1.0.13

• Measure merge cost against the union of the parts' original meshes instead of the hulls being merged — merge error no longer accumulates across successive merges, and decompositions generally produce fewer hulls at the same threshold (#110). Note: this changes default output; on a 61-mesh V-HACD benchmark, average hull count went from 34.10 to 31.49 with max concavity essentially unchanged.
• Don't abort when a cutting plane passes near a triangle vertex between the two internal tolerances — such cases now fall back to the closest vertex instead of throwing from inside the parallel region, which crashed the process (#110)
• Add a COACD_WITH_OPENMP environment variable for pip builds to force the std::thread backend (e.g. dual-libgomp environments such as ROCm/HIP torch), document the parallel-backend CMake options, and honor WITH_STD_THREADS=OFF when OpenMP is requested but not found (#109, thanks @alexis779)