I would strengthen this to just short of actually requiring tests. They
don't have to fit a specific test runner: a Scheme script that prints
messages when tests fail will suffice. This is 2015, and unit tests
should be a standard part of practice. Exceptions would be when the
sample implementation is trivial and/or not really meant to be used,
just a proof of concept. (My upcoming SRFI on ephemerons will have
a completely portable implementation except it won't actually work
with any GC.)
What should my reference implementation include?
It should implement all the features described in the SRFI document.It should also include automated tests. Having them will help implementors, and that will increase the likelihood that your SRFI will be incorporated in Scheme implementations. It will also help users understand how your SRFI is to be used.However, if the reference implementation is trivial or not really meant to be used, i.e. it is just a proof of concept, it's okay to omit tests. That should be a rare case.
No specific test framework is required, but both SRFI 64 and SRFI 78 are available.