nx_lw_cache_gate.nx

nx_lw_cache_gate.nx -- gate for the packed dequant-once cache in nx_f32_lazy_weight (the m=1 decode lever: stop re-dequanting the same weights every token; cached matmuls are pure packed __f32x4_dot bands). Exact-f32 regime construction as in nx_q4k_matmul_x4_gate (dense d=1.0/sc=1 blocks -> integer dequant values 0..15; int A; all |sums| < 2^24), so bit-exact comparison between the CACHED path (x4 dot order) and the scalar streaming oracle is legitimate. Checks (8): 1 scalar oracle OK (k=512, n=37, m=3) 2 first dispatch fills + matches oracle; pk_state=1; used bytes exact 3 second dispatch (new A) matches fresh oracle (cache reuse + shared-pool delta-wait across fill+dot submissions) 4 m=1 decode shape via dispatcher matches oracle 5 budget refusal: fresh weight + tiny budget -> pk_state=-1 AND result still matches oracle (streamed scalar-pool fallback) 6 stale k=1 shape surfaces NX_LW_ERR_INNER (guard before cache) 7 decode-shape speedup: cached reps vs streamed reps (same data, bit-exact) -- floor cached >= 1.5x streamed; first-call fill overhead printed for the honest record 8 cached and streamed big-shape outputs bit-identical lineage_id: lw_cache_gate_v1