What is unclear is whether the driver is likely to have seen it in time if the car had no autonomous mode. Humans, when paying attention even for long periods of time, are still way better at recognizing objects than computers.

My expectation is that this is exactly the problematic case for stastical vs concrete risk analysis. The automated system as it is today is generally safer than humans, as it's more predictable and reliable. However, there are individual situations where even a less-statistically-safe system like a human forced to pay attention by having limited automation can avoid an accident that the automated system can't.