This talk about extensibility is fueled by our empirical evidence that any serious system inevitably runs into the problem of modifying data types while retaining compatibility with deployed software (hardware).
You can see this discussed seriously in-depth in UAVCAN v1.0 and ArduPilot.
At the same time, we need to manage the complexity, which is why we are not actually talking about arbitrary modification of types but we limit ourselves to appendability only. Complexity is also the reason why I am hesitant to seriously consider the introduction of different binary representations, especially within the same transport (Classic CAN and CAN FD are the same in the eye of UAVCAN).
Having seriously analyzed the alternatives discussed here so far, I am inclined to believe that the approach described in my latest post here seems to be the closest to the optimal balance between the complexity and flexibility.