The most noticeable thing about that is that the worst percentage-wise
cases appear near the bottom end of the range.  And indeed inspection
of individual entries showed that trivial cases like

SELECT (ROW(1, 2) < ROW(1, 3)) AS "true"

were hurting the most percentage-wise.  After some study I decided that
the only thing that could explain that was the two rounds of
construct-an-upper-rel-and-add-paths-to-it happening in grouping_planner.
I was able to get rid of one of them by discarding the notion of
UPPERREL_INITIAL altogether, and instead having the code apply the desired
tlist in-place, like this:

        sub_target = make_subplanTargetList(root, tlist,
                                            &groupColIdx);

        /*
         * Forcibly apply that tlist to all the Paths for the scan/join rel.
         *
         * In principle we should re-run set_cheapest() here to identify the
         * cheapest path, but it seems unlikely that adding the same tlist
         * eval costs to all the paths would change that, so we don't bother.
         * Instead, just assume that the cheapest-startup and cheapest-total
         * paths remain so.  (There should be no parameterized paths anymore,
         * so we needn't worry about updating cheapest_parameterized_paths.)
         */
        foreach(lc, current_rel->pathlist)
        {
            Path       *subpath = (Path *) lfirst(lc);
            Path       *path;

            Assert(subpath->param_info == NULL);
            path = apply_projection_to_path(root, current_rel,
                                            subpath, sub_target);
            /* If we had to add a Result, path is different from subpath */
            if (path != subpath)
            {
                lfirst(lc) = path;
                if (subpath == current_rel->cheapest_startup_path)
                    current_rel->cheapest_startup_path = path;
                if (subpath == current_rel->cheapest_total_path)
                    current_rel->cheapest_total_path = path;
            }
        }

With that fixed, the scatter plot looks like: