[DRE-maint] Bug#979731: rdoc should be Multi-Arch: foreign

[Helmut Grohne]

Package: ruby
Version: 1:2.7+2
User: debian-cross at lists.debian.org
Usertags: ftcbfs
Control: affects -1 + src:ruby-tiago

This is a difficult one, please sit down and take a cup of coffe or
equivalent.

I've attempted cross building ruby-tiago and it failed installing
Build-Depends in mysterious ways. dpkg would print introduce an error
message, but then the actual error message would go missing. Jessica
Clarke quickly identified the cause and Guillem Jover quickly fixed it.
Now dpkg says that the dependency on rdoc cannot be satisfied.

The next bug happens to be in apt. apt and dpkg disagree on dependency
resolution. When an M-A:allowed package (such as ruby) provides a
package (such as rdoc), apt wrongly treats the provided package as if it
were M-A:foreign. As such, apt thinks that a host architecture
dependency on rdoc can be satisfied with a build architecture ruby, but
dpkg disagrees. The dpkg interpretation looks more reasonable to me.
Let's pretend the apt bug was fixed.

Now satisying a host architecture rdoc needs a host architecture ruby.
Of course such a ruby cannot be run. That's not what we want here. So we
actually want a build architecture ruby. I've asked around how rdoc
works and it seems like it parses Ruby and C sources rather than
executing and introspecting them (like e.g. Python does). As such rdoc
would be a prime candidate for being marked M-A:foregn - if it was a
real package.

And there we are. Providing virtual packages from a M-A:allowed package
is practically broken and we'd like ruby to remain M-A:allowed while
marking rdoc M-A:foreign. That's not gonna work without adding more
binary packages.

So I propose a number of ways foward.

The obvious approach is to introduce a real M-A:foreign rdoc package.
It would depend on ruby:any. It would seem kinda obvious to actually
ship the rdoc binary in rdoc, but that'd result in ruby having to depend
on rdoc. That'd result in a circular dependency. So maybe just leave
rdoc as an empty meta package instead.

The other way round would be saying that /usr/bin/rdoc belongs to the
interface of ruby:any. When you want to use rdoc, you depend on ruby:any
(not rdoc). That'd imply dropping the provides. No binary package
depends on rdoc. Three source packages Build-Depend on rdoc. So while
this seems like a regression initially, it actually solves the issue
with changing just four source packages.

How should we move forward?

Helmut