See for example T43393, T35590 for some tickets reporting the same problem, this has been considered a known limitation so far.

Using the distance over the surface was the intended behavior, and removing it will make connected proportional editing a lot less useful. Maybe there can be an option for this, or if we do decide to remove this then e.g. editmesh_set_connectivity_distance() should be removed entirely as well.

But the real problem is that the way this distance is computed is not exact and influenced too much by the particular topology, and that's what should be fixed. In general the correct solution would be to use Geodesic distances (there are many research papers about this that tend to be quite complex), or there may be some simpler heuristics than work well enough to get a smoother distance here.