I simplified my patch and pushed it.  I had to discover again that I needed to inject customized command objects into dist.command_obj, like you found out a few months ago when we had a private email discussion :)

> Using set_undefined_options in install_* is definitely not safe,
> because it fails if parent [command] has not set skip_build.
We can be sure (by looking at the code) that bdist always sets skip_build to something (False by default), so it’s not unsafe.

> Also skip_build has to be set to None in install_* (set_... needs
> that to work) which obviously breaks it.
Yes, skip_build on the subcommands needed to change from None to False, but I don’t see how it is a breakage; it’s just an harmless change needed for this bug fix, so I did it. :)

Thanks to everyone involved in this bug.