Memory hardened is not susceptible to bot nets if the memory requirement is adequate. That is the fud gpu farmers push. Higher memory requirement would eliminate that non issue.
Then this would seem to be the way to go, since functions like Equihash are already available for use as drop-in replacements for SHA256. A randomized GPU PoW would be much more work to implement.
Could someone fork Bitcoin-core and make the necessary changes so we can have a bitcoin-equihash-testnet?
Also might be worth considering a fork for sha256+equihash, so we can evaluate one of the popular propositions here (i.e. so as not to totally destroy the friendly miners (BitFury, BTCC, et al.)).