ibu-boost: A GBDT Library Where Splits Are Absolutely Rejected, Not Just Relatively Ranked

I built a small gradient-boosted decision tree (GBDT) library based on the screening transform from "Screening Is Enough" (Nakanishi 2026). The paper was originally written for Transformers, but the core idea—replacing relative comparison with absolute-threshold rejection—maps naturally onto GBDT split selection. This is an independent implementation applying the screening concept to improve how trees choose their splits. Instead of just ranking candidate splits from best to worst, the library uses absolute thresholds to outright reject poor splits, potentially making the algorithm faster and more efficient.