Combine sys_platform and platform_system in marker algebra

[charliermarsh]

Summary

This PR adds a representation to the marker algebra to capture "known pairs" of sys_platform and platform_system values.

For example, we want to be able to detect that sys_platform == 'win32' and platform_system == 'Darwin' are disjoint.

There is some risk here, in that if there are platforms for which, e.g., sys_platform == 'win32 but platform_system is not Windows, we'll get things wrong. I'm trying to weigh whether that's a real risk.

In theory, we could maybe also include os.name here?

Closes #7760.
Closes #9275.

[Vinno97]

In theory, we could maybe also include os.name here?

Yes. I was just about to comment on #9275 about this.

For example, jupyter-server has a dependency on pywinpty for os_name == 'nt'. If want linux-only lock-file, a full dependency string would then have to be: platform_system == 'Linux' and sys_platform == 'linux' and os_name == 'posix'

[charliermarsh]

Yeah. I think it's a bit more challenging because there isn't a 1-to-1 mapping between os.name and sys.platform for Linux and macOS (they both use posix).

[charliermarsh]

(Not impossible but different than what I've encoded here.)

[Vinno97]

Since os_name == 'posix' could also be true for BSD and other more esoteric POSIX operating systems, perhaps os_name == ' posix' should entail os_name != 'nt', with its subsequent platform_system and sys_platform restrictions?

[charliermarsh]

On reflection, what I have here is not correct. It fails this test (i.e., it thinks these are disjoint):

assert!(!is_disjoint(
    "sys_platform == 'cygwin'",
    "platform_system == 'Java'"
));

[charliermarsh]

Somewhat losing confidence here because it turns out Android also returns "Linux" for platform_system? Despite sys_platform returning "android".

[Vinno97]

It seems like Poetry has gone though quite the rabbit hole for this themselves: python-poetry/poetry#5192. But even after quite some pull requests, they closed the isue with this comment:

Closing since there haven't been much improvements/fixes anymore in the last time and I can't come up with a good solution for merging sys_platform and platform_system. - python-poetry/poetry#5192 (comment)

It appears this may be an unresolvable issue, though perhaps it's feasible to just cull known impossible environments? I.e. we know that platform_system=='Linux' will never occur together with sys_platform=='win32' or sys_platform=='darwin'.

[charliermarsh]

I think it's absolutely possible to solve in our marker system, I just don't have the right solution here.

[konstin]

This is a tough one, my intuition is to encode know mismatch rather than known pairs. There's a tendency to subdivide platforms to accommodate new developments (the development of rust target triples is a good reference since it's more upstreamed than the python build vars), so i would encode tier 1 and tier 2 pairs that never match, such as darwin sys_platform and Windows in platform_system.

[Vinno97]

@konstin that's a good point and probably catches the majority of the standard cases with a smaller chance if breaking future cases

[charliermarsh]

Yes that's also what I want to do. Here's a new idea: we encode all mutually exclusive pairs, and for each pair (A, B), we add and (not A or not B) to each expression. In practice, it's probably easier to have a method like is_impossible alongside is_false that just does a single operation to check this for all pairs at once.

[charliermarsh]

Closing in favor of #9444.