> The combination of st_dev and st_ino from stat() should be unique on a machine

It should, but it seems no longer to be the case. I believe there was an attempt to get a sysctl flag in to force the kernel to return the same inode for all files to see what breaks.