> identify the salient behavior that is hardware-specific and needs to remain invariant under refactoring, and write unit tests for just that?

I don't think you really can. There will be things that can be different between runs (e.g., a DMA transfer) and not necessarily mean bugs. I don't think there are many fully stateless drivers where that kind of effort would both cover a significant part of hardware behavior and yield good enough code coverage.

Can you record the register writes, DMA transfers of a successful run, then run the kernel in an emulated environment and ensure it makes the same writes, and presents the same results to userspace given the same data returned back? This is probably a lot more practical for a mouse than a high-bandwidth webcam or display where the "expected" logs can run into the megabytes or gigabytes for a few seconds of recording.