nx_image_load_raw.nx

nx_image_load_raw.nx -- load real images via raw-binary header + pixels. The substrate bridges to real images via a SIMPLE raw format that any preprocessor (PIL, ImageMagick, ffmpeg, custom script) can produce: bytes 0..7 width (little-endian i64) bytes 8..15 height (little-endian i64) bytes 16..23 channels (little-endian i64) bytes 24.. raw pixel data, row-major, width * height * channels bytes This is the LAST-MILE primitive — once it lands, nx_color_skin_mask / nx_region_segment / nx_eye_detect / etc all consume real PNG renders via this conversion bridge: # Python preprocessing (svc-quality does this once per image): import struct, PIL.Image, numpy as np arr = np.array(PIL.Image.open(path).convert("RGB")) h, w, c = arr.shape with open("/tmp/img.raw", "wb") as f: f.write(struct.pack("<QQQ", w, h, c)) f.write(arr.tobytes()) # nxc2-compiled detector reads /tmp/img.raw, runs the full pipeline, # writes verdict.json. svc-quality reads verdict.json, emits to UI.