nx_nofloat_muon.nx

nx_nofloat_muon.nx -- the FULL Per-Head Muon optimizer LOOP in sovereign no-float (K3 F405; operator 2026-07-19 "beyond SOTA, novel, evidence-driven"). Muon = MomentUm Orthogonalized by Newton-schulz (Keller Jordan / Moonshot). One optimizer step on a weight matrix: B_t = mu * B_{t-1} + G_t (heavy-ball momentum buffer) U_t = NewtonSchulz(B_t) (orthogonalize the momentum -- the proven nx_nofloat_muon_ns core) W_t = W_{t-1} - lr * U_t (apply the orthogonalized update) PER-HEAD: attention weights are nheads independent hd x hd blocks; each is orthogonalized SEPARATELY (that is the "Per-Head" in Per-Head Muon -- a per-head spectral step, not one global orthogonalization). PURE INTEGER Q16 throughout -> BIT-EXACT DETERMINISTIC optimizer: a float Muon drifts (non-associative gradient/momentum accumulation + float NS); ours is identical every machine, every run = the novel TRAIN exceed. Composes nx_nofloat_muon_ns (mn_orthogonalize). license_tier: ORIGINAL No hw writes (Rule 26).