with this patch the scaling was too tight as under the "classic" style was being right at the edges of the pie chart which makes the text go out of the axes
• but with v2 setting there is a margin so there is enough space for the text again

If my understanding is correct @dstansby (or anyone else) can (self-)merge.