nx_nofloat_deepstack_gate.nx

nx_nofloat_deepstack_gate.nx -- HARD-EVIDENCE gate that the no-float transformer scales in DEPTH: a 2-BLOCK (depth-2) transformer LM trains end-to-end on next-token prediction, pure integer Q16. Each block is the verified pre-norm block (attn+RoPE+causal-softmax+SwiGLU+residuals); stacking = composing blk forward N times (NO new ops -- depth is pure composition). Trained with the Q16 AdamW optimizer. A1 depth-2 gradcheck wrt Wlm : the LM-head gradient == finite differences. A2 depth-2 gradcheck wrt E : E's gradient flows back through BOTH blocks (the deep-backprop proof) == FD. D neg-control teeth (E, largest component) ; C bit-exact. B the depth-2 LM TRAINS : next-token CE drops substantially (AdamW on E+Wlm through the 2-block stack). (informational) depth-1 vs depth-2 final CE printed for transparency. Evidence -> knowledge/status/nofloat_deepstack.log. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0