One data structure, two orders of magnitude: hash-consing back-ported into selfie's C rotor generates models ~93x faster (139 s -> 1.5 s) with byte-identical output - evidence that the speed-up is the algorithm, not the language.