nx_f32_vae_decode_tiny.nx

nx_f32_vae_decode_tiny.nx -- first end-to-end sovereign f32 VAE-DECODE (latent -> RGB), the R4/R7 integration milestone of the sd-server -> Nishi migration. Composes the now-gated sovereign f32 bricks into the canonical VAE-decoder shape: latent[1,Cz,Hl,Wl] -> conv_in (Cz->Cm, 3x3 pad1) nx_f32_conv2d -> ResBlock (Cm) nx_f32_resblock (GN->SiLU->Conv x2 + residual) -> upsample (2x) nx_f32_upsample -> conv_out (Cm->3, 3x3 pad1) nx_f32_conv2d -> RGB[1,3,2*Hl,2*Wl] This is the "decode half" of txt2img, proven to COMPOSE + run end-to-end + be deterministic on the software-f32 tier with no third-party. The full Flux/Z-Image decoder is this shape scaled up (16-ch latent, more ResBlocks, 3 upsample stages) with REAL dequantized GGUF weights -- the next rungs wire nx_gguf_load_model + nx_q4k_to_f32 in. Here the weights are fixed small values to prove the pipeline. license_tier: ORIGINAL