I guess they must be using the lock in a careful way to avoid any contention, or to distribute people who need the lock across many versions of a data structure.

That's wouldn't be the same as a 'simple mutex', would it? Which is absolutely not wait-free (even if I don't understand how the OP is using locks.)

Your original criticism was that a process would crash while holding a lock.

I guess we need to wait for the arxiv to come back and read the paper for the details.

Yes - a lock can redirect people trying to acquire it, rather than queueing them. So processes need a lock to access a resource, but there are enough resources for everyone to get their own without waiting, and enough that if a process crashes holding one it doesn't cause anyone else to wait. That's not a 'simple mutex' though.