👍 to removing do_event() from the public API, but it is a useful helper for testing. Why not keep it around (rename to _do_event()), and change the function definition to MouseEvent._from_ax_coords(...)._process()? That way it would avoid having to remember to have to call _from_ax_coords() and _process() across the tests.

The problem is that right now do_event takes a (not-standardized) widget method name as parameter. I guess, sure, we could completely change the signature and have a private do_event("button_press_event", ax, xycoords, button) helper method, but if we really think that MouseEvent._from_ax_coords(...)._process() is not ergonomic enough, perhaps that rather points to the fact that that API should be reworked anyways (which is not something I really want to go into right now)?

I guess, sure, we could completely change the signature and have a private do_event("button_press_event", ax, xycoords, button) helper method, but if we really think that MouseEvent._from_ax_coords(...)._process() is not ergonomic enough

I'm +/-0 here. MouseEvent._from_ax_coords(...)._process() is not really elegant:

• We now have an asymmetry in that some events can be created via the constructor, and some via the class method

  KeyEvent._from_ax_coords("key_release_event", ax, end, key)._process()
  KeyEvent("key_press_event", ax.figure.canvas, "escape")._process()

  Note also the difference in passing ax vs. ax.figure.canvas
• explicitly calling _process feels a bit low-level. That was more elegant in do_event.

OTOH I don't have a good API either.

We now have an asymmetry in that some events can be created via the constructor, and some via the class method

KeyEvent._from_ax_coords("key_release_event", ax, end, key)._process()
KeyEvent("key_press_event", ax.figure.canvas, "escape")._process()

Note also the difference in passing ax vs. ax.figure.canvas

Sure, we could require the caller to also pass ax.figure.canvas when using data coordinates (in which case xy could just be given as KeyEvent("key_release_event, ax.figure.canvas, key, *ax.transData.transform((x, y)))) but that feels even more redundant to me. I guess I could inline that at all relevant places if you really prefer it.

explicitly calling _process feels a bit low-level. That was more elegant in do_event.

That point I do feel quite strongly about: directly calling the event handlers, like do_event did, actually abstracted too much, as shown by the changes I had to make to the tests; going via manual event emission is the proper way to simulate what really happens, and does catch additional relevant issues.