feat(pypi): add native local wheel override support to bzlmod pip.parse

[meteorcloudy]

This feature allows developers to easily test locally built Python wheels (e.g., wheels built locally before running integration tests) without needing to publish them to an index or manually edit requirements_lock.txt.

This is a proper implementation of the local_wheel_dist_folder feature used in JAX/TF/XLA.

• https://github.com/jax-ml/jax/blob/d72694890a8dd1911435edc8f5732af0c58054f5/WORKSPACE#L50-L62
• https://github.com/google-ml-infra/rules_ml_toolchain/blob/11bbb592cb4d7cbc8ae07d6a4e5ad58ee38605f2/py/python_repo.bzl

How to Use

In your MODULE.bazel file, configure pip.parse with the new local_wheels dictionary mapping:

pip = use_extension("@rules_python//python/extensions:pip.bzl", "pip")

pip.parse(
    hub_name = "pip",
    python_version = "3.15",
    requirements_lock = "//:requirements_lock.txt",
    experimental_index_url = "https://pypi.org/simple",
    local_wheels = {
        "my_package": "dist/my_package-0.1.0-py3-none-any.whl",
        "libtpu": "dist/libtpu-*.whl",
    },
)

Details

• Maps individual package names to local wheel paths relative to your workspace root. Supports middle wildcards (e.g. dist/libtpu-*.whl) and automatically selects the newest version if multiple matching wheels exist. Silently falls back to remote PyPI if the local file is missing.
• Discovery the workspace root via @@//:MODULE.bazel
• Effective only when pip.parse is evaluated by the root module.
• Effective only when index_url is used (which is now by default)

[gemini-code-assist]

Code Review

This pull request introduces support for overriding PyPI dependencies with local wheels located in a specified distribution folder, restricted to the root module. It adds new attributes to the pip extension for specifying the distribution folder and inclusion/exclusion patterns, implements logic to collect and filter local wheels based on ABI markers, and updates the wheel repository generation to use local file URLs. Feedback from the review highlights potential issues with ABI marker generation when patch versions are present and warns against using lexicographical sorting to resolve multiple matching wheels, suggesting instead to fail on ambiguity.