nx_par_pool.nx

nx_par_pool.nx -- SOVEREIGN persistent worker POOL via PIPES (robust fix; busy-spin deadlocked on LICM, ledger #21). Forks PP_NW workers ONCE; dispatches MANY matmuls via per-worker go/done PIPES -- kernel- BLOCKING read/write (no busy-spin, no LICM hazard, no idle CPU burn). Data (A/B/C) in sys_mmap_shared (set up BEFORE fork). Dispatch = write 1 byte to a worker's go-pipe; the worker computes its fixed range then writes 1 byte to its done-pipe; shutdown = close the go-pipe (read returns 0 -> worker exits). Pure syscalls (pipe2/fork/read/write/close/wait4) -- NO libvulkan/CUDA/pthread. Proves: fork-once + blocking dispatch => N-core speedup that HOLDS across many calls without per-call fork. license_tier: ORIGINAL