For the conversion to an Arc compiler, it's similar enough to Scheme that I don't think it'll be a big issue, especially as it's implemented on MzScheme. Quite a few tricky Arc functions are just MzScheme functions under a different name (e.g. threading, I/O stuff). That brings up a point: you really want a compiler for the subset of MzScheme that Arc uses, not Scheme itself (although a Scheme->Arc compiler would be cool too).

Is it too much work for a summer? Depends on what you want to achieve. A completely non-optimising Arc compiler, based on code ripped from other projects, would probably be really easy to implement. Writing a Scheme compiler in one semester is standard coursework on some CS courses. So I guess the important questions are:

1. What good Scheme compilers written in Scheme already exist?

2. How much actually needs to be converted? (Most of the runtime will probably be implemented in C anyway, so no conversion necessary).

3. Given the spread of possible compilers you could convert, do any of them look like a summer's worth of work?