I did a hot-loop performance cleanup run, because the hot loops were getting slower with new features, and I found a big optimization opportunity in clip-near call sites:

clip_near slices off the part of a triangle that pokes behind the camera’s near plane, but almost every triangle is fully in front of it with nothing to clip. The pipeline still called it for all of them, and in the compiled binary that common case just copied all three vertices to memory and back, millions of times a frame. The fix was to check inline whether the triangle is fully in front and skip the call entirely if so. Result on a 2.77M-triangle model: median frame time dropped from 21.1 ms to 19.2 ms, about 8.7% faster.