Problem
The draft implementation in #37 can repackage upstream CUDA 12.4 and CUDA 13.3 sidecars with exact native/core provenance, but it is not releasable from runner-only evidence. The dependent Dart consumer is leehack/llamadart#340; on-demand runtime download remains leehack/llamadart#339.
Acceptance criteria
- integrate current native main without force-pushing and rerun exact-head wrapper/provenance checks;
- validate both CUDA 12 and CUDA 13 on representative physical NVIDIA hardware;
- load a representative GGUF and exercise real generation/kernel execution for each family;
- verify same-run core/sidecar hashes, loader behavior, manifest inventory, and coexistence without mixed DLL families;
- recheck tracked CUDA/speculative-decoding correctness risks relevant to publication;
- publish only through a separately approved native release, then update the Dart consumer in ownership order.
Existing work
Draft PR #37 contains the current implementation and historical runner evidence. Keep it draft until the hardware and publication gates above are satisfied.
Problem
The draft implementation in #37 can repackage upstream CUDA 12.4 and CUDA 13.3 sidecars with exact native/core provenance, but it is not releasable from runner-only evidence. The dependent Dart consumer is leehack/llamadart#340; on-demand runtime download remains leehack/llamadart#339.
Acceptance criteria
Existing work
Draft PR #37 contains the current implementation and historical runner evidence. Keep it draft until the hardware and publication gates above are satisfied.