nx_pool_stale_gate.nx

nx_pool_stale_gate.nx -- STRESS REPRODUCER for the 2026-07-15 run-to-run greedy divergence (proven by exact-repeat 1.5B ruler runs: identical binary+args+weights, different outputs at near-tie forks; 0.5B byte-stable). Hypothesis class: a pooled op's nx_pool_wait(done0+bands) occasionally returns with a band UNFINISHED (early wake / counter anomaly / ctx hijack) -> the next stage reads a stale band -> perturbed logits flip ties. This gate hammers the EXACT usage idiom of nx_nofloat_llm's pooled kernels (delta-wait + shared ctx arena + shared activation scratch, band-parallel disjoint writes) at high op counts and CHECKS three invariants after EVERY wait: I1 STAMP every band wrote its per-op serial stamp (no missing/unfinished band) I2 VALUE every band's dst equals the exact expected sum of THIS op's activations (no stale-input read, no hijacked-ctx compute) I3 COUNTER tasks_completed advanced by exactly +bands across the wait (no over/under-count) Any violation prints op#, band, got/want and the gate goes RED = reproducer captured. A clean 200k-op run is evidence the race is NOT in this idiom (pool+chan+wait) and the hunt moves up/down a layer. T1: 200k ops x n_workers bands, zero violations T2 NEG-CONTROL: sabotage mode (one band deliberately not submitted, wait shortened) MUST be caught by I1 Return from main (pool reap). No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL