It seems like a good candidate for inclusion as it has a rather minimal dependency footprint.
Although we probably need to wrap the benchmarks into something that doesn't pull in criterion and half of hackage. Maybe we can have a criterion shim that provides the same API (with minimal overhead)?