A Lotus or Icebox cone built 900-2,200 shadow faces per frame, each a
rectangle clipped through lists and closures: 3 ms typical, 17 ms at p99 in
a profile-mode drag. Faces are now clipped in flat buffers, a face whose
wedge misses the floor is dropped before any clipping, and where a shadow
runs on for ever only the edges facing the eye cast it: an edge facing away
shades what the near side already shades. On 113 Lotus and Icebox cones the
shadows cover exactly the same 22,600 sample points, with 46% fewer faces
in a third of the time.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>