nx_ta_parallel_grad_gate.nx

nx_ta_parallel_grad_gate.nx -- KEYSTONE for FAST model training: the f32 tensor autograd (ta_*) trains DATA-PARALLEL on the sovereign thread pool. Each example in a batch runs its OWN forward+backward on a PRIVATE (tape,vals,grads,st) arena; the per-example gradients are then reduced. The soft-float training that bottlenecks the neural reader (and every bigger model the benchmark endgame needs) parallelizes across the 16 hardware workers here -- the conv2d pool already measured 6.76x wall-clock. The liar-kill is STRONG: serial and parallel compute each example's gradient with identical math, so the per-example gradients must be BIT-IDENTICAL (f32 raw equality). ANY race, arena collision, or shared-state bug in the parallel path perturbs a bit and fails. (No f32-associativity fuzz: we compare PER-EXAMPLE grads, not a reordered sum.) T1 parallel per-example grads == serial, BIT-IDENTICAL across all B examples x all NP weights T2 the pool actually ran B tasks (completed counter advanced by B) on > 1 worker T3 deterministic (re-run parallel == identical) expect_exit: 0 license_tier: ORIGINAL Sovereign: nx_thread_pool + nx_autograd_tensor + nx_syscalls.