Fascinating behavior. Since this is an edge, would it ever occur in human-written code? Could there be a situation where a computer would end up generating such code?
(The idea was to do DI using HLists, to get something like Spring but that worked at compile time. It worked, but as the program got larger the compile times became too long for it to be practical)
It's a different matter with generated code. For a while I had a program that generated OCaml code (from descriptions of Linux kernel structs). That would result in 100000+ line generated OCaml programs which frequently caused very long compile times (eg 20 minutes and more), with the time being spent in the type inference part of the compiler.
Where it does show up, annoyingly, is in error messages. You actually see this all the time with GCC's STL error messages (C++ templates provide a parameterized type system very similar to Haskell or ML, but without the inference, at least until C++11). STL containers are usually parameterized not only over the key and value, but over the comparator and allocator. Nest a couple of them and you suddenly have a type with several dozen variables. GCC expands the whole type out, without typedefs, when printing error messages, which means that you will sometimes get compile errors that are 10 pages long. Clang fixes this by preserving the typedefs, without which you wouldn't want to write such a type. Haskell has the same problem, except worse because it infers types, but GHC has gone to great pains to format error messages so they are at least readable if verbose.