> Some of which is not actually UB because the implementation defines it

No - if something is UB in the spec, it's UB. The implementation will do something, sure, but what it does is not fixed and may even change based on compiler version and optimization level.

> DWORD-sized memory access is atomic on Windows because Microsoft said it is

Well, Intel said it is. Mind you I don't think there are any 32-bit native architectures where aligned dword access isn't atomic. Unaligned, on the other hand ...