This is wrong. This DFSM does not halt: node with arrow pointing back to the node on empty input.
I think you also need to have at least one transition for every symbol in the alphabet from every state for it to be a valid ε-NFA, and then it can be reduced to a DFSM, giving it a bounded running time.
DFAs can run in bounded time, because every symbol takes exactly one step to be processed, and there is only one way to process a string.
That also means NFAs can run in bounded time, by reducing them to DFAs first, but that's not a very satisfying explanation.
A NFA rejects a string if it can not end in an accepting state. It's not intuitively straightforward to see whether that's the case, but it can be done in bounded time by exhaustively checking possible paths (which is possible with a finite number of states and a finite input string).
You can't always perform that trick with Turing machines, because they can have infinitely many reachable states.
If they're simple enough, they can be simulated by a brain, so a brain has more computational capacity. If you can make them as complex as you want, then they can match (and exceed) the computational capacity of the brain, just by incorporating all states a brain can have and pre-computing the transitions. For practical purposes it's almost always best to treat brains (and other computers, for that matter) as Turing-complete, but I think that strictly speaking they're closer to DFSMs.
But the point is that almost all predictions about people's decisions involve time bounds. When there's a time bound, those decisions become computable. "Will this program return 1" cannot be solved in general, but "Will this program return 1 within the next hour" can be.