Generally 👍 on the concept.

Questions:

• SVG is harder to compare visually than PNG. I assume it's smaller though. Is the added complexity worth the savings compared to just using PNG?
• Is it reasonable/necessary to check all fonts?

I agree pngs would be simpler, but...
From a quick test (on the single test case used here): the current svg baseline is 2.8K (and git should internally gzip it down, which results in a file just below 1K). "Naively" switching to png results in a much larger (7K) file. Somewhat surprisingly, even passing pil_kwargs={"optimize": True} (https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html#png) only reduces it to 6.3K; passing the file through optipng does crush it down to 3.5K though. But git shouldn't be able to further compress down the resulting file via gzip.
Also, I expect svgs to diff much more efficiently if we change e.g. precise glyph positionings.

Checking all fonts is likely mostly relevant for test_mathfont_rendering only, I guess?

OK, so only removing SVG and PS wouldn't gain us much because PNG uses the most space anyway.

I'm very hesitant to use optipng for testing. First that'd make it a hard dependency for testing. And second more importantly, we'd bake the compression algorithm into our tests. If optipng changes somehow it could break all our tests.

It likely has to be SVG if we want to save space.

If the sqrt code itself needs to change (as opposed to its test), it should at least be in a separate commit. I'm not sure why those changes are happening as there's no context for it.

Done.
The changes to sqrt were mostly to make sure that the new machinery works in a concrete use case.

Conclusion from the dev call today:

• Take font spec improvement (first commit); it's independent anyway
• Limit baseline images to SVG
• Don't use fonts in the SVG but keep the fonts being converted to paths.
  Concern is on the added complexity; and the production code uses converted paths, so testing these is closer to the production case. We live with the larger SVG size due to paths in test images for now. Tom is planning to prioritize the separate repo for test images, so image size is a little less critical.

Don't use fonts in the SVG but keep the fonts being converted to paths.

I don't have hard numbers right now but IIRC embedding the paths makes the SVGs much larger, defeating the purpose of the PR... in that case we may just as well stick to png only for ease of use.

You're right. The existing svgs under test_mathtext are roughly 2x the size of the corresponding png. Given that, the pragmatic solution is to stick to png only. So let's

• Take font spec improvement (first commit); it's independent anyway
• Limit baseline images to PNG

milestoning as 3.4 as a dependency for #18916

Aren't we doing #19261 instead?

Yup, milestoned the wrong one :p