Actually I am thinking to add some machinery to add "as if" comparison tests, which generate two figures and checks that they are indeed identical (in this case, we would be comparing bar([1, 2], ["k"]) and bar([1, 2], ["k", "k"])). This would alleviate the current issue of whether we should try to avoid adding more test reference images which are quite bloating the repo.

Well that sounds indeed cleverer that the genuine image test I was going to write... Looking quickly at test_axes, it seems like none of the parameters of type 'scalar or iterable' is actually tested for ax.bar: this would be a nice addition to have the kind of test you are thinking about.

You can already write this "by hand" (https://github.com/matplotlib/matplotlib/pull/9305/files#diff-663c17d0729218d60dbb862c286a17d4R5327), but ideally the machinery should be such that triage_tests.py can display failures correctly.

For this the simplest test is to pass a scalar and then just iterate over the returned artists and check the edge color is what you expect.

tests in