- CUDA Graph: node enumeration/params, exec-update, capture-dependency, multi-stream,
  upload, explicit graph construction via Runtime and Driver proxies.
- cuBLAS: strided/array batched GEMM + Ex, device scalar pointer mode, epilogue pointer
  translation, stream state cache, pipelining, libcublas.so.11 ABI version script.
- Allocator/storage: caching reuse verified, allocation/event/stream leak detection,
  abnormal session cleanup reclamation.
- Big data: Driver SHM bulk path (pinned arena, chunking) + SPSC ring D2D fast path.
- NVRTC: server-side PTX->cubin JIT cache keyed by code hash/device/driver version.
- Governance: per-IP/identity/unauthenticated session limits, per-user quota + global
  memory cap with alloc/free accounting, virtualized cudaMemGetInfo, ticket-based fair
  scheduler, process-shared stats block, worker exception reclamation.
- Validation: reproducible report in validation/ (ctest, concurrency/D2D, native vs
  proxy CSV, RPC stats CSV, runbook).