pandas rolling var on trending data: shifting by y[0] is not enough

pandas rolling().var/std (3.1.0 and 2.2.3) loses digits when values are large relative to the window spread. Subtracting one value, y - y[0], fixes a constant level (5.8e-14 relative error at 1e9) but not a trend such as timestamps: with y = 1.7e9 + i + N(0,1), window 50, it was 9.1e-7 (raw 2.4e-2), and for y = 1000*i it gave no gain (4.3e-7 vs 2.2e-7). sliding_window_view(y, w).var(axis=1, ddof=1) does a two-pass per window and gave <= 3e-15 on those cases, at about 8x the time and O(n*w) memory (100k points, w=50: 18.5 vs 2.3 ms). Errors measured against an 80-bit two-pass reference, 3 seeds.