Your best bet on removing the "prickly edge" is to artificially interpolate data along the "red-valued" boundary (parabolic, hyperbolic?) to add one (or two) dense perimeters along that edge - if that's appropriate or what you want.
Triangulation-based methods will return some peculiarly shaped triangles in that region from the data you have, that is what it is, it feels counterintuitive and "unfair" but if you zoom in and look at the triangulation you will see what happens (you'll have to do an explicit triangulation with delaunay - likely on a much smaller sub-grid and manually plot the triangulation). To some degree you might change that behaviour somewhat by scaling your x and y-values differently but that's a crutch.
