tests/AABB: de-flake dynamic-vs-brute-force timing assertion - #2560

Problem

The AABB: dynamic test (tests/include/igl/AABB.cpp) gates a speed comparison on F.rows() > 300:

if(F.rows() > 300) {
  REQUIRE(t_dynamic_tree < t_static_tree);
  REQUIRE(t_dynamic_tree < t_brute_force);
}

Each t_* is a single-point query (one barycenter) averaged over MAX_RUNS = 10 runs — only a few microseconds. On small/medium meshes the tree's O(log n) advantage over the O(n) brute-force scan is within timing noise (effectively a wash around ~1e3 faces), so on shared CI runners the t_dynamic_tree < t_brute_force assertion flakes.

Observed on an unrelated PR's CI (ubuntu-24.04 Static tutorial tests):

tests/include/igl/AABB.cpp:229: FAILED:
  REQUIRE( t_dynamic_tree < t_brute_force )
with expansion:  0.0000473499 < 0.0000459433   ← lost by ~3%
with message:    decimated-knight.obj          ← 1000 faces

The meshes that hit this comparison are {decimated-knight.obj 1000, decimated-knight.mesh 3544, truck.obj 4770, bunny.off 69451, elephant.off 157160}. The tree only robustly beats brute force on the last two (~100×); the smaller ones are noise-dominated.

Fix

Raise the comparison threshold from 300 to 10000 faces, so the timing assertions only run where the asymptotic advantage clearly dominates measurement noise (bunny + elephant). The dynamic tree's correctness (indices, distances, structure) is still verified on every mesh — only the timing comparison is gated.

Passed 5/5 local runs (1,170,443 assertions each).

Found while investigating a flaky CI failure on #2558 (the flake is pre-existing and unrelated to that PR).

🤖 Generated with Claude Code