It seems the linker is playing tricks on the full bot:
Sometimes it works by cleaning the stage2 dir completely, but there
may be some extra bits that we're not considering. However, I can't
see why any of the commits on the first failure did anything to the
relocations on AArhc64...