Hey Butter,
It's a weird case. I've driven through there on AP (both 1 and 2) and it always seemed pretty solid to me. The lane markers are white-on-white through that whole section down there, which is generally not a good thing, but aside from early AP1 I never really noticed any lane holding issues there. I mean - the markings aren't terrible, they're just not great. I've driven in it pouring rain with lots of vehicle spray and straight into the morning sun and on flat cloudy days and it's never been an issue. It must be just about the most heavily travelled AP route in the world, so you'd expect them to have at least this one working.
I haven't tried it on 2018.10.4 yet. Here's a photo from a 'bad' day a couple of weeks ago at about 10:45am. It's a bit before the section where the accident occurred (maybe a mile or two?).
View attachment 290810
So that was on 2018.6.1 and the lanes look pretty terrible under those conditions but the lane holding was really good. This photo is at low speed but traffic was intermittently going 75 and it was no problem. And on a good day the marking are easy to see (it's the rain that ruins them). So it seems odd to me that lane marking on the day of the accident, which was clear, were bad enough to be a big contribution to the problem. Of course glare could be a lot worse at 9:30am but I've driven it in full on glare and AP hasn't been particularly wonky.
Because I've seen people trying to swap lanes there at the last minute I initially assumed that the driver tried one of those maneuvers and got tripped up somehow. Traffic at that location at 9:30 am is a nasty combination of fast and dense and everyone is in a rush at that time of day so if you were to try to swap lanes there at the last minute you could find your target lane closed off before you got into it. As an Apple engineer on his morning commute he probably wanted to go left (towards Apple) but his car ended up on the right side of the barrier. Initially I thought that meant that he was in a different lane from the one he wanted to be in and was trying to change. But that must all be wrong since he was on AP and didn't make any steering inputs in the seconds prior to the accident.
There's the theory that AP incorrectly locked onto the wrong side of the lane split and ended up driving straight down the split. It nicely explains how the car would hit the barrier head on but slightly to the right since the barrier is towards the left of the split zone. But I've never, personally, seen that flavor of screw up and I've seen so many screwups that I sort of feel like I've seen them all - many times. Every time I've seen a lane split failure it was because AP changed it's mind about which side it should be in after the start of the split. It doesn't end up riding the split it just changes over to the other side. This particular split is really, really long so that kind of fail wouldn't have caused an accident here, but being so long it might look more like a lane once you get down it a ways. But if AP isn't fooled at the start then being fooled later and moving over is going to be a lane change. This isn't a curve - it's dead straight there. I've never seen failure to hold a lane on a dead straight road with marked lanes.
But maybe 2018.10.4 has a new failure mode that I haven't seen yet. If I still have 2018.10.4 the next time I drive through there I'm definitely going to try it on that exact path. Another thing that is weird about this theory is that, if you're paying the least bit of attention at all, you'd notice that all the cars around you are starting to go one way and your car is going the other way. Peripheral vision reflexes make it hard to ignore that phenomenon. I wish I could say that the split zone feels different there but so many people drive across it that it's just as clean and smooth as the road. And of course all of the bots dots are worn off and there's no road boundary vibration imprint on the pavement either.
I don't drive with the distance set to 1 so there might be some behavior that happens with close following that I wouldn't have seen. (I never use anything less than 4 and usually stick with 7.) The FOV of the main and narrow front cameras is not all that wide and it could be occluded by a vehicle ahead if it is both large and close, so maybe being on 1 could lead to a situation where the lane markings were blocked by the car ahead? Traffic often slows at that location, so maybe the following distance between the X and the vehicle ahead would get even closer, but that's not consistent with hitting the barrier at 70mph. It's also not consistent with the claim that the vehicle had a clear view of the barrier for several seconds.
I have a hard time seeing how a driver who's familiar with the route and the vehicle and paying even a modicum of attention is not going to notice something weird going on for several seconds in that location. It's not the kind of casual drive where your attention wanders because nothing is going on. The vehicle is going to be accelerating and braking to deal with traffic and would have just finished a gradual turn to the left. It seems like the wrong time of day to be falling asleep, and the wrong situation to be intoxicated. Heart attack? Stroke? That seems like too big a coincidence when you combine it with an AP failure. Everything seems unlikely. But then accidents are rare so maybe the unlikely is to be expected.
What seems to be left is something going on in the car that took the drivers complete attention for many seconds and which was engrossing enough that he didn't notice the visual or audio warnings.
I look forward to seeing the results after the investigation is completed.