nx_gen_lora_repr.nx

nx_gen_lora_repr.nx -- DECIDE the adapter representation by measuring it, not by arguing it. The hot matmul consumes Q8_0 blocks + per-block f16 scales. nx_lora_fold produces f32. So an adapted run has exactly two shapes, and they trade accuracy against per-step cost: A) FOLD-AND-REQUANTIZE — dequantize Q8_0, fold, requantize to Q8_0, run the normal kernel. One-time cost, zero per-step cost, hot loop untouched. This is what a "merged checkpoint" is. But it quantizes a weight that was ALREADY quantized once. B) RANK-FACTORED DELTA — keep the original Q8_0 blocks untouched and add scale * (h @ down^T) @ up^T in f32 at matmul time. The base carries NO new error. Costs 2*rank*(in+out) per token per step forever -- what sd.cpp pays. ★ BOTH ARE GRADED AGAINST THE EXACT f64 REFERENCE, NOT AGAINST EACH OTHER. Comparing A to B answers "do they differ", which is the question nobody asked; comparing each to the truth answers "which is right". This is the same two-way-comparison trap that made the oracle's quantization error look like ours for most of this lane. Usage: nx_gen_lora_repr <gguf> <lora.safetensors> [multiplier_milli] license_tier: ORIGINAL