I think I've come up with a potential fix, though I need to sleep before I will have the tuits to code it.

The mechanism I envisage is to plot the circles larger than scale. This will cause points that are further away than the specified limit to be grouped artificially. You then replot just the points in each group, at the correct scaling, on a clean canvas using alpha-blending, to eliminate the false hits.

The first pass will reduce the problem to a few small, isolated groups making the second pass only need to deal with a few points at a time, so the slowdown shouldn't be significant.

I'll try to post a corrected solution later.