The case where the gap was containing a point that was
lying inside (or on the edge) of a neighbouring polygon
used to fail, as we were using that neighbour's points
to snap to (since it was also at distance 0).
This was leading to flaky wrong results.
