nx_pteam.nx

nx_pteam.nx -- PERSISTENT BARRIER TEAM (fork-join parallel-for), the pool- scaling fix. MEASURED (2026-07-10): the channel pool scales only ~4.5x on 14 threads (gcc OpenMP 8.8x) because workers pick tasks off the MPMC channel via CONTENDED CAS -> serialized start. A fork-join team fixes it: each worker owns a FIXED band index and reads it from a shared descriptor, so all workers start SIMULTANEOUSLY -- no channel, no per-task queue, one generation bump instead of 14 sends+wakes. Idle = futex-sleep on the generation word (0 CPU), same eventcount discipline as the futex pool. API: nx_pteam_new(n) ; nx_pteam_run(t, fn, ctx) calls fn(band, ctx) once per band in [0,n) across the team and returns when all bands finish. SINGLE-DRIVER (one thread calls run at a time) -- matches the LLM forward. genealogy_id: openmp_static_forkjoin + eventcount_futex lineage_id: substrate_pteam_v1