The overriding principle is "get maximally close to equivalent properties to a single chain". This includes optimistic confirmation via fork choice rules and traditional BFT finality through Casper-FFG cycles.
A 50/50 network partition would likely have the same consequences as one on a single chain: nothing gets finalized (until either the network comes back together or the leak mechanism kicks in), but unfinalized chains on both sides keep growing.
In Ethereum sharding, there's no notion of shards being mapped to specific long-term node clusters or geographies etc; it's not like there's a Shanghai shard, Toronto shard, etc. All shards take their (frequently reshuffled) validator sets from globally dispersed nodes.
