memra-validate — shared validation-protocol core (Phase D extraction, ARCHITECTURE-H100.md §5).
Pure host logic, zero engine/CUDA dependency: the pieces every gate bin and bench duplicated (kernel_check, gdn_bench, fa_sanitize, dtype_gpu_check5, the N=5 bench protocol). CPU kernel references stay with their kernels; THIS crate owns the protocol: deterministic test vectors, error measures, tolerance banding, N-rep medians, and the ALL-GREEN tally contract.
Extraction law: moved code is verbatim (bit-identical vectors and measures) — the
pr generator here is the kernel_check/dtype_gpu_check5 variant; fa_sanitize's
16-bit variant intentionally stays local to it (different distribution = different
test vectors = a silent gate change).