It appears that in 1.5.3 the image was being created in Cartesian coordinates and then clipped to the projection. Is this what you want?

Yes.

That seems like fundamentally incorrect behavior, doesn't it? Wouldn't most users expect that if anything, imshow with a projection would warp the image to that projection? It's never done that, though. In 2.0 its behavior with a log scale on one or both axes has changed so that it does map the image to the scale, but I think that its behavior with a non-affine projection has always been undefined. If so, it should raise an error. Alternatively, I suppose it could warn and drop back to the 1.5.x behavior. Comments, @mdboom?

I was looking through past PRs... did this change occur in #5718?

Oh, OK! I modified my test code to use ax.imshow(..., extent=[0, 1, 0, 1], transform=ax.transAxes):

#!/usr/bin/env python
from matplotlib import pyplot as plt
import numpy as np

lon = np.linspace(-np.pi, np.pi)
lat = np.linspace(-np.pi/2, np.pi/2)
lon, lat = np.meshgrid(lon, lat)

ax = plt.axes(projection='mollweide')
ax.grid()
ax.imshow(lon, extent=[0, 1, 0, 1], aspect=ax.get_aspect(), transform=ax.transAxes)
plt.savefig('test.png')

And it works now:

Oh, OK! I modified my test code to use ax.imshow(..., extent=[0, 1, 0, 1], transform=ax.transAxes):

Was this just a latent bug in my own code that was revealed by the transition to 2.0.0? If so, then we can close this issue.