It's not a (provable, guaranteed) CP system. Bitcoin attempts to provide a global consensus across a large number of actors, but its attempt is based on proof of work, specifically that it's probably difficult to compute hashes with a certain property (leading zeros, last I checked). It in no way guarantees consistency in the face of partitions.

Consider the simplest case: 1/2 of bitcoin users/miners are temporarily split off from the other half, for say a week (all atlantic/pacific fibers are broken at once, all satellites fall out of the sky, other huge catastrophes occur all simultaneously). Each half would append their own blocks to the blockchain happily, and depending on the chain lengths added, once the partition went away there would be no good way to reconcile. Thus: not consistent in the face of partitions.

> Each half would append their own blocks to the blockchain happily, and depending on the chain lengths added, once the partition went away there would be no good way to reconcile.

Why do you say that? The chain with the greatest total difficulty wins.

Which means that all the people who wrote to successfully and relied on the other chain over the last week suddenly lose all their transactions. That's not what anyone means by CP.

That is not an acceptable form of reconciliation in a CP system.

Bitcoin is AP not CP.

The hard problem is consistency, in terms of tradeoffs and implementation. Pure AP is easy.

What if both happen to have the same total difficulty? Granted, it is unlikely, but so is unlikely a failure of non-paxos system with lease timeouts on the order of ~10 minutes.

See, for example, Garay et. al: https://eprint.iacr.org/2014/765.pdf