BugFix #14329

[Wintreist]

Closes #14329

Fixed the issue of getting the really closest label to the test. I abolished the priority of base classes over labels, but supported getting the usefixtures mark in reverse order (i.e. from the base class to the heir), since this is a unique case when you really need to call fixtures sequentially from the base class.

I did not add myself to the Authors, as I do not plan to be a pytest developer yet, I fixed the problem, as it interferes with my main project. Thanks

[bluetech]

Thanks, the changes LGTM. I pushed a few tweaks.

I'll let others a few days to review, then merge.

[RonnyPfannschmidt]

Under strict pedantic this is a breaking change
Im a bit torn for the classification

[bluetech]

Under strict pedantic every bug fix is a breaking change 😀

I think this case is sensible for a minor release. I don't actually expect the MRO thing to come up that much...

[bluetech]

Maybe the usefixtures change could be a bit more disruptive, but it will be annoying to separate the two...

[RonnyPfannschmidt]

The order of items changes slightly

That likes to break things

[bluetech]

The order of items changes slightly

What items are you referring to here? The mark list items, or the Items (tests) order changing due to usefixture changes?

That likes to break things

Would you prefer to defer this PR to pytest 10?

[RonnyPfannschmidt]

Thanks for the patience here, and sorry it took this long to get to a decision.

I'm going to land #14630 instead and close this one — but your test goes in with it verbatim, with a Co-authored-by trailer, and you're in AUTHORS. The diagnosis and the regression test were the hard part; the disagreement is only about which layer the fix belongs in.

The concrete reason: reversing get_unpacked_marks also reverses Class.own_markers, and PyobjMixin.obj does

self.keywords.update((mark.name, mark) for mark in self.own_markers)

which is last-wins. With base-first own_markers that puts the closest marker into keywords; reversing flips it to the base class one:

@pytest.mark.foo("base")
class TestBase: pass

@pytest.mark.foo("child")
class TestChild(TestBase):
    def test_it(self, request):
        request.node.parent.keywords["foo"].args[0]  # "child" on main, "base" with this PR

Nothing in the suite catches it — test_unpacked_marks_added_to_keywords uses distinct mark names per level. It is one reversed() away from being fixed, but it is the kind of fallout that makes me want the fix in the iteration layer rather than in the stored order: own_markers is public and plugins read it.

#14630 leaves own_markers alone and overrides iteration on Class instead. Same user-visible fix for #14329, same test.

For the record, one of my claims about this PR was wrong, and I've corrected it on mine: I said reversing breaks parametrize ID ordering in a way an iteration-layer fix avoids. It doesn't. _genfunctions consumes definition.iter_markers("parametrize"), so both approaches reorder inherited class level parametrize IDs identically (TestChild::test_it[1-x] becomes TestChild::test_it[x-1]). That is now called out in the changelog on #14630, since it invalidates pinned ID selections and the --last-failed cache.

@bluetech you approved this one — say the word if you'd rather go this way instead, reopening is cheap.