Personally I've been interested in replacing the GC part of maru with a structural method of managing memory (http://concurrency.ch/Content/publications/Blaeser_ETH_Diss_...), this would require extending the language with the Composita primitives, but not really sure how this'll go yet.
The Lumen-C compiler sounds good! Yes the tricky part is the self compilation, I'm wondering if you'll just end up with a maru-like system if you follow this down?
Alternatively the other way of doing this is via an explicitly generated runtime as-per Ferret (https://github.com/nakkaya/ferret)?
decisions decisions :)