Introducing a naive "letrec-set" field to the various kinds of "set"
record turned out to be a bad idea. Making the set that a letrec does
to stomp on the recursive binding be a "not-really-mutating" set
causes the backend to assume there's no change, and to emit code that
copies the dummy-init value around. Sure, there's no box, but also
there's no linking of any mutually recursive procedures going on. I'll
think some more about it and revisit it later. It doesn't look like
it'll be easy to avoid the boxing: the alternative is to go into a
bunch of environments in lambdas that are already constructed, and
stomp on their letrec'd bindings. Nasty. Perhaps Kelsey's CPS/SSA
technique will help here - a little analysis to identify mutually
recursive /procedures/ could go a long way.

The Mono "ilasm" and "al" seem crippled compared to the Portable .NET
versions of the same. Newmoon requires the pnet versions to compile
properly. (Update: I take that back. Sure, there are features that are
missing from the Mono versions of the tools, but now I have corrected
the syntax of my generated IL (oops!), the Mono ilasm seems to be
quite happy.)