The situation you describe is internally inconsistent. You state:
The machine is capable of running several slightly different simulations simultaneously, the difference being some action that the user of the machine can do.
and
Since the machine can't practically simulate the action of the user (anything can be done in different ways) it actually simulates it's own 'actions' (i.e. printing output)
If the machine can't predict the scientist's actions, then how is it simulating the results of printing that message? The procedure would seem to be as follows:
- Simulate printing the message.
- The scientist could do absolutely anything here and I have no way of predicting what specifically will happen.
- Predict with perfect fidelity all second and higher order effects that will happen as a result of (whatever happened in step 2).
- Make recommendation
You have here a machine that can accurately simulate the next several decades after printing that message, but not the next several minutes.
But even assuming that it can simulate the scientist as well doesn't fix things. If the machine can accurately predict what the scientist can do, then the scientist doesn't have to destroy the machine, since what the machine predicted was that the world in which it printed that message is the ideal one, not necessarily that the world in which it was destroyed is. The scientist does whatever they want to, and either the computer simulated the scientist's reaction correctly, in which case this leads to the good outcome, or it didn't, in which case the simulation is flawed and the scientist shouldn't be making decisions based on it in the first place.