Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors