Logged In: YES 
user_id=764593

Any hash will have some collisions.  If there is going to be 
predictably bad data, this is probably a good place to have it.

The obvious alternatives are a more complicated hash (slows 
everything down), a different hash for embedded tuples (bad, 
since hash can't be cached then) or ignoring some elements 
when determining the hash (bad in the normal case of different 
data).

I would also expect your workaround of data rearrangement to 
be sensible almost any time (X, (X, Y)) is really a common 
case.  (The intuitive meaning for me is "X - then map X to Y", 
which could be done as (X, Y) or at least (X, (None, Y)), or 
perhaps d[X]=(X,Y).)