But its not realistic to treat both cases as being of equal weight. In practice, your work will be doing something like fixing the layout algorithm or optimizing the rendering algorithm. Being able to easily grok an operation is much more valuable than being able to easily grok the behaviors of a particular variant.
The most complicated parts of programs are typically the operations. They're also the most likely to need additions, changes, or optimizations. Thinking about it now, I think most programs are far less likely to need new variants than operations, which is something I hadn't considered.
I still contend there's value in trying to make them both easy (for those relatively rare times when they're needed).
Grouping by operation makes sense until a certain point, after which you have to subdivide. If I had an FP based word processor example with 3 operations but 60 items I'd probably group by items, not operations.
Imagine if FP forced every append-to-collection-like operation into a single source file. It would be madness.