I've changed my mind on this a few times after fixing a broken algorithm implementation, but I'm starting to think a two pass approach to polygonization - fixed-step De Casteljau and then line simplification (using Visvalingam-Whyatt) - is almost always better for curve quality compared to adaptive subdivision. adaptive subdivision (in three-text's case the algorithm from Anti-Grain Geometry) makes locally optimal decisions per curve segment, but the oversample-then-simplify approach distributes points more evenly across the whole contour. and the cost to compute either is trivial. but it's possible I'll change my mind again, maybe when multiple glyphs with very different curvature are involved...