The basic goal is to allow depending on a build-time generated wheel (i.e. py_wheel without having to go through the zip/unzip process of creating a wheel archive. Wheels may be large in size, so the process of staging, archiving, and extracting them is unnecessary overhead.

So the idea is py_wheel provides (prefix, files) information, which can then be used by consumers. This then allows e.g.

py_wheel(name="foo_whl", ...)
py_wheel_as_py_library(
  name = "foo_whl_as_pylib",
  srcs = [":foo_whl"]
)
py_test(
  name="foo",
  deps = ["foo_whl_as_pylib"],
)

Where py_wheel_as_py_library returns PyInfo() et al pointing to what the whl would extract to (i.e. a directory with the contents of the wheel).

Looking at py_wheel, this looks quite doable, since Starlark is already generating most (if not all?) of what goes into the wheel, and the wheelmaker.py program is mostly just zip's the files up.