Bits or permutations? Spending a fixed inference budget on weight precision or on option-order averaging in small local language models