Ok, with that counterexample in place I’m now again leaning the other direction: that the right approach is using the simplest algorithm that works.
My impression is that the non-commutativity and non-associativity of merging are effectively randomly stumbled upon through accidentally having/putting transactions in the right order. And these accidents can’t be prevented, and moreover trying more subsets will inevitably mean there is a small chance of finding something better still. However, that doesn’t mean it’s worth spending time on even if it’s a small cost. That time could be better spent on directly trying more things in the linearization algorithm.