nx_gen_adaln_gate_verify.nx

nx_gen_adaln_gate_verify.nx -- SOVEREIGN adaLN gated residual join, verified vs the oracle. out == inner * tanh(gate) + residual where `gate` is one chunk of a block's adaLN modulation vector, broadcast across all tokens. This is the join at the end of each half of a modulated DiT block. Usage: nx_gen_adaln_gate_verify <model> <inner> <adaln> <n_chunks> <gate_idx> <residual> <out> [rows] The three tensors are NAMED, not derived from a scope with fixed op names. The first version hardcoded attn_norm2/blk_in/mid, which meant it could only ever check the ATTENTION half of a block -- the identical join at the end of the FFN half needed a second organ. Naming them makes one organ cover both halves, and any architecture that gates a residual this way. The CHUNK COUNT AND INDEX ARE ARGUMENTS because they are architecture facts, not universal ones: Z-Image packs 4 chunks [scale_msa|gate_msa|scale_mlp|gate_mlp], Flux-style blocks pack 6 [shift|scale|gate] x2. Baking either in would make this organ silently verify the wrong slice the first time it met another model -- and a wrong slice still produces plausible numbers, so it would not announce itself. This organ needs NO model weights: every input is an oracle fixture. That makes it the cheapest possible check of the chunk ORDER, an assumption every downstream organ inherits. license_tier: ORIGINAL