In adding tests for this I confirmed that out text tests are not strict enough to catch major issues. The second commit on this PR changes what symbols are in the test and it still passed locally.

@tacaswell: Look at all those nice accentsand symbols:

Thanks for fixing this. Is there a way to set a per-test tolerance in the mathtext tests? I imagine we don't want to jack up the strictness for all tests, but we really need an accent test.

I gave it a shot and made a new category in test_mathtext calledaccent_test with a few accent tests and lowered their tolerance to 16. Using baseline images from this PR, the tests now fail, even though it seems that most of the discrepancy comes from the modified spacing when accents are shifted rather that the accents themselves.

The branch is here:https://github.com/zblz/matplotlib/tree/fix-accent-tests

@tacaswell : do you want me to do a PR on that or will you cherry-pick them here?

I think we do want to push the threshold up on all of them.

I started to do some work last night to sort out what fails when the threshold is push down and some of themshould fail for example\doteq is used in test 01 which has been happily passing.

fun thing to do, apply the following threshold

diff --git a/lib/matplotlib/tests/test_mathtext.py b/lib/matplotlib/tests/test_mathtext.pyindex 745588a..71d9d0a 100644--- a/lib/matplotlib/tests/test_mathtext.py+++ b/lib/matplotlib/tests/test_mathtext.py@@ -151,7 +151,7 @@ for fonts, chars in font_test_specs:                                          def make_set(basename, fontset, tests, extensions=None):                                             def make_test(filename, test):                                                                       @image_comparison(baseline_images=[filename], extensions=extensions,-                          tol=32)+                          tol=10)         def single_test():                                                                                   matplotlib.rcParams['mathtext.fontset'] = fontset                                                fig = plt.figure(figsize=(5.25, 0.75))

python tests.py -s --processes=4 --process-timeout=300 matplotlib.tests.test_mathtext 2> math-fail.txt

withopen('/home/tcaswell/other_source/matplotlib/math-fail.txt')asf:lns= [ln.strip()[51:-5].split('_')forlninf.readlines()ifln.startswith('FAIL:')]df=pd.DataFrame({k:vfork,vinzip(['font','number'],zip(*lns))})

Gives

In [104]: df.groupby('number').count()Out[104]:         fontnumber      00         901         905         906         507         409         915         116         217         224         226         227         930         439         242         244         245         350         951         154         955         356         558         160         661         362         264         972         9

Some of the baseline images (like 01) are clearly wrong.

Brain dump of my notes from the tests that have 9 of 9 fail with atol=0, 63 (?!) of them have at least 1 failure.

*** 00 (fixed) - \dots is really a symbol, should change*** 01 (fixed) - baseline image is wrong, built against bad rendering of \doteq*** 05 - spacing issues*** 06 - spacing issues*** 07 - spacing issues*** 09 - spacing issues, looks like baseline is correct*** 26 (fixed) - accents in the wrong place*** 27 - spacing issues, looks like baseline is wrong*** 34 (fixed) - length of frac line was a bit too long, shortened - probably due to issues with right pad on super*** 39 - wide accents slightly shifted, not clear to me which one is   correct*** 42 - spacing issues*** 44 - frac length issues - binomial does not look right - superscript looks too close to base letter - probably due to issues with right pad on super*** 50 - spacing issue, looks like baseline is correct*** 54 - spacing issues*** 56 not clear either version is correct    r'${\int }_{1}^{x}\frac{\mathrm{dt}}{t}$',*** 60 spacing issues from right pad on superscript*** 64 - spacing issues*** 72 (fixed) - clearly wrong

@zblz I can pull the last 3 commits if they are conflicting with what you have been working on which I have not been following as closely as I should have been.

@tacaswell: They will definitely conflict with PRs#4872 and#4873, for they change spacing of symbols and position of sub/superscripts respectively. Withtol=10 the tests will likely fail with those changes (good!). I removed the baseline images from#4873 because I am still finetuning some corner cases.

If you merge this PR I can rebase on these without baseline images and then regenerate them.

I pulled off the last three commits, I would rather fix them up all at once then have multiple generations of the test images in the repo.

The changes to 01 do not conflict with your work. There will be a merge conflict on 72, but that should not be too hard to clean up.

@matplotlib/developers@mdboom is on vacation, can anyone else review the changes to the parser logic?

The changes to the parser seem like a bit of a hack, but I don't know if that should hold up this PR. I think longer term, the mathtext parser is due for a simplification or overhaul -- it's grown rather crufty with special cases over the years. But perfect is probably the enemy of the good here. At leastsnowflake is autogenerated here so it should work even as new symbols are added.

I picked the name snow flakebecause it is a hack 😉

I spent a fair amount of time trying to write parsing rules that would not match depending on the next character, but I don't have enough experience with parse/lexxing to be sure if I did not sort it out because I don't know what I am doing or because it can't be done.

In either case, this either needs to be replaced by a better solution or merged for 1.5 as my last attempt to touch the parser really really broke it (and it turns out that comment was right).

I'm happy to merge this once it's rebased and tested again...

@zblz@mdboom Both are rebased on current master and all of the test updating shoved into one commit.

The second round of updated test images is cases where all 9 tests failed for that test at a tolerance of 15

Changed my mind about including tests at a higher threshold that is currently required, I don't have the bandwith to check that the new images are any more correct that the old.

ping@mdboom