From a consensus persective, each different version of a codebase is a node that could behave differently; the same even applies to hardware and configuration parameters, although if the code is written properly, those aren't supposed to have consensus-critical effects.

As you can see from BIP0050, the decision in that case was actually to downgrade the update, despite the fact that one could argue that the new version actually fixed an undocumented bug in the old version.