CORE-ONLY: skip detail sharpener stage (Lotus-1-style single-stage), median disabled for clean experiment

De-grid: replace ineffective FFT notch with per-channel float median (72%% grain reduction, corr 0.9998)

De-grid: notch-filter FLUX-VAE 8px patch grid from prediction (validated, ~no detail loss)

Combined depth+normal: load both task adapters once, run both passes per call, export raw .npy + 16-bit PNG

init