Non-universal resolver should ignore requires-python upper bound

[konstin]

Without universal resolution, we fail to resolve colabfold 1.5.5 on python 3.12, since it only supports python 3.9 to 3.11:

$ echo "colabfold==1.5.5" | uv pip compile -p 3.12 -
    × No solution found when resolving dependencies:
    ╰─▶ Because the requested Python version (3.12.0) does not satisfy Python>=3.9,<3.12 and colabfold==1.5.5 depends on
        Python>=3.9,<3.12, we can conclude that colabfold==1.5.5 cannot be used.
        And because you require colabfold==1.5.5, we can conclude that the requirements are unsatisfiable.

We generally ignore upper version bounds (#4022), the universal resolver also resolves succesfully:

$ echo "colabfold==1.5.5" | uv pip compile -p 3.12 --universal -
Resolved 62 packages in 23ms
[...]
colabfold==1.5.5
[...]

We should change the regular resolver to ignore this requires-python upper bound, too, and make it clear in the docs that we ignore all requires-python upper bounds.

[charliermarsh]

Why is the extra relevant?

[konstin]

We should error both with and without extra due to the mismatch in the python requirement, after finding uv behaves different depending on whether an extra is present (#5043), i'm just testing both cases. Fwiw without the extra the resolution is simple (it's a stub, functionality-wise), with the extra it's a complex machine learning tree.

[charliermarsh]

This is correct though. We do not respect upper-bounds on requires-python.

[charliermarsh]

uv pip compile -p 3.12 - is resolving for Python 3.12, while uv pip compile -p 3.12 --universal - is treated as requires-python: ">=3.12". This is mentioned in the documentation.

[konstin]

I've changed the issue to changing the regular resolver instead of the universal resolver. We should also make it clearer in the docs that we ignore all upper bounds on requires-python, at least i couldn't find that information when looking for it.