This was discovered whilst looking at regression suite running prefetch2.ecl on larger clusters. See : https://track.hpccsystems.com/browse/HPCC-21301
This simplified query running on a 16 way thor illustrates the issue:
In Thor the super index parts are read such that all part 1's are read on the 1st slave, all part 2's are read on the 2nd slave, etc.
The result is that 1st sorted result is found, which happens to be from the 2nd subfile of the superkey.
Whereas in Roxie (and in hthor) the 1st match it finds is from a later part of the 1st subfile.
For Thor I think this is the correct behaviour.
For the other engines, perhaps it's undefined.