Well that explains the thinking behind the test, but I’m still not sure it matches the Spec.
The Spec makes it clear that the various collapse rules “Activate cite grouping and collapsing”. So you don’t collapse without grouping. And grouping, doesn’t depend on prior sorting, it is a sort of limited kind of sorting.
With grouping
With cite grouping, cites in in-text citations with identical rendered names are grouped together, e.g. the year-sorted “(Doe 1999; Smith 2002; Doe 2006; Doe et al. 2007)” becomes “(Doe 1999; Doe 2006; Smith 2002; Doe et al. 2007)”
So I guess … I’m digging my heels. I think my behaviour is per spec 
Frank’s is perfectly sensible in its way, if anything is sensible about collapsing (!), but it ignores the grouping step which is implicit in any style that sets collapsing.
ETA: Frank’s addendum suggests grouping only applies when each author’s works are from different years. But the spec contains no such limitation. It simply depends on the output of the first rendered name being the same.