nx_vae_tile.nx

nx_vae_tile.nx -- overlapping tile split + alpha-feathered blend. Ships VRAM-track V-001 per docs/VRAM_OPTIMIZATION_REALISTIC_TRACKING.md: VAE tiling with overlap blending. ~1-1.5 GB saved during decode at ZERO quality cost when overlap >= 16 and blending is feathered. Substrate primitive -- model-agnostic. The caller orchestrates: tiles, meta = tile_split(latent, h_tile, w_tile, overlap) for each tile_i: out_tile_i = decode_2d_block(tile_i) -- caller's model image = tile_blend_feathered(out_tiles, meta, overlap, H, W) The decoder runs on one tile at a time, so peak VRAM is roughly (model weights + ONE tile of activations) rather than (weights + full-image activations). For Z-Image-class latent 96x128 -> image 768x1024 (8x upsample), per-tile decode of 64x64 latent -> 512x512 image drops decoder working memory by ~6x. Pure NishiLang i64 row-major. Single-channel buffer (multi-channel = caller calls per channel; for fused CHW caller passes the flattened buffer and strides). Q10 fixed point for blend weights. Quality envelope (honest measurement basis): * overlap >= 16 px and alpha-feather -> sub-perceptual seams * overlap = 8 px and alpha-feather -> visible boundary at high-contrast edges * unfeathered blending (constant 1) -> hard seams; refused. We ship feathered only; unfeathered is documented as deliberate non-shipping per the no-silent-corruption cardinal. Algorithm (alpha-feathered blend): For each tile, weight at position (tr, tc) inside the tile is w_y = min(tr, overlap, tile_h - 1 - tr) // distance from w_x = min(tc, overlap, tile_w - 1 - tc) // tile edges weight = (w_y + 1) * (w_x + 1) so weight rises from 1 at the tile boundary to (overlap+1)^2 at tile interior (overlap >= edge distance saturates). Output: image[i, j] = sum_t (weight_t * tile_t[r, c]) / sum_t weight_t where the sum is over all tiles t containing (i, j).