The difference is that the interior of your car is not typically an adversarial environment.

The exact moment that you need a seatbelt is the same moment your car's interior becomes an adversarial environment.

"Adversarial" in this sense is meaningfully different than "dangerous" - at no point is your car trying to outsmart you.

You're right, the car isn't trying to outsmart anyone - physics is trying to outsmart engineers.

Nah, physics doesn't need to try :D

That really depends on your threat model. Every effort to reduce your attack surface increases the effort needed for your adveraries. Your "won't help very much" is only true for the trusting trust problems in the paper, i.e. if you need to survive state sponsored attacks