Oh. So this is indeed a case of "If you're running TDT, I screw you, otherwise you walk free". The way I understand this, TDT one boxes because it should.

If TDT cannot perceive the difference between the original problem and the simulation, this is because it is actually the same problem. For all it knows, it could be in the simulation (the simulation argument would say it is). There is an infinite regress, solved by the fact that all agents in all simulation level will have taken the same decision, because they ran the same decision algorithm. If they all 2-box, they get $1000, while if they all 1-box, they get the million (and no more).

Now, if you replace "an agent running TDT" by "you" (like, a fork of you started from a snapshot taken 3 seconds ago), the correct answer is always to 1-box, because then the problem is equivalent to the actual Newcomb's problem.