nx_f32_lora_linear.nx

nx_f32_lora_linear.nx -- sovereign LoRA-adapted linear (the Aligned-LoRA / fine-tune-fewer-images lever). From DiT-IC's "Aligned LoRA Adaptation" (ingested 2026-07-01): fine-tune a frozen weight W [out,in] with a low-rank adapter B [out,r], A [r,in] (r << in,out). Forward: out = W x + B (A x). Only r*(in+out) params train instead of in*out -> huge param/VRAM/data saving (the "fewer images, faster training" lever), and LoRA merges into W for zero-overhead inference. Gate: LoRA forward == (W + B@A) x within f32 tolerance. license_tier: ORIGINAL