GiST rtree logic is not right

It occurred to me to wonder whether contrib/rtree_gist fixes the rtree
bug documented here:
http://archives.postgresql.org/pgsql-general/2004-03/msg01143.php

The answer is unfortunately "no". In the regression database,
install rtree_gist and do:

regression=# create table gist_emp4000 as select * from slow_emp4000;
SELECT
regression=# create index grect2ind ON gist_emp4000 USING gist (home_base);
CREATE INDEX
regression=# select count(*) from gist_emp4000 where home_base << '(35565,5404),(35546,5360)';
count
-------
2144
(1 row)

The correct answer is

regression=# select count(*) from slow_emp4000 where home_base << '(35565,5404),(35546,5360)';
count
-------
2214
(1 row)

Now this is noticeably better than the rtree implementation, which finds
only 1363 rows, but broken is still broken :-(

This is doubtless not as high priority as the concurrency stuff you
are working on, but it'd be good to fix anyway. I was thinking of
proposing that we move rtree_gist into the core --- but the case for
it would be stronger if it worked correctly ...

regards, tom lane