Yes, you are right. So the question is how long to support systems
with the old cabal 0.10. This is the one included with the previous
haskell platform (and thus lots of linux distro's), which is less than
a year old. But it's also pretty old, since there weren't any cabal
releases for a while.

That's a very awkward situation. At least in the future, Johan and I have a proposal to make this class of problem more avoidable by introducing a regular release schedule. See the thread that starts here for details: http://www.haskell.org/pipermail/cabal-devel/2012-August/008987.html

For the state of things today, it's not obvious to me what to do.

It's burdensome to ask package authors to remove stuff from their packages because it can't be handled by a broken version of cabal, especially since there's no upper bound on how long that broken version will be floating around. We'd essentially be giving up on this feature semi-permanently, which would make me sad because it's so useful.

Just as unappealing is the idea of breaking builds for people who, through no fault of their own, are using the broken cabal. However, at least this class of people has the incentives aligned to do something about their problem: either upgrade cabal-install or their distro.

The other question is how useful test suites in a released package
are. Aren't they much more useful (and used more often) in source
repositories?

They're certainly useful in source repositories, and we have historically chosen not to make a distinction between what's in a source repo and what gets shipped to end users via cabal, which makes sense to me.