Since most production functions are quasiconcave over inputs, negative selection is a cheap method of increasing expected return. You lose some outliers and also people who would be good in those rare domains with quasiconvex production functions, but our system is optimized for the average case.

In the college admissions example, a top school wants to admit undergraduates likely to become successful doctors/lawyers/businesspeople and alumni donors, not gamble that the smart kid with a few Bs in high school is going to revolutionize a scientific field in 15 years. Even if they did, their undergraduate institution would be only the third most important on their CV, after their current institution and where they got their PhD.

This is a good example of individual selection being suboptimal from a group perspective. Each top school would prefer that some other top school gamble on said smart kid, and then if they have the chops for research they can try to grab them when they apply to graduate school or go on the academic job market. Positive selection on undergraduates is just not a smart strategy from an individual institution's perspective since most undergraduates will be going into more conventional fields.

Depending on the payoff scale, a TDT agent will cooperate if it believes that the other agent has some (high enough) chance of being a TDT agent. In other words, raise the sanity waterline high enough, and TDT cooperates.

TDT / superrationality will defect probabilistically given a high enough payoff for defection, even against a known-TDT agent.

In short: TDT and superrationality theories aren't as simple as some here make them out to be, and the one-shot prisoner's dilemma has hidden depths for smart players.

The problem goes away if you allow a finite present value for immortality. In other words, there should be a probability level P(T) s.t. I am indifferent between living T periods with probability 1, and living infinitely with probability P(T). If immortality is infinitely valued, then you run into all sorts of ugly reducto ad absurdum arguments along the lines of the one outlined in your post.

In economics, we often represent expected utility as a discounted stream of future flow utilities. i.e.

V = Sum (B^t)(U_t)

In order for V to converge, we need B to be less than zero, and U_t to grow at rate less than 1/B for all t > T for some T. If a person with such a utility function were offered the deal described in the post above, they would at some point stop accepting. If you offered a better deal, they would accept for a while, but then stop again. If you continued this process, you would converge to the immortality case, and survival probability would converge to P.

Of course, this particular form is merely illustrative. Any utility function that assigns a finite value to immortality will lead to the same result.