The difference is that attempting to add code to a live system that mishandled an input (either from a non-upgraded node or an upgraded one) would result in a type-error and reject the upgrade.
You could still use Erlang's crashing and supervisor technique, but you would have the additional benefit of using static typing across a distributed system (where each node may or may not have received the upgrade yet).
The problem is whether the function can or cannot handle the content of the new message and that's completely orthogonal