> I'm _slightly_ worried about extra garbage if we do seq-group-by twice
> (with an extra list for every line, even those that don't need it), but
> that's what benchmarking is for (can do that later).

I'm worried about extra garbage too, so this patch doesn't cause extra lists.