That's the mother of all assumptions. They might get worse as a result of the engineering response. You are assuming that there is a perfect understanding of the problem space. There isn't. This is not a 'mere matter of engineering' at this stage.
The point is that the understanding of the problem space increases with each incident. This doesn't mean that regressions are impossible (they are probably inevitable), but each regression also increases understanding of the problem space.
A death isn't the same as running into a stationary emergency vehicle of course, but AFAIK Tesla hasn't taken any meaningful engineering response to that issue. Assuming they will to other issues seems overly generous.
OTOH, Waymo seems to have only crashed into one bus, assuming it would move over.
https://www.iihs.org/topics/fatality-statistics/detail/state...
Here: https://www.iihs.org/topics/fatality-statistics/detail/state...
Another aspect that is important but easy to miss - deaths caused by an autonomous vehicle are learning experiences - once a couple have happened there will be an engineering response and they will stop happening.
Compare that to ordinary drivers, where each death is just a pointless waste of a life. There is an argument in favour of putting these cars on the road even if they were more dangerous than humans. The faster we push through the painful learning experience the sooner the payoffs of superhuman driving will start being felt in the death statistics.