nx_compute_runner.nx

nx_compute_runner.nx -- executes a validated ComputeGraph. The brick that closes the loop: with this in tree, the substrate has a working ComfyUI replacement at the compute layer. Build a graph -> validate -> toposort -> RUN -> get real tensor output -> verify via nx_numeric_oracle. All pure NishiLang. Dispatch: NODE_INPUT -- copies caller-supplied input tensor into the tensor store at the node's output slot NODE_CONST -- copies caller-supplied const tensor (weight blob) into the store NODE_OUTPUT -- alias-copies its input tensor into the store so external readers can collect graph outputs NODE_TENSOR_OP -- dispatches by op_code to elementwise / shape ops implemented inline NODE_KERNEL -- dispatches by kernel_kind to nx_blas_i64 (and future nx_conv / nx_attention modules) Tensor lifecycle: * One tensor per output port of each node. * Indexed by tensor_idx = node_id * NX_CN_MAX_OUTPUTS + port. * Allocated by the runner; freed by sys_mmap unwind at process exit (no GC in v1; tight footprint kept by careful sizing). Audit: Caller passes an optional *TraceCfg. If non-null, the runner emits one CONTINUITY_CHECK-style span per node (kind = generic "node executed"). Span attrs carry node_id + op_code + verdict. genealogy_id: comfyui_executor + jax_pjit + tvm_runtime + tensorflow_executor lineage_id: substrate_compute_runner_v1