I too have noticed a slow decline in some features that actually worked pretty well. It appears that Tesla has split the production base into two very different sw designs. FSD is all wrapped up in neural nets with a closed loop feedback to the Mother Ship. This is nice except that there no regression testing prior to push. They seem to do a quick check, not an extensive set of checks. The reports suggest two steps forward, one step back.
The original track seems to bumble along adding cute features (color cars) but does not progress. In fact, I see a two steps back, zero steps forward phenom. Could this be managed by the B team? Hope not.
I think a better choice would be to have two levels of FSD. The lower feature level would operate as Autopilot does now. The upper level is FSD Beta. This would allow each car to be a data collector so that even the Autopilot capable base contributes to improvement. Now you would have one code base with the A team on it. Simply stated two feature levels, one sw design/team.
Right now, I doubt Autopilot standard has any feedback, nor does it operate a learning mode process. If Tesla cannot do the above, at least operate FSD in learning mode on all the cars so that they benefit from a fleet experience.
OK, enough.
SR