I'd like to compute the identity function. Let me denote by 1 the presence of a rock, and 0 the absence of a rock. I'll set up the problem (the input) by placing a rock appropriately, and after ten seconds, I'll check to see the if a rock is present. If so, the output is 1, otherwise it is 0.
This is a theory, a valid interpretation, highly reliable, scales well, and it commutes. Why isn't it a computation by their definition?
Now it is a computation because you are the computing entity that encoded a problem into the system and later decoded the answer. They paper essentially establishes the distinction of a system just evolving under the laws of physics and using this evolution to compute something by assigning an interpretation to the state of the system. I could use the same stone to compute something different, for example the boring trajectory of a stone placed on a surface in a gravitational field, but without specifying what you are computing the stone is not computing anything.
It does not fit their definition because the whole point of the paper is that a physical system is not computing anything unless an entity is using the system to compute something by encoding the problem into the state of the system and later decoding the answer from the state of the system.