antecedent-kernels
Borrowed matrix/vector views and scalar, portable-optimized, and (optional) architecture-specific kernels. Contains no causal semantics. 1 and §23.2.
Public dispatch selects once per batch via KernelPolicy (force_scalar,
allow_portable_optimized, allow_arch_simd). Scalar is the correctness gold
standard; portable is always compiled. Arch SIMD is selected only when
arch_simd_available() is true (requires a justified simd-runtime feature).
Kernels include masked reductions/covariance, gather/copy, standardization, pairwise L1, contingency accumulation, bootstrap weighted sum/mean/dot, and partial correlation.