> (3) is an instance of seeing "a = b and b = c" and deducing "a = c" from > that. (4) does not follow that pattern, so it's outside the realm of what the > planner can deduce.
I see, that makes sense. I assumed there was something more complex going on in the background. What about converting (1) into (2)? I know the planner does something kind of similar, in converting SELECT * FROM (SELECT * FROM x WHERE a) WHERE b into SELECT * FROM (SELECT * FROM x WHERE a AND b) I guess in this case it would have to know about unique indexes to prove that if the primary keys are equal, all the other columns are too. My intention in trying (3) was to take that burden of proof off the planner.
