Exactly, my previous car (BMW) once saved me in the fog by emergency braking for something I wasn’t able to see yet.

If you were driving at a speed at which the low visibility would have gotten you into into an accident due to some obstable you weren't able to see yet, you were driving too fast. Simple, isn't it?

The whole idea is they should be safer than us at driving. It only takes fog (or a painted wall) to conclude that won’t be achieved with cameras only.

Well, I do still think that cameras could reach "superhuman" levels of safety.
(very dense) Fog makes the cameras useless, A self driving car would have to slow way down / shut itself off. If they are part of a variety of inputs they drop out as well, reducing the available information. How would you handle that then? If that would have to drop out/slow down as much, ~~you gain nothing again~~ /e: my original interpretation is obviously wrong, you get the additional information whenever the environment permits.
And for the painted wall. Cameras should be able to detect that. It's just that Tesla presumably hasn't implemented defenses against active attacks yet.

You had a lot of hands in this paragraph. 😀
I like to keep spares on me.

I’m exceptionally doubtful that the related costs were anywhere near this number.

cost has been developing rapidly. Pretty sure several years ago (about when tesla first started announcing to be ready in a year or two) it was in the tens of thousands. But you're right, more current estimations seem to be more in the range of $500-2000 per unit, and 0-4 units per car.

it’s inconceivable to me that cameras only could ever be as safe as having a variety of inputs.
Well, diverse sensors always reduce the chance of confident misinterpretation.
But they also mean you can't "do one thing, and do it well", as now you have to do 2-4 things (camera, lidar, radar, sonar) well. If one were to get to the point where you have either one really good data-source, or four really shitty ones, it becomes conceivable to me.

From what I remember there is distressingly little oversight for allowing self-driving-cars on the road, as long as the Company is willing to be on the hook for accidents.

It's the only sensible approach. Not just is the notion that "humans use just their eyes too" completely wrong (otherwise how would be able to tell that something is off with the car "with our butt"?), computers are not even remotely close to our understanding and rapid interpretation of the world around us or cooperation beyond of what's pre-programmed, which is necessary to deal with unforeseen circumstances. Cars must offset this somehow, and the simplest way to do so is with vast sensor suites that give them as much information as possible. Of course many humans also utterly fail at cooperation and defensive driving, but that's another problem.