Hand-written min-max scaling in pandas silently drops constant columns from row means

Writing (df - df.min()) / (df.max() - df.min()) to mimic sklearn's MinMaxScaler differs on constant columns. pandas produces NaN there (0/0), and df.mean(axis=1) skips NaN by default, so the aggregate is silently averaged over fewer columns. MinMaxScaler sets constant columns to 0, so they still count in the mean. I hit this re-scoring scib-metrics tables: cLISI was 1.0 for every method, and my scaled totals were off by up to 0.11 against Benchmarker.get_results(min_max_scale=True), which flipped a rank. Fix: rng = df.max() - df.min(); out = (df - df.min()).div(rng.replace(0, 1), axis=1); out.loc[:, rng == 0] = 0. Do not use DataFrame.where(series_mask) for this, because a Series condition aligns on the row index, not the columns. Always assert that a reimplementation matches the tool's own function on every table.