If the AV needs up-to-date maps just to do something basic like lane keeping at L4 reliability, then it is not good L4. You can't have a system that downgrades to L3 every time it encounters a change in the map. That is not going to work.
Without sparking (yet another) debate of the differences of L2, L3, and L4, when it comes to autonomous systems driven primarily by machine learning, I don't think there are necessarily bright lines. In machine learning, everything comes down to confidence levels. A neural network won't tell you that something is, e.g., a dog or a cat. It will tell you that it thinks there is an 92% chance it's a dog, and a 5% chance it's a cat.
A system may be designed and have all functionality to be L3, but if aggregate confidence levels aren't generally high enough, it may behave erratically and be deployed by the manufacturer as an L2 driver assist, as is the case of with Tesla FSDb. Similarly, a system may be designed as an L4 system, but instead of geographic, its limited domain may be environmental - e.g., visibility, reliability of map data, lighting, etc. The system may only operate at the confidence levels to be L4 when the environment is ideal. It that a "good L4?" Maybe not, but I think you would agree that the thing that makes Waymo L4 over Tesla's L2 is not necessarily design goals of the system at this point but availability of data (sensors, maps, GPS, etc.). Tesla having limited sensors and low resolution mapping has hampered it's ability to become a true L3 system. And, IMO, without additional sensors and/or data, it can never be L4. Accordingly, that's why I say V2V and V2I could provide additional data to these types of systems so that their confidence levels, when such data is available, are high enough for these system to operate at L3 (or maybe L4).
Will it happen? I doubt it. But could it add to the confidence levels of the NNs and thus the performance of the system? Absolutely.
We already do somewhat have V2V communication. It comes in the form of traffic data. Routing decisions are being made based on this information. While this data is nowhere near as fine grained as what I believe you're talking about, the concept is the same.
The problem with something like Waze is that there is tremendous latency, so while it may help for routing around problems on a long trip, it can't really add much to autonomous vehicles. But imagine that you are driving down the highway and it starts to rain hard and visibility is reduced. The system may be designed to fallback to L2 in those conditions. However, if the car was in direct communication via V2V with one or more cars in front of it on the same road travelling in similar vectors, that may be enough information for it to have the confidence levels to continue to operate at L3 despite the reduction in visibility.