To clarify, the 1% window was a heuristic that we recommended in the first paper a a good default for most workloads. However we showed that sometimes a larger window was needed, and concluded that as future work.

The follow up paper, which you also cite, explored adaptive techniques to resolve this deficiency. From there we introduced the idea of using hill climbing by monitoring the hit rate change over a sample period and adjusting the step direction. Caffeine incorporates this with a step size decay to allow for convergence. The implementation that you have does not include this improvement, but it is very simple to add and works surprisingly well.