Sure, that’s an entirely valid execution. But this is about what exact pairs of values (x, y) are observable by each thread. Some are allowed, others are not, by the semantics of atomic loads and stores guaranteed by the CPU. The starting or joining order of the threads doesn’t matter, except insofar that thread starting and joining both synchronize-with the parent thread.

In general, in the presence of hardware parallelism (ie. always since 2007 or so) the very real corner cases are much more involved than "what if there’s an interrupt" and thinking in terms of single-threaded concurrency is not very fruitful in the presence of memory orderings less strict than seq_cst. It’s not about what order things can happen in (because there isn’t an order), it’s principally about how writes are propagated from the cache of one core to that of another.

x86 processors have sort of lulled many programmers of concurrent code into a false sense of safety because almost everything is either unordered or sequentially consistent. But the other now-common architectures aren’t as forgiving.