Can you try turning down the tolerance and seeing how many other tests fail and in what way? It would be good to know if there are any regressions.

        @image_comparison(baseline_images=[filename], extensions=extensions,
                          tol=32)  # <-- Change here

@tacaswell did an overview of the mathtext tests with lower tolerance in #4887, but ended not using it because of lack of time in checking the images. Quite a lot of them fail with tol=15, so we should probably take a look at them at some point.

I didn't mean to lower the tolerance permanently, necessarily, just lower it enough to see if this change causes any regressions. I can probably find some time to sort through them soon.

I did a run with tol=0 following the method of @tacaswell in #4887 using v1.5.0rc3 and I get the following mathtext tests failing:

• 00: baseline image is wrong, symbols are not properly spaced.
• 03: dot decimal spacing is right in baseline and not in master, should be fixed by this PR
• 07: baseline image is wrong, symbols are not properly spaced.
• 09: spacing is wildly different, I imagine the baseline is wrong since you recently did a PR on mathtext spacing to match TeX?
• 10: files are almost identical (RMS 0.008 in stix svg).
• 18: baseline is wrong (symbol spacing)
• 19: idem
• 22: idem
• 25: idem
• 26: accents are shifted, the baseline looks worse to me
• 28: baseline is wrong (symbol spacing)
• 33: idem
• 35: idem
• 38: idem
• 39: accent location, same as 26
• 41: stixsans svg fails with RMS 0.006
• 44: baseline is wrong (symbol spacing)
• 45: idem
• 47: idem
• 48: idem
• 52: idem
• 60: idem
• 63: idem
• 64: another spacing issue with \, as in 09. baseline looks wrong to me: 2 \, a should render to a very small space, and the baseline spacing is pretty big.
• 65: spacing difference after ! (baseline has no space, master has it as per the _punctuation_symbols spacing)
• 67: baseline is wrong: symbol spacing and spacing after comma
• 69: accent position, as in 26, 39
• 71: different spacing between symbols
• 72: idem

There are a lot of them related to recent PRs in which we did not replace the baseline images because the tests where not failing (#4887 for symbol spacing, #4403 for spacing symbols such as \,). If we replace these and the accent ones we could drop the tolerance significantly to be able to catch spacing issues in future changes.

Regarding this PR, I don't think there are any relevant regressions. I believe it only affects test 03 where the baseline image is closer to this PR than the master render. Most of the other uses of decimal separators in the tests suite (tick labels mostly) are not in a mathtext enviroment, so this PR does not affect those.

I have removed the new test because test 3 should have caught this problem if the tolerance was lower. When #5307 is merged, that test will be enough.

Just a note to myself and others: To avoid recreation of too many baseline images, this should be merged before #5307, the test images regenerated in #5307, and then merge #5307.

Backported to v2.0.x as 60b60fb