Judging from the code, the main idea is to implement the Arc (or Arc-like) language using the standard techniques used to implement new languages in Racket. This approach will make it easy to use the tools/infrastructure of Racket without any changes (correct highlighting, arrows between a variable and its binding, scope aware renaming of variables, profiler, etc.). The new language compiles via macro expansion to Racket which is then compiled and jitted as normal.
If I recall correct the original implementation of Arc was written as an interpreter - and thus the Racket set of tools are not available.
The main benefit is that the new language will run faster than standard Arc.
As a bonus it will be possible to use module written in Rark from Racket code - and use modules written in Racket in Rark modules.
A more developed variant with these goals, that got discussed a fair bit in the arc forum, is https://github.com/arclanguage/arc-nu. It hasn't received much kicking on wheels by anyone other than its author(s), though, so be prepared for some rough edges if you decide to try it out. In the long term I think it might be much nicer than the original prototype.
If it's helpful, the original implementation of Arc was actually quite similar to this - a bunch of macros that compiled down to MzScheme (which eventually became PLT scheme, which eventually became Racket!).
I found Arc version 3.1 on arclanguage.org. It does compile to (Mz)Scheme - but it does not use the MzScheme macro expander. It reads Arc program as an s-expression. Compiles the program and produces an mzscheme program also represented as an s-expression. The "problem" is that s-expressions doesn't track source location of identifiers. That made it impossible for the old Arc implementation to reuse the MzScheme tools.
Some the libraries were written using Arc-macros that expanded into simpler Arc constructs, but MzScheme macros weren't uses in that process.
Note: PLT Scheme was the names used when MzScheme, DrScheme and a few other tools was distributed together.
If I recall correct the original implementation of Arc was written as an interpreter - and thus the Racket set of tools are not available.
The main benefit is that the new language will run faster than standard Arc.
As a bonus it will be possible to use module written in Rark from Racket code - and use modules written in Racket in Rark modules.