nx_nofloat_weighttie_gate.nx

nx_nofloat_weighttie_gate.nx -- CAP-NF-WEIGHTTIE: tied input-embedding / output-projection (GPT-style parameter efficiency), pure no-float Q16. The SAME embedding tensor E is used both to embed tokens AND as the output projection (logits = hn . E^T via matmul_nt). One leaf node nE is referenced by two ops, so the autograd tape ACCUMULATES gradients from the embed path + the output path into E -- elegant and exact. Removes the separate Wlm (saves dm*V params). Trained on the 3-category grammar. T1 the tied LM LEARNS (held-out CE << uniform ln(8)=2079). T2 the tied LM is NEAR-OPTIMAL (CE ~= floor 964) -- tying does NOT hurt quality, at fewer params. T3 TEETH: an UNTRAINED tied model has ~uniform CE -> the low CE is from training, not the architecture. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0