nx_zimage_real_proj.nx

nx_zimage_real_proj.nx -- a REAL Qwen transformer projection on REAL weights, sovereignly. sd-server -> Nishi migration: the first genuine transformer-forward computation on the actual model. Dequantizes token-0's real embedding (token_embd row, 2560 Q6_K) AND the first 4 output-neuron rows of blk.0.attn_q.weight (the real Q-projection matrix, in=2560->out=4096, Q6_K, at file offset ~335MB), then computes the attention Q projection Q[o] = sum_i embed[i] * Wq[o][i] with our own f32 mul/add. Real Qwen weights -> our sovereign matmul -> real Q values. This is the same Qwen that Z-Image uses to encode image prompts AND (in the elder fork) to chat -> the shared-LLM reuse, now flowing through OUR code. Read is a bounded 384MB prefix (covers metadata + token_embd + the start of blk.0); we only dequant the slices we need (embed row + 4 Wq rows), not the full multi-GB tensors. license_tier: ORIGINAL