Engine agreement and throughput (DESIGN.md section 5)

Scripts used for section 5 of docs/DESIGN.md. They are measurement scripts, not part of a package.

Environment: a uv virtualenv with Python 3.13 and numpy scipy numba jax refnx genx3 torch; CUDA runs used a separate environment with torch 2.11.0+cu128. rl_abeles.py is ReflectoLearn's engine at b6fb0c4:

git -C ReflectoLearn show b6fb0c4:src/reflectolearn/physics/abeles.py > rl_abeles.py
cargo build --release --manifest-path rs/Cargo.toml
python -I compare.py . agreement.json        # section 5.1, 2000 q points
python -I bench2.py . throughput.json numba-f64-par --jac   # section 5.3, one engine per process
python -I bench2.py . throughput.json rust-12
python -I jac_torch.py . throughput.json x

Engines for bench2.py: numpy-f64 numba-f64-par rust-12 jax-cpu-f64 jax-cpu-f32 torch-cpu-c128 torch-cpu-c64 torch-cuda-c128 torch-cuda-c64 refnx genx (--jac for numba, JAX f64, torch c128). bench.py holds the engine definitions; its own main is the earlier pilot at 2000 q points.

agreement.json and throughput.json (quiet window) are the outputs quoted in DESIGN.md; throughput_pilot.json is the earlier run outside a quiet window. compare.py reads the Rust curves from rust_out/, which bench.py writes.