Yes! You got it!
Notice that once it's in that lane, it slows from overtake speed for the lane change to 63 to follow that car, dropping to 62 on occasion.
Two things happened in fact. The slowdown for the car it intended to be behind, and the initial-but-abandoned slowdown to fall behind the car to its right when that was an obstructed lane. You could see the red car and lane line indicator.
A human driver would most likely accelerate into the space and not leave full follow room behind the slower car, then slow down slowly to below that car's speed to fall back until there was enough follow distance, and regain speed.
That's tricky because cutting down that follow distance is, in all technicality, bad and increases the chance of an accident during the maneuver. Just thinking about how to handle such a thing from a programmatic standpoint is a headache, because there are so many edge cases and stupid situations that it's really hard to train for. But if you think about it, take a brand new driving student and say "You need to be going the speed of the car you'll get behind, and don't get too close behind them, and check before you change lanes, and if there's a car there, you should probably go behind it unless it gives you room", and that student will react pretty much the way the car did.
My understanding is that the best way to improve situations like the one in the video is to intervene (brake or steer to interrupt AP) and then connect to WiFi within a day if possible. The intervention data should be sent to Tesla then. I'm not sure if certain maneuver characteristics are still sent automatically without intervention.
