This was PG 9.1 BTW.
Let me also add that I tried recasting one of the offending queries several ways (in (SELECT ... IN (SELECT..., joins, join on subquery in FROM clause, etc) and got the same plan each time (except the nested IN SELECT forms lead to semi joins rather than full joins). That's good in that the planner's able to build a canonical form from each of my twisted efforts to fool it that's equivalent to the deeper meaning of what I was trying to do. It's bad in that each plan had a sequential scan of (gulp) 69 million rows!!!!
