Token-level Generalized Advantage Estimation (GAE-λ), Schulman $k_3$ KL penalties, group advantage normalization, and online-softmax block reductions sit on the inner loop of RLVR rollout post-processing. Requiring PyBind11/CMake wheel compilation complicates lightweight container and CI portability, while pure Python loops incur interpreter overhead.
We ship csrc/veritas_kernels.cpp with extern "C" ABI entrypoints compiled on demand via clang++/g++ -O3 -std=c++20 -shared -fPIC, keyed by the SHA-256 digest of veritas_kernels.cpp inside .veritas_cache/, and paired with an exact NumPy fallback in VeritasKernelBridge.
<1ms cached dynamic library loading after first compilation, and Hypothesis-verified numerical parity (< 1e-5) between native C++20 and NumPy fallback paths.