Interestingly, the test reveals that the Agg backend doesn't handle the font quite right.

Correct output, screenshot from a PDF viewer.

I'll trust you on the ttf parsing, but otherwise the patch looks reasonable.

Oops, made a typo with a push and accidentally closed this.

If someone wants to check the code against documentation, composite glyphs are described at https://docs.microsoft.com/en-us/typography/opentype/spec/glyf#composite-glyph-description and 2.14 fixed-point format at https://docs.microsoft.com/en-us/typography/opentype/otspec181/otff#data-types (search for F2DOT14).

Interestingly, the test reveals that the Agg backend doesn't handle the font quite right.

If I run outside the test framework, it appears to be okay?

Interestingly, the test reveals that the Agg backend doesn't handle the font quite right.

If I run outside the test framework, it appears to be okay?

Yes, I couldn't reproduce this. It might have been some problem with an intermediate version, as I might have forgotten to compile the extension when running some test.

Interestingly, the test reveals that the Agg backend doesn't handle the font quite right.

If I run outside the test framework, it appears to be okay?

Yes, I couldn't reproduce this. It might have been some problem with an intermediate version, as I might have forgotten to compile the extension when running some test.

To clarify, I can reproduce if running in the test suite, by removing the explicit extension from the comparison settings.

This is because tests disable hinting with plt.rcParams['text.hinting'] = 'none'; if you run the script outside the test suite, but add that, then you get the weird streaks. Any setting other than 'force_autohint' or 'auto' (same thing to FreeType, latter is what we call our default) will do it, so it seems to be something wrong in the font hinting itself.

The test failures are because of unrelated Pandas deprecations; apparently Pandas 1.1.0 is now available

This is already fixed on master; you should rebase if you're having trouble with it.

It looks like you may have accidentally reverted some things in that last rebase?

It looks like you may have accidentally reverted some things in that last rebase?

You're right, it looks like I lost the suggestions committed via the Github web UI. I committed those again.

See #18181 for an alternate solution, though (which can reasonably avoid adding a new font file in the repo, and in any case seems more robust in the long term, as it can Type3-subset any font that FreeType can outline-decompose).