Isn't this the default style already?

Yes for image_comparison, but no for cleanup. It looks to me like that is a bug in cleanup.

I am having dejavu about this and some gordian knot we failed to deal with like 6mo ago....

Fortunately, it doesn't seem to matter much. style='classic' is needed two places in test_ticker, and a few more in the specific backend testers. The tests are working as they are. I don't think it is worth spending any time on right now.

unclear what you mean by save...maybe "will be used when the tick values have 4 or more digits"

I agree that it is hard to figure out how to describe this, but it is not a matter of the number of digits the tick values have, it is a matter of how many fewer digits they can have if the offset is used. That's what is meant by "saved". I am using the original terminology of @anntzer from his comment in the code, I believe--though maybe that comment predates his refactoring.

Hmm, maybe then use retain instead? Same basic meaning, but doesn't have associations with saving a figure.

I introduced that terminology (see last chunk of https://github.com/matplotlib/matplotlib/pull/5785/files). It's not "retain", it's the leading digits that are not retained thanks to the use of the offset (i.e., exactly what @efiring said).

I'm so thoroughly confused since save is usually synonymous with retain and yet it's apparently not in this case and I'm not quite sure what it is a matter of how many fewer digits they can have if the offset is used. means in a practical sense.

Sorry, I meant an example of what setting the offset does to the plot. Like I went back to the original issue and the offset code changed [1999-2000] changed the ticks to -10-10, but I don't know what the generic form of that is...basically, I'm also wondering if it makes sense to explain this functionality in terms of what it saves.

Oh sorry. After writing out a description (as best as I could understand it), I'm not sure whether it will help clarify the docs, but maybe I'm just thinking along the wrong lines.

Let's just say there are two ticks, which can be written as abcd and efgh (where each of those letters is a single 0-9 digit). Then pick some other value between them jk00 so that the ticks are abcd - jk00 => -nopq and efgh - jk00 => rstu. If you can pick jk00 such that n == o == r == s == 0, then the offset is used.

The new setting means the "other value" must be jklm and additional conditions of p == q == t == u == 0 must also be satisfied (which it wouldn't be possible in this case, thus fixing dates.)

No worrys, as it sounds complicated. Is there a typical use case for this sort of offsetting or a name for it? (Like it's not scientific notation, which is also partly why it's so confusing)

It seems like it encodes ticks to reduce their size by offset or more digits.

Example: an axis is showing years from 1990 through 2010. This could be shown with ticks from -10 to 10 with an offset of 2000. That is classic mode: 2 digits can be chopped off of the tick labels, so the offset is used if useoffset is True. With the new setting, saving those two digits in the tick labels is not enough, so the offset won't be used, and users will be happier. But if you have an axis from 20001 to 20003, then it will be given an offset of 20000, and the ticks will go from 1 to 3. The tick labels have been shortened by 4 digits, so they will have a better chance of fitting in the frame without overlapping.

A version of this was put in place long ago. Its purpose is to allow mpl to handle a wider range of cases by default, without excessively long tick labels.

While that explanation is too long in docs, it's by far the clearest. Thanks!

So I think it's something like "Computes offset when ticks have more significant digits than offset_threshold. Displays tick-offset at each tick location to reduce tick label size to offset_threshold."