Fix a lurking width-inference bug; improve adjacent style (#298)

ceil(log(x) / log(2)) does not, in general, round to ceil(log2(x)).
I noticed this because of #297.