I believe you mean well. And you are correct about Waymo. However, you are just not seeing it. It is not remotely possible to get near L4 and you will eventually harm thousands trying to get there relying on public shadow and safety driving, deep learning and gaming based simulation. The solution is to switch to simulation based on DoD tech. Which will result in a legitimate real world physics based all model type digital twin. (Be glad to show you s demo. I know your simulation is not remotely capable of this.) And to use dynamic sense and avoid with targeted deep learning.
Proposal for Successfully Creating an Autonomous Ground or Air Vehicle
Regarding your list of reasons justifying disengagements — I assume you list of reasons for disengagements are not at completed legitimate L4 but in the development and test phase? If so you are making my point that safety driving is reckless and untenable in the public environment. While they will not go to zero you are making excuses for using the wrong approaches. And while monitoring/alarm systems and training help you cannot resolve time critical scenarios. And since you have no idea when they will occur you have to assume they always will. And how do you train and test accident scnerios that cannot be avoided without the safety driver refusing to disengage and in many cases committing suicide?
Regarding geofencing. It’s mostly hype. See my article here