nx_lowrank_train.nx

nx_lowrank_train.nx -- M2: does a LOW-RANK bottleneck IMPROVE quality (not just preserve it)? The MLA/ inductive-bias thesis (operator: "quality UP, VRAM down -- not just quant shrinking quality"). Trains models on a NOISY rank-2 target, scored on a CLEAN held-out test set, swept over many SEEDS (robustness) AND over bottleneck RANK (the U-curve -- guards the over-claim "smaller is always better"): FULL: y = W x (W is D x D = 64 params) -- can overfit the noise LOWRANK: y = Wu (Wd x) (Wd r x D, Wu D x r = 2*r*D params, the rank-r bottleneck) CLAIM (precise, not over-claimed): the OPTIMAL bottleneck rank MATCHES the data's true rank R0=2 -- too small (r=1) UNDERFITS, too large (r=4) OVERFITS like full, and at r=R0 it beats full at HALF the VRAM. Sovereign autograd (nx_tgrad_core tg_* + AdamW ad_step; the proven _t4 recipe). nx_cc UNTOUCHED (operator: it's the parallel effort). license_tier: ORIGINAL