Module broadphase 

Broadphase — a hand-rolled f64 median-split AABB BVH over one operand’s triangles, queried by the other operand’s triangle AABBs to replace the O(|A|·|B|) all-pairs scan.

The BVH is a conservative FILTER only: it returns the set of AABB-overlapping triangle indices (the exact pairs the old all-pairs+bbox loop produced). The float SAH/centroid sort never decides topology — callers canonicalise the candidate pairs by exact (i,j) keys before processing, so the arrangement topology (and the pinned determinism manifests) are byte-identical.

What the tests module’s “brute-force” checks can and cannot catch (#2830)

bvh_candidate_pairs_match_brute_force and ray_and_point_candidates_are_conservative_supersets read as independent oracles, but their reference computations call the SAME predicate functions the production paths call: Bvh::query and the tests’ own brute() both call overlap() (query at line ~175, brute at line ~262); Bvh::ray_candidates / Bvh::point_candidates and the superset test both call seg_hits_aabb() / aabb_contains(). These tests therefore validate the TREE — median-split construction, node-bounds unions, pruning/traversal — never dropping a candidate the shared predicate would accept; they cannot validate the predicates (overlap, seg_hits_aabb, aabb_contains) themselves. A sign or inequality-strictness bug in overlap() shifts both the BVH’s own candidate set and the “brute-force” reference identically and both tests stay green (mutation-confirmed: changing overlap()’s <= to < leaves both tests passing). Predicate correctness needs its own fixture, derived independently of these functions.

Structs

Bvh

Functions

candidate_pairs

All AABB-overlapping (i, j) pairs between a and b, sorted by (i, j) — a drop-in for the all-pairs+bbox loop with identical output order.

tri_aabb