# GPU Scaffold
P183 adds a default-build GPU backend scaffold without CUDA, wgpu, HIP, or any
other GPU runtime dependency.
The scaffold is intentionally conservative:
- it implements `ComputeHardware` as `GpuScaffoldBackend`;
- it advertises `DeviceKind::Gpu` and `MemorySpace::Device`;
- it exposes dense integer capability metadata only;
- it does not advertise `padic:fixed_precision` or `sheaf:finite_site`;
- it has no executable kernels and makes no performance claim;
- it returns explicit fallback reports for unsupported p-adic, sheaf, and
transfer paths.
The scaffold uses the P182 `DeviceBuffer` and `TransferPlan` model for transfer
rejection evidence. Host-to-device transfers currently produce structured
fallback reasons, so future accelerated paths must add real memory support
before claiming GPU execution.
This phase is a contract boundary, not an accelerator. Its purpose is to make
future GPU work participate in planner capability checks while preserving the
CPU oracle as the semantic reference.