D70 — Native macOS with an optional Metal backend

Decision

Native macOS uses Apple Clang and an optional Metal cost-to-go backend; field_gpu, compiled_backend, and backend-neutral device indices replace CUDA-specific routing APIs. CUDA remains the Linux backend and CPU references remain universal. Metal kernels are embedded, compiled once per device, and use reusable per-thread shared buffers, ordered dispatches, and an atomic convergence flag.

Alternatives

CPU-only Mac support; CUDA translation layer; a full GPU router or GPU placer

Reason

Direct Metal supports Apple silicon without changing routing decisions. Integer CPU equality and real-board byte equality gate the port. GPU availability is not a speed guarantee; Metal reports working-set headroom rather than claiming free VRAM.