nx_nofloat_k3stack.nx

nx_nofloat_k3stack.nx -- MULTI-LAYER K3 stack with ATTENTION RESIDUALS in no-float (operator 2026-07-19 "logically integrated, mature"). Stacks the K3 block at DEPTH, and closes the K3-specific Attention- Residuals gap: instead of a plain x+f(x) residual, each layer's input is a CONTENT-WEIGHTED ATTENTION over ALL preceding layer outputs (arXiv 2603.15031: "learned input-dependent attention over preceding layers... earlier-layer representations stay reachable deeper in the stack"). Minimal faithful AttnRes: score_s = |mean(h_s) . mean(h_{l-1})| ; w_s = score_s / sum ; r_l = sum_{s<l} w_s h_s ; h_l = block(r_l) Pure integer Q16, deterministic AT DEPTH -- a float stack drifts per-layer and compounds; ours is bit-exact through N layers. Composes nx_nofloat_k3block (the K3 block) + weight-tying across layers. license_tier: ORIGINAL No hw writes (Rule 26).