> plain old L3 latency can be worse than misprediction.

true but the instruction fetcher is driven by the branch predictor, so a misprediction to rarely executed code in practice also implies a cache miss.