That's actually a really good idea and something that could easily be implemented. Who knows, maybe that will be on a future software update.
The problem is so well-known they must have plenty of data for it already --
there are so many Tesla employee/drivers that any data campaigns can just involve them.
I personally haven't noticed it so much since the advent of "Tesla Vision". Although
I drive the latest FSDbeta now, I actually expect it more for city driving; Tesla annotates it
as "false slowdowns" in the release notes. Triggers: over-sensitivity to pedestrian intent
(but "VRU kinematics" improves with every release), changing execution plans in
the middle of a turn or an intersection due to planner bugs, etc.
One unnerving possibility is that for certain complex situations, compute power just plain
runs out (at least for combinations of any given software release and HW3), and the software
temporarily freezes/hangs with a petit-mal seizure until the next appropriate hard-realtime
interval obtains.
Now didn't that all sound like I know what I'm talking about? Ha, I are just a retired engineer!
Seriously, if someone ever can ask this at a cocktail party to a sufficiently high-up engineer
at Tesla, they may just spill the reason.