[antonsoo]

Integer vertices keep their dtype during edge subtraction and cross products in face_areas. Products can overflow and unsigned arithmetic can wrap, giving wrong face areas and, through vertex_areas, wrong vertex masses. For example, an int16 right triangle with 300-unit edges returns 12,232 instead of 45,000.

Widen integer coordinates before subtraction, then convert the resulting edges to floating point before taking cross products. The 64-bit case uses Python integers to retain small edges even at a large integer origin. Float inputs keep the existing arithmetic, and inputs are not mutated.

Add two regression methods covering all eight integer dtypes, both windings, float controls and translated 64-bit coordinates. All 13 project tests pass on Python 3.12/NumPy 2.5 with locally built Debug bindings and on Python 3.9/NumPy 1.26 with released 1.4.0 bindings and the current Python source. The patch changes no C++ code.

An additional audit of the included Bunny scan, using an independent Heron reference with 70-digit arithmetic, restores 24 incorrect area, mass and integration configurations. All 16 previously correct configurations remain byte-identical, including eight float controls. git diff --check passes.