I think that the CP/AP dichotomy is a good example of how we treat fundamental, but not hard tradeoffs as hard.

E.g. often we have a fundamental tradeoff between latency and throughput, and it's impossible to get 100% in both metrics. However, we can still do very well in both, and that's what matters in practice.

We have the same thing with CAP. You can build a CP system with high availability.

Fundamental tradeoffs in CAP are latency/time for consistency, since real world networks are always partitioned. You are just using a different meaning of availability when talking about high availability, not the one from CAP.

When I speak about Datomic, my most frequent questions are about CAP theorem, which Datomic seemingly subverts to do seemingly impossible things (like aforementioned Spanner): http://www.dustingetz.com/:datomic-cap-theorem/