> ...powered through emulation under a modern CI server...

I have a 486 PC sitting in my living room. For shits and giggles, I've cobbled together a FAT12 boot loader that runs a program directly off a floppy and played around from there.

And even by that little that I played around so far, I managed to run into more than one issue where something would work perfectly fine in Qemu, but not on the real hardware. Bochs appears to be more faithful, but also not 100% exact.

Btw. did you know that Windows 9x has an interesting TLB invalidation bug that apparently went unnoticed for decades and now triggers in KVM on AMD Zen 2 and newer CPUs? (see: https://github.com/JHRobotics/patcher9x)

AFAIK, part of the reason Linux no longer supports i486 is that it made CMPXCHG8B a hard requirement (and also RDTSC). You would need to maintain a completely separate implementation of a bunch of low-level locking primitives. I'm somewhat skeptical how well that will work when your testing relies entirely on emulation.

> ... someone else should do this, of course.

of course ;-)