45% of validators sign on fork A
  45% of validators sign on fork B
  10% of validators sign both A and B

1) to validate, instead of just sitting doing nothing

2) to choose the chain with the most likely chance of being the next main, otherwise they will be slashed

As soon as even the slightiest imbalance appears between the two forks, it will bias validators to choose one against the other and thus exponentially make it the preferred fork.

I think this is actually the most commonly overlooked assumption. The protocol seems very robust, *assuming* perfect information (aka all incentives are known to the system).

I encountered this problem a few years back when trying to design a protocol for a decentralized prediction market. It's very hard to account for hedges or huge bets on other markets.

These would have to have different network or chain IDs to avoid slashing events. One of them, presumably B, would have had to actually change this number from whatever the previously standard/agreed-upon chain ID is.

I could imagine this scenario happening in a contentious split or attack, but it would create two very clearly different blockchains with different goals and needs, like what happened with Eth and EthClassic, or Eth and EthPoW.