Execution Planner

The public defaults are streamed_edges="direct" and execution_profile="capacity", with allow_fixed_workspace=False. The Python frontend owns artifact discovery and compatibility diagnostics; native code provides contract validation, backend capability, graph-size and memory information. Fixed-workspace candidates require explicit frontend opt-in. The current detailed graph-time plan report applies to MH-0 MACE/MACEField.

For standard compact MH-0 models, speed is a capability-selected throughput bundle rather than a synonym for retaining all intermediate state. It retains harmonic gradients for small graphs and switches to Y-only harmonic storage at 40,960 directed edges when qualified on CPU or CUDA FP32. CUDA FP64 and HIP retain harmonic gradients. Both variants reuse eligible MH-0 state and recompute readout state. FP32 uses compact geometry; FP64 retains Cartesian geometry. Capacity selection uses the same throughput candidate first and may also select Y-only storage when the estimated device-memory budget requires it. Debug plan identifiers remain exact even on hosts where no device-memory query exists.

The public direct-execution matrix is maintained in Execution Support Matrix; architecture and planner invariants are described in Direct Execution Architecture. Exact debug-plan identifiers remain private implementation details in symmetrix.calculator.