CPU and CUDA backends

The default build uses deterministic pure Rust CPU propagation. Start there for reproducibility and to establish a reference before trying acceleration.

Option	Values	Default
--backend	cpu, cuda, gpu	cpu
--precision	f64, f32	f64
--threads	Worker count; 0 selects the available pool	0
--max-paths	Ranked paths stored per corner	5
--max-state-gib	Estimated propagation state limit per corner	16

CPU f64 propagation is designed for bit-identical results across worker counts, with deterministic ID tie-breaking. The default precision is f64; a higher thread count does not imply a different numerical model.

Build and select CUDA

cargo build --locked --release -p ruststa-cli --features cuda
target/release/ruststa --version
target/release/ruststa run --workspace ./ws --base-rev rev1 \
  --backend cuda --precision f64 --max-paths 5 --format json

The backend dynamically loads the NVIDIA driver and NVRTC at runtime, so a CUDA toolkit is not required to compile the Rust package. A working device and runtime libraries are needed to execute CUDA kernels.

gpu chooses an available compiled GPU backend after a device self-test. If CUDA was not compiled in, the device is unavailable, or self-testing fails, the command fails with BACKEND_UNAVAILABLE (A0501, exit 1).

CPU/CUDA f64 propagation has a bitwise equality contract. Opt-in f32 propagation has an explicit tolerance of max(1e-4 ns, 2e-5 × |v|) for |v| ≤ 100 ns. These are test contracts for propagation, not a speedup claim or a statement that every analysis phase executes on the GPU.

CUDA verification needs a real device. Missing hardware is unverified coverage. Metal/Accelerate is a family architecture direction, not a shipped RustSTA backend. See compute decisions and measurement guidance.