Ok, what would you have me benchmark? Why don't you do it yourself?
If you need access to a complete hash, it is completely clear that returning a reference to a hash which has been created already is much more efficient than causing the hash to be re-created from a list. If you claim that is not true, then you need to provide some benchmarks yourself! That duck just doesn't fly.
I guess we are talking about returning ONLY the keys from a hash table, that which is created in a sub? Or what? The sub can make a list of those keys itself or it can pass a ref to its internal hash for the calling sub to do the same. That computational effort will be similar.
What are you talking about? What kind of benchmark do you propose?