Which input plane each result face lies on was found by scanning every input plane whose
normal x was within the resolution's angle: quadratic on axis-aligned parts (two thirds of the
faces have normal x = 0), so it was capped at 16 384 faces a body and 2 M comparisons, and the
join of the 20 level-2 sponges of a level-3 Menger sponge (7 326 faces) was refused past the
comparison budget.
The planes are now indexed by their unit normal, in cells twice the bound on two matching
normals' difference (asin(angular) < 2 angular), and in each cell by their offset from a common
centre, searched in the window linear + |a.o - c| |n_b - n_a| that
d_b - d_a = (b.o - a.o) . n_b + (a.o - c) . (n_b - n_a) gives, plus rounding slack. The index
offers a superset; same_support still decides each match, and candidates are taken in the same
order as before, so every answer is unchanged. A plane with a normal not of unit length is
offered to every query. Both caps are removed: results are bounded by inputs capped where they
entered. The descriptor-size budgets on kept roles are unchanged.
Rig, --known-failing, results compared case by case with develop's: booleans (129) and shell
(28) identical except booleans/5126-join-menger-sponge-level-3, which now passes in 38.9 s on
aarch64 and leaves KNOWN_FAILING.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>