PR summary

Closes #27763.

I have not fully understood what was going wrong, but I did discover that when fig._get_renderer is called, fig.dpi gets reset to 72. As the renderer had deliberately set it to something else here, this is probably not helpful! Anyway, it is intuitive to me that the renderer used to work out where things are should be the same renderer that does this part of the drawing. Taking a simplified version of the #27763 example,

from mpl_toolkits.axes_grid1.inset_locator import inset_axes
import matplotlib.pyplot as plt
import numpy as np

fig, axes = plt.subplots(1, 1, dpi=100)

x = np.random.randint(100, size=(10, 200))
c = axes.imshow(x, origin='lower', cmap='Reds', aspect='auto')
axins1 = inset_axes(axes,
                    width="10%",
                    height="30%",
                    loc="upper left",
                    bbox_to_anchor = (0.045, 0., 1, 1),
                    bbox_transform = axes.transAxes
                    )

fig.colorbar(c, cax=axins1, orientation="horizontal")

fig.savefig('test_single.pdf', bbox_inches='tight')

With main:

With my branch:

The alignment is still not perfect (as also shown in the new test image), but it looks identical to what I get with v3.7.1 (before I started breaking things 🐑). I note that several commenters on the issue seemed happy with v3.7.1 behaviour. Also I cannot see any offset if I increase the dpi to 300:

I propose to open a new issue to track the remaining small offset.

PR checklist

• "closes #0000" is in the body of the PR description to link the related issue
• new and changed code is tested
• [n/a] Plotting related features are demonstrated in an example
• [n/a] New Features and API Changes are noted with a directive and release note
• [n/a] Documentation complies with general and docstring guidelines