On the speed detection: TSR is not just improving the accuracy of the speed detection. How do I know? The Tesla read a sign on the side of the freeway, and thought it was the speed limit for the freeway. It wasn't. It was the speed limit for the frontage road next to the freeway, and I was about 30 mph over that speed. So whatever they're doing, they're not just doing a tree search weighted with the currently known speed.
But yes, that's 101-level.
Fair enough that the hard issue is the amount of time the lane analysis needs to be right. Here's something though that doesn't make sense to me in your problem description: why in god's name would you ever consider a discrete change in assessment of where the lane is to be valid? Specifically, why would you perform an action based on a sudden change in where the lanes are? Granted, I'm making an assumption about continuous lanes, but the scenario you're describing is that the car is humming along in the right lane just fine and dandy, and then, due to bug/memory corruption/light glare off of the cameras, the algorithm thinks it needs to move exactly one lane over. Yes, it's hard to get the TP rate up to 5 nines, but then again, the decision process should never be such that a jump in lane condition results in an immediate action.
There's a separate problem with line markings disappearing, but I contend that that's a problem even human drivers struggle with.
Finally, I don't think that I can solve this - I've seen it solved: by Google and Audi, specifically. Now, how "solved" this is? Google has a few 100k miles under the belt of its autonomous cars, Audi quite a bit less. But both have navigated in traffic, with passengers inside. My incredulity doesn't come from me having solved it, it comes from having used other people's work 20 years ago to answer the question "what am I looking at", and having seen it progress from edge detection in a jpg to cars driving on their own in a fenced-off terrain 10 years ago, and now have seen driverless cars on the road. Doing what you're describing as not solved.
Maybe you should talk to the Google engineers. Or the Audi ones. Or the Tesla ones. I hear Tesla opened up all its patents anyway. Maybe they opened up how they do their lane detection and decision process.