cc @timhoffm as the author of #22587 (and API lead).

Thanks for bringing this up. I think we need a discussion what we mean by clear().

I don't have the full history present and think we might have been inconsistent here.

From a conceptual point of view. IMHO clear() should reset all stylistic attributes (including tickdir, which feels the same category as linewdith to me). #22587 is a rewrite to explicitly not reset tick and ticklabel visibility. This is because visibility is more structural than stylistic. It's e.g. set by shared layouts and resetting visibility would make clear() unusable for shared Axes. Whether that's the clean way would be a separate discussion, but since we don't have a persistent concept of layout connected to tick visiblity, not resetting visibility is the only viable approach. I don't think these arguments hold for other tick properties.

To be checked: We might have been inconsistent and even changed behavior unintendedly or at least without warning. So independent on what is a good behavior, we need to check the past and current behavior and also make sure we have consistency, or if that is not possible at least a documented behavior over time.

My intuition is that clearing the axes removes artists, puts the axes back to its default axes limits, and clears label. Ie all data added to the axes would be removed.

OTOH, axes styling - tickdir, tick length, which spines are shown, spine styling, etc would not be reset as these are independent of the data on the axes. That is what I believe the behaviour of Matlab is, and I think it is the behaviour Matplotlib has had until recently.

Note that I opened #23808 to restore the behaviour of matplotlib 3.5, if you guys want to stay conservative.

I'm inclined to be conservative and go back to the 3.5 behavior rather than adding a behavior change note at this point.

I think it is better be consistent through time but inconsistent through the API (unless we can make the API fully consistent in one go or at least have a clearly articulated plan of where we are going).

I'm inclined to be conservative and go back to the 3.5 behavior rather than adding a behavior change note at this point.

Agreed, but still we need to consider two which properties this should apply. For example, 3.5 also keeps tick_params(... width=10) under clear() whereas 3.6rc does not. I suppose this holds for a lot of the tick_params.

We should add a figure comparison test

params = {direction='in', width=5, ...[lots of other params to persist]}
fig_test.tick_params(**params)
fig_test.clear()
fig_ref.tick_params(**params)

to make sure we are don't have an unintended regression through the clear refactoring.

This was fixed by #23834; discussion for what to do long term is in #23839.