nx_dedither.nx

nx_dedither.nx -- RECOVER A 1990s DITHERED IMAGE WITHOUT TOUCHING THE ORIGINAL BYTES. WHY THIS EXISTS (measured 2026-08-07, page3.com/13Nov96.gif): a 280x403 GIF weighing 81,953 B is a PHOTOGRAPH crushed into a 256-entry palette. The encoder of 1996 traded colour depth for file size by scattering alternating palette entries across smooth regions -- so a plain studio backdrop is stored as high-frequency speckle. The archive serves those bytes FAITHFULLY, and faithful is exactly why the backdrop reads as white fuzz. The fuzz is not our defect and not decode damage: header valid, trailer 0x3B present, extension preserved 1:1 from the 1996 HTML, no re-encode anywhere in the pipeline. *THE ARCHIVE WAS RIGHT AND THE PICTURE STILL LOOKED WRONG. Those are different problems and only the second one is fixable. Rule 25 says build intelligence, never strip features -- so this NEVER rewrites the stored asset. It emits a SEPARATE enhanced render beside it. The original remains the record. HOW: a sigma filter (Lee 1983) -- average only over neighbours whose colour is already CLOSE to the centre. Dither speckle is by construction a set of near-neighbours straddling the true value, so it averages away; a real edge has far neighbours, which are EXCLUDED from the mean and therefore survive. A plain box blur would also lower the noise metric -- and would destroy the picture. That is precisely why the selftest below refuses to pass on noise reduction alone. RULER: mean absolute local deviation (MALD) vs a 3x3 mean, x1000, integer. Reported BEFORE and AFTER so the claim is a measurement and not an adjective. usage: nx_dedither <in-image> <out.png> [radius=2] [threshold=40] nx_dedither --kat (selftest: 4 teeth, 2 of them anti-vacuity) argv[1]=input (any format nx_img_bytes_to_rgb decodes), argv[2]=output PNG. exit 0 ok | 2 decode/write failure | 3 usage | 1 selftest RED license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0