I'm not sure that it does make a lot of sense -- we allow (mutually) recursive functions, even though they come with an efficiency penalty. Why should we not allow (mutually) recursive modules, even though they too come with an efficiency penalty. This is even an example where the efficiency loss is *only* at compile time, and only happens once, so it's somewhat a better situation than allowing mutually recursive functions.
I'd say it falls very heavily into the ghc-bug category, not the spec bug category (even if there's reasons for the bug existing in ghc).
Perhaps it would be better for GHC to allow compilation of cyclic inclusions via a flag? -fcyclic or something? Or, to do it by default unless a -fno-cyclic flag is raised? It does seem strange that the only way to compile cyclic modules is to hack together a build using hi-boot files. Regards, Chris.