You would want as much as possible to make the timeout not based on runtime -- but on an abstract model of the computation which doesn't actually depend on any machine particular. "Hey your program didn't compile on my machine -- try a faster cpu" would be a nightmarish interaction.

You want constraints about how long these programs are allowed to execute to be tight enough for developers to learn a sense for what is reasonable to do with these features -- as opposed to waiting for runtime.