Thought I'd think out loud for suggestions while I think about how to phrase a believable RMA request.
Machine is an i5-6500 / H170, 16GB(2x8) DDR4, RX 460, blah blah. I have isolated the problem down to either the motherboard or CPU. Changing the PSU or GPU (or yanking the GPU entirely) has no effect, using a completely fresh install of Windows 10 has no effect. Clearing CMOS has no effect. Memtest86 V7.3 has yet to find a problem with RAM (given the how quickly and often the crashing happens, I would expect it to easily find something if RAM was involved). I have no equivalent spare RAM, CPU, or motherboards to swap out for testing. All temperatures seem normal. Problem has manifested last Tuesday after user had it for four months, has been fine up until that point. User was running the machine in a closed cupboard though, so they may have cooked it.
Here's what's happening: shortly after booting into Windows 10, it will crash. Timing is random, but while playing with it crashing didn't seem to take longer than 10 minutes. Using it "normally", it doesn't seem to make a difference what you're doing, eventually it will BSOD, and there are so many different bugcheck codes that it's not even worth mentioning them. It does seem to be able to boot to desktop without any trouble. Before it crashes, it may or may not choke on whatever it's doing (e.g. fails to read or run something). BUT - if I start Prime95 in test mode fast enough, it becomes relatively stable. Not *completely* stable, it will still fumble tasks/writes/whatever every now and then, but it no longer seems to BSOD (Prime95 itself runs mostly OK, though it has detected one rounding error so far). Loading the RX 460 instead (i.e. generating a load on the PSU without loading the CPU much) does not have the same effect, it will BSOD as usual.
You may think, aha, this is a C6/C7 PSU underloading problem. I don't think so - as I said, I have tried two different PSUs that are stated to be "Haswell compatible" by the manufacturers. Motherboard UEFI has an option for a dummy load (irritatingly, not for C6/C7 - I suspect that those C-states are not even enabled at any point).
Does it seem reasonable to think it's a bad motherboard? Perhaps it cooked a VRM?