Perhaps the simpler check would be self.canvas is self.canvas.figure.canvas? Note that while figure.canvas can change (hence the introduction of _changed_canvas), I think canvas.figure does not ever change (at least in practice... it may even be possible to spec that), so the proposed check is effectively "does the figure I initially got still use the same canvas as initially?" Also none of these attributes should ever be None, AFAICT.

Alternatively, we could additionally record self._figure = ax.figure in the constructor and check self._figure.canvas is self.canvas; again this should sidestep the None problems?

Alternatively, we could additionally record self._figure = ax.figure in the constructor and check self._figure.canvas is self.canvas; again this should sidestep the None problems?

I think that would still miss the "middle removed" issue as we have references that skip generations and we should avoid adding more hard-references than we strictly need....

But I will change to self.canvas is self.canvas.figure.canvas

Hmm, we are back to not being careful enough...but only an azure?!

That's a bit fishy, but perhaps just change all the self.ax.draw_artist(artist) by artist.draw(self.canvas.get_renderer())? (which should bypass the problem of whether self.ax.figure is None or not -- anyways, having draw_artist be an Axes method always seemed a bit weird to me (I do realize that historically we cached the renderer on the axes, but that's not really the logical place to have it)).