I would err on the side of caution and say 2tps myself.
Maybe you should err on the side of doing the actual fucking math instead.
With 1mb blocks you can do roughly 7.76 tps when a transaction spends one output to two adresses: the payee, and the change address.
nobody wants to have to hard fork every year
That's stupid, because that's not what the guy was suggesting. His idea can be implemented in a single hard-fork implementing a *rule* not yet another hardcoded magic number.