STARK/hash-based proof implementations aren't usually zero-knowledge. Why? "Efficiency" is the usual but somewhat surprising answer: compared to NTTs, re-randomization doesn't seem costly. A colleague raised an interesting alternative explanation: engineering complexity.