This improves performance on the test.yaml attached to #158 by about
25% on my machine as compared to the previous commit (0.25s -> 0.20s),
as measured by `time build/util/parse < test.yaml > /dev/null`.