Thought I'd try one more thing before IntEnum. Automatically converting the style code into an enum by making style_literal a separate function seems to be a better approach? (It solves the double conversion problem, too.)

Sure, that works too.

Some backend tests have failed—I don't suppose they are related to these changes? (I remember that they had passed after commit b7dee27.)

Indeed, these are irrelevant.
Probably this could use a test using check_figures_equal comparing genfrac and dfrac?

Yes, I could add a test for that. One point, though: \genfrac needs rulesize, which only accepts floating-point numbers, whereas \dfrac uses a rule size of self.get_state().get_current_underline_thickness(). Not sure how to proceed.

If I set the rule size to 1 (e.g. \genfrac{}{}{1}{0}{2x}{3y} compared with \dfrac{2x}{3y}), the test passes (it fails for 0 and 2). But that doesn't look like a good test?

One point, though: \genfrac needs rulesize, which only accepts floating-point numbers, whereas \dfrac uses a rule size of self.get_state().get_current_underline_thickness().

I am open to suggestions!

I agree this is a bit annoying :( Probably I would cheat a bit and, in the test, use

thickness = matplotlib._mathtext.TruetypeFonts.get_underline_thickness(None, None, fontsize=<whatever>, dpi=<whatever>)

and insert that value into the test genfrac.

Rebased on the current main branch, since a temporary fix for the failing macOS tests appears to be in place.

Fixed merge conflicts and rebased again. There are a couple of back-and-forth changes, so I will combine everything into a single commit.

The test failures are due to an unidentified image error. Not sure what went wrong.

I squashed the commit. Thanks!