Right, it's not because of the non-determistic DISTRIBUTE, because all rows originate on slave 1 as it's defined (with the inline dataset)...

It's because the query itself can have different results depending on the number of slaves.
In particular 2 slaves will always give 112839, any other number of slaves will give 11389.

The distribute by "r / (thorlib.nodes()+1)" , on a 2 slave system is always going to place the 1's,2 & 8 on the 1st node and 3,4,9 on the 2nd.
If 1 slave they'll all be on same slave so no issue, if >=3 slaves, values <=slaves will be on 1st node and others will either be grouped on higher slaves, or sequenced over higher slaves.
There would be the similar problem if there was an r value of 12 and slaves was 3.

I'll fix the test so consistent.

Two thor slaves

Haven't seen before, my guess is that both results are 'correct', but the result is non deterministic.
Result 3 - is a global rollup with a DISTRIBUTE in front of it, which means the rows will arrive at the rollup in a non-deterministic order.

How many slaves were in this setup?