I think you should pull the lever, even if this ended after the entire human population was on the track and the experiment doesn't go on infinitely. Hear me out:

When a person pulls the lever with a chance of 50% and in one case they kill 2 people and in the other case 0, the kind of average outcome is 0.5 * 2 + (1 - 0.5) * 0 = 1. Now let's consider the last person in the chain of decision-makers. They would have 2^33 people on the tracks, or about the entire human population. To make the expected outcome be exactly one person, they'd have to pull the lever with likelihood x so that x * 2^33 + (1 - x) * 0 = 1 which would lead to x = 1/2^33 or about x≈0.0000000001. So only if the last person directs the train towards the people with less than this tiny chance, the expected outcome is smaller than 1. This chance is incredibly small, and far far smaller than I'd guess the actual percentage is. Think of the percentage of people that are psychopaths, or mass murderers, or maybe even just clumsy. If you evaluate the percentage as someone flipping that switch as anything above 1/2^33, you should therefore flip the switch yourself. You can guarantee that the outcome is 'only' one death, whereas the average outcome of just the last person likely exceeds 1 by a huge amount.

I really wanted to calculate the percentage so that the expected outcome is 1 even if every person in the chain flips the switch with that chance, but wolfram alphas character limit let me down :(