nx_parallel.nx

nx_parallel.nx -- parallel_for / parallel_map / parallel_reduce. Built on top of nx_thread_pool: each primitive partitions its iteration range into n_workers chunks and submits one task per chunk. Wait until pool.tasks_completed has advanced by n_chunks (snapshot-based, so multiple parallel_* calls compose). Why partition rather than per-iter-submit: * Per-iter submit pays ~20ns chan_send + dispatch per index; for cheap bodies (e.g. element-wise add) that dwarfs the work. Chunked submission amortises chan overhead across ~ITER_COUNT/n_workers items. Static even partition for MVP -- range is divided into roughly equal contiguous chunks. Dynamic / work-stealing partitioning is the L7-evolution follow-up once we measure imbalanced workloads suffering >5% throughput loss. Public API (single namespace; sealed strategy enum): nx_parallel_for(pool, begin, end, fn(i64)) nx_parallel_map_i64(pool, in_ptr, out_ptr, n, fn(i64) -> i64) nx_parallel_reduce_i64(pool, ptr, n, init, fn(i64, i64) -> i64) Composes against: [[nx_thread_pool_shared_queue]] (workers), [[vyukov_mpmc_channel]] (task queue), [[atomic_intrinsics_real_amo]] (per-chunk reduction sum), [[nx_hw_dynamic_probes]] (chunk count defaults to nx_hw_worker_count).