Red Hat's Enterprise CMS actually takes your third strategy to an almost absurd extreme. it's probably complete overkill for most of us, but kinda cool. :)
i went to a talk that Rafael Schloming gave on "Versioning Structured Content" where he explained Red Hat's approach. the idea is to basically keep a log of every single event at a very fine level of granularity and to ensure that every low level operation is invertible (ie, for every add, there is a corresponding delete, etc.). then, to revert to an older version, or another branch, you just construct a stream of events that will get you there. this is basically the same as how RDBMSes handle transactions internally. the only difference is that when a transaction is rolled back in an RDBMS, it can dump the log info. in a CMS, you potentially have to store everything as branches.
his slides are online.