nx_nofloat_muon_gate.nx

nx_nofloat_muon_gate.nx -- gate: the FULL Per-Head Muon optimizer loop (momentum + NS-orthogonalize + apply), deterministic, per-head. Honest scope: proves the OPTIMIZER STEP mechanics + the novel exceed (bit-exact), NOT a full training convergence run (that rides the GPU seat F101). T1 MOMENTUM heavy-ball: repeated aligned grads grow the buffer beyond a single grad (geometric sum) T2 ORTHOGONALIZED update: U = NS(momentum) has far lower orthogonality error than the raw momentum T3 UPDATE APPLIED EXACTLY: W_after == W_before - lr*U elementwise (mechanical correctness) T4 DETERMINISTIC: two identical per-head runs -> bit-identical weights (the exceed: float Muon drifts) T5 PER-HEAD INDEPENDENCE: perturbing head-0's grad changes head-0's W but NOT head-1's W T6 STABLE: 12 steps stay bounded (no blowup) license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0