nx_fnet_mlm_gate.nx

nx_fnet_mlm_gate.nx -- GATE for MODEL-002: a tiny MASKED LANGUAGE MODEL on the FNet mixer. FNet's Fourier mix is BIDIRECTIONAL (every position sees every other), so its native LM objective is masked-LM (BERT-style), NOT causal next-token (which would let the model see the answer). Here: one position of a 4-token palindrome [a,b,b,a] is replaced by a MASK token; the model fills it in -- which requires MIXING the mirror position into the masked one (a pure bag-of-words model cannot). No attention anywhere. tokens(+MASK) -> EMBED(trained) -> FNET mix -> relu FFN -> softmax-CE over the vocab at the masked slot G_train masked-token accuracy >= 7/8 AND final loss < first loss. G_repro bit-exact: train twice -> identical accuracy + final-loss bits. Evidence -> knowledge/status/fnet_mlm.log (FNETMLMGATE authored=organ ... verdict=GREEN). license_tier: ORIGINAL