“Shared, identical context between an evaluator and a generator model is a more crucial factor for causing in-context reward hacking than the length of the context.”