One thing to decide is whether to apply the line gap for a single line. Previously, I don't believe we ended up doing that, because the line spacing was added before the second and subsequent lines. This PR currently always adds it, which means the text bounding box is taller in all cases. This causes such changes as the fancy bbox patches in boxarrow_test_image.png to all be larger, or e.g., every multi-artist legend to be larger. To minimize such changes, we can decide to drop the line gap for single-line text entries, if we want?

Discussed on the call today and decided not to add line gap for a single-line text. This should be closer to what we used to do, and will minimize the number of changes that occur due to this.

I've now pushed fixes to the text_placeholders fixture, which should reduce the number of test image changes. Also fixed some tests that explicitly check values of text locations/sizes/etc.

There is one (set of) regression for which I have not pushed the image updates (and so why tests are failing). All the vector multi-font tests now have a reduced line height:

vs before:

This is because the test uses Computer Modern as its primary font, with fallback to DejaVu Sans for the other characters (which is on purpose to test fallbacks). While the ascent/descent metrics for Computer Modern are larger than the "lp" measurement, that font does not have a nonzero line gap. Additionally, since we don't do the fully-correct thing of looking at all font metrics for a used line, while we do see the ascent/descent of DejaVu Sans, we don't know that DejaVu Sans has a larger line gap. Since we aren't applying any other line spacing, all the lines now look a bit cramped. This last thing cannot be fixed until we move to processing all text in the backend where we can inspect all glyphs at a time.

I see three options for handling these:

1. The easiest fix here would be to set a fixed line spacing in the tests themselves.
2. If we think this might be something that users might complain about, then we can hard code a line gap for Computer Modern only to match the !~120% line spacing that we used to do. (But keep in mind that Computer Modern is the old default font, not the current one, and this would only affect multi-line text.)
3. Just ignore it entirely.

Separately, I did mention that I would run the constrained/tight layout tests without placeholders so that we could compare the results. I did that, and posted the results to a gist here: https://gist.github.com/QuLogic/254ee85ecaa778dd44635924d706acea

Consensus on call is to ignore for now and let it be cramped for now. If we fix this later a few tests would be acceptable.

OK, pushed the images to the text-overhaul-figures branch, and dropped them from here. Will merge after CI is okay.

So it looks like some of these tests are failing on platform/arch-specific ways. What I'll do for now is merge this one since it's holding up everything, and put in a smaller PR upping the tolerances on the ones that are broken.