Cairo linear-gradient endpoints were placed at the rounded rect's
geometric corners (3,3) and (21,21), but the *visible* corner pixel
of a rounded rect with radius 2 is the outermost point of the arc,
inset along the diagonal by r * (1 - 1/sqrt(2)) ≈ 0.586 units.
That meant t=0 and t=1 of the gradient landed in the cut-off corner
regions, and the rendered corners sampled at t≈0.033 / 0.967 — about
3.3% in from each endpoint, ~8 RGB units short of the true previous /
active colors (visible on saturated pairs like pure red → pure blue:
the "pure red" corner rendered as ~rgb(194, 10, 0)).
Shift the gradient endpoints inward by the arc inset so t=0 maps to
the visible TL corner pixel and t=1 to the visible BR pixel. The
gradient now spans the visually rendered area exactly; saturated
endpoints render flush.
337737b59f