Tailoring fusion-based error correction for high thresholds to biased fusion failures