Fix pinning expressions for prerelease builds

[isuruf]

No description provided.

[isuruf]

cc @mbargull

[mbargull]

Ah, didn't see this before, thanks for the cc!
I think we have multiple issues here:

1. apply_pin_expressions does not respect certain corner cases like pre-releases, epochs and thus needs fixing
2. apply_pin_expressions should not be run on build string-pinned deps (because you also get things like python 3.10.0 h1234567 => python >=3.10,<3.11.0a0 which is not correct; see conda search --info conda-forge::libpython-static)

For 2.: Maybe something along this (in conda_build.render)

 def get_pin_from_build(m, dep, build_dep_versions):
+    match_spec = MatchSpec(dep)  # from conda.exports import MatchSpec
+    if match_spec.get_exact_value('version') and match_spec.get_exact_value('build'):
+        # exact pin; don't apply pin expressions
+        return dep 
     dep_split = dep.split()

might work. I haven't tried this one yet, though.

[isuruf]

If 1 is fixed (at least the pre-release part) as in this PR, then 2 becomes not so important. 2 would be nice to have, but not necessary. Note that fixing 2 alone is not enough for our python issue if we start building downstream packages with rc0.

[mbargull]

Note that fixing 2 alone is not enough for our python issue if we start building downstream packages with rc0.

Yep, I know. Either we fix 1 and get a new release with that fix or we'd have to mark the RC as python=3.10.0=rc_h123457 or similar as a workaround (and then of course always have it remain in the pre-release channel, i.e., never main label).

[isuruf]

Either we fix 1 and get a new release with that fix

Which is what this PR does

[mbargull]

Either we fix 1 and get a new release with that fix

Which is what this PR does

yeah, that was mostly meant regarding "and get a new release" ;).

[isuruf]

This is ready for a review

[mbargull]

@isuruf, I've added some commits for docs/news and adjusted the tests (I've duplicated the test to add one that specifically uses python -- that's technically not necessary, but to ensure we don't regress in "special handling cases" like with python). Let me know if you are fine with these changes or revert/amend as you see fit.

[isuruf]

Looks good. Thanks

[mbargull]

Merging this since all CI failures are unrelated to the PR.

@chenghlee, @jezdez, if we want to go ahead with a RC-based migration on conda-forge ( conda-forge/python-feedstock#481 , conda-forge/conda-forge.github.io#1499 ), we'd need a new conda-build release for this. I'm not aware of any other things we'd need though; @isuruf might know.

[isuruf]

Thanks @mbargull. I need #4248 as well for building GCC.