> maybe all this should just be thrown away and we should invent an FS with ACID semantics.

You're describing WinFS, which looked into and ultimately abandoned Microsoft 20 years ago. I'm sure other groups have looked into this as well, but there's no such thing as free lunch.

> I'm all for gradual improvements but at one point and on we should zoom even further out and pick our battles well.

That sounds a lot like picking up more battles, yet we all still have 24 hours a day. Recursively trying to perfect lower layers will have you like Hal changing the lightbulb https://youtu.be/AbSehcT19u0