nx_asset_autotag.nx

nx_asset_autotag.nx -- PERCEPTUAL NEAR-DUPLICATE detection (R3 of the universal org-tooling arc). Makes the catalog know which media are NEAR-duplicates -- the ones exact-CID dedup CANNOT catch. Content-addressing (ar_cid) gives EXACT dedup: identical bytes collapse to one CID, but ONE changed pixel / a recompression / a resize yields a DIFFERENT CID while the image is visually the same. A PERCEPTUAL hash fingerprints what the media LOOKS like, so near-identical media land a tiny Hamming distance apart and cluster together (org_research.tsv CONFIRMED perceptual-autotag: perceptual hashing groups near-dups that exact-hash/CID misses; machines pre-tag, humans curate). ZERO NEW ALGORITHMS -- pure COMPOSITION over the media-studio kernels: * perceptual fingerprint (dHash, brightness/recompression-robust) -> nx_phash (nx_phash_dhash) * bit-distance + near-dup clustering -> nx_simhash (nx_simhash_hamming) dHash takes a grayscale (w x h, row-major u8) buffer and is FORMAT-AGNOSTIC by design, so at_phash fingerprints any decoded grayscale media plane (the decode step is the caller's; this layer is the fingerprint + grouping, kept pure/deterministic). No hardware/persistent-firmware writes (Rule 26). license_tier: ORIGINAL