Yes, PG implements multi-versioning, and the semantics are very similar to Oracle's though the implementation is very, very different.
It was the introduction of multi-versioning in PG 6.5 and the removal of fsynching the transaction log on read-only transactions that in combination convinced me that PG was usable in reasonably busy web/db environments (I'd rejected 6.4 and earlier versions as being unusable in practice).
As far as what was "wrong" with your statement(s) I think you misunderstood the meaning of Sebestiano's statement. The cursor query is indeed executed when you open it, not when you fetch it, and your timings showing where the "work" is done doesn't counter that fact.
This isn't necessarily a high-cost step in Oracle because it is essentially just marking a transaction ID of some sort that is used to figure out what data to return when the fetch is executed. I don't know exactly how Oracle does this internally (i.e. the actual form the transaction ID takes) but conceptually it is easy to understand.
