One reasonable assumption is of course to assume that sleep works as it does on virtually all modern OSs, in which case, you're correct, but that is certainly not the only way it can be. You could easily imagine a specialized SortOS that executes one thread on one process and outputs messages at a specifically scheduled time after a program begins executing, in which case the result would be radix sort with extraneous waiting, rather than priority heap sort with extraneous waiting.
Moreover, that is essentially the whole point of the field of algorithms. Your job as an algorist is to abstract the algorithm away from the implementation details of wholly separate systems, and make all of your assumptions explicit. If you don't specify it, you can't analyze it. And, if an algorithm has to assume an entire, specific OS and scheduler is implemented under it, then you have gone about your job as an algorist completely wrong.
That is the appeal to practicality. But as if to prove my point, you _do_ end up making assumptions. You are assuming that the scheduler is backed by a priority queue, which is certainly not universally true. You don't get to a hard mathematical bound by waving away a detail as important as the data structure that backs the core sort mechanic. If your approach is to say "y'know, like a _normal_ scheduler", then you should stop and rethink your approach to the problem.