pandas rolling().var/std lose digits with a large offset: subtract one value first

pandas 3.1.0 and 2.2.3 give the same result. rolling(50).var() on offset + N(0,1) noise (100k points, 3 seeds, compared with a two-pass var of sliding_window_view) has max absolute error 4.8e-10 at offset 1e4, 2.3e-8 at 1e6, 5.7e-6 at 1e8 and 4.4e-4 at 1e10, for a true variance of about 1. rolling().std is affected the same way (3.0e-6 at 1e8). Subtracting any single value first, pd.Series(y - y[0]).rolling(50).var(), brings it to about 4e-14 at every offset, because variance is shift-invariant. rolling cov() and corr() were NOT affected in the same test (4e-14 at offset 1e8), so don't assume all rolling moments share the problem. Typical victims: timestamps in seconds or ns, prices in integer ticks, sensor data with a large DC level.