Note that attackers don't actually have to spend that input if they disrupt the protocol. They can use the same one over and over, even many times simultaneously.
You're missing my point. If you detect them disrupting the protocol the participants blacklist that input and never participate with it again (ideally the protocol would be such that they got proof of the disruption that they could show to others). But since success consumes the input you don't have a problem with long term identity linking. Yes, simultaneous is a possible but even in the worst case a given input could only be used to jam each group of users once. If recovering from a jamming is just a few packet round trip times and signature operations to redo the mix, this isn't so bad.
In any case, I've probably given you enough of my thoughts on this. Go out and code. Practice trumps theory. As I said in my initial post, I suspect that it might be better to just start out with very weak dos protection and then adapt to the actual threats rather than trying to predict all possible threats in advance. Unlike a global consensus system there is no great harm in evolving this technology.