You've hit the nail on the head. The real difference is input latency.

Tweaking acceleration curves will not help overcome latency of 2-3 frames which is common on Linux compositors (33-50 ms of delay, causing target overshoots).

I'm thinking either sensitivity is lowered as deceleration begins, or deceleration is over-emphasized as you say. I can move really fast then stop completely dead even as I'm sure my by-no-means-steady hands are not that precise and are probably still very slightly moving as I'm "stopped".

This is a perfect description of input latency. You stop moving your finger over the target, but the cursor keeps moving for another 33-50ms, causing you to overshoot.

Very interesting, TIL. That's certainly a simpler explanation