The chart shows cond_number (log scale) by degree, one panel per method family - this is solver/dtype-independent, a property of the raw features themselves, and it's what actually causes the solver/dtype sensitivity below: PCC past degree ~4 and RPCC at degree 5 exceed float64's precision entirely (cond_number > ~1e15), so their 'mean' pivot deltas there are largely numerical noise, not a genuine solver comparison - trust the pivot most where cond_number is still comfortably under that line. The pivot's Δ-vs-row mode compares every solver against the lstsq_numpy baseline, one method column at a time; filter agg_level, weights_finite or predict_finite to isolate unstable runs directly.