nx_gen_embed_verify.nx

nx_gen_embed_verify.nx -- SOVEREIGN F16 linear-with-bias (an embedder), verified vs the oracle. y[t][o] = sum_i x[t][i] * W[o][i] + b[o] The DiT's front-end projections -- x_embedder (patchified latent -> hidden) and cap_embedder.1 (text-encoder output -> hidden) -- are stored F16 WITH A BIAS, unlike the Q8_0, bias-free projections inside the blocks. Two differences, either of which silently produces a finite wrong tensor if assumed away, so this is its own organ rather than a flag on the block linear. Usage: nx_gen_embed_verify <model_id> <gguf> <x_name> <w_tensor> <bias_tensor|-> <y_name> [rows] pass "-" for <bias_tensor> when the projection has none Weights come from the GGUF, activations and the reference from the oracle fixtures: this grades ONE op, so its inputs should be the oracle's, exactly as the per-op bench does. license_tier: ORIGINAL