# Feature-Gated Accelerated Pilot
P187 adds the first non-CPU execution pilot behind the `accelerated-pilot` feature.
## Scope
The pilot backend is `GpuDenseI64PilotBackend`. It advertises a GPU-like hardware target and supports one narrow operation:
- dense integer domain
- dense CPU representation as the host-side tensor boundary
- single-step pointwise `add`
- `i64` execution through `execute_i64_add_with_cpu_oracle`
The implementation intentionally avoids CUDA, wgpu, HIP, streams, device allocation, and external runtime dependencies. It is a feature-gated pilot kernel boundary, not a performance claim.
## Semantic Guardrail
Every pilot execution runs a CPU oracle comparison:
1. Clone the input `TensorStore`.
2. Execute the same graph with `CpuScalarBackend`.
3. Execute the pilot dense i64 add path.
4. Compare every checked output exactly against the CPU oracle.
5. Return `GpuDenseI64PilotReport` with checked outputs, match status, and evidence.
The report keeps `preliminary_runtime_ns` optional and currently unset. Benchmark and performance evidence are deferred to later phases.
## Explicit Non-Support
The pilot rejects:
- p-adic domains
- finite-site sheaf / cover-local paths
- reductions
- matmul
- fused steps
- any non-integer domain
Unsupported paths return explicit backend errors instead of silent fallback or accidental execution.
## Validation
The default build remains dependency-light and offline-testable. The pilot is validated separately with:
```bash
cargo test --offline --features accelerated-pilot --test accelerated_pilot
```
Default CI-style validation still uses:
```bash
cargo test --offline
```