Interesting, it appears that prefix-intersection merging is not associative.
That is, given 3 linearizations A, B, C of a given cluster it is possible that not all three of these are equally good:
- merge(merge(A, B), C)
- merge(merge(A, C), B)
- merge(merge(B, C), A)
This holds even when the inputs, intermediary results, or outputs (or any combination thereof) are post-processed.