nx_concept_embedding.nx

nx_concept_embedding.nx -- typed concept manifest + storage contract. Closes the STORAGE half of the image-ingestion loop: nx_image_feature_extract N images -> N feature vectors nx_image_feature_centroid N vectors -> centroid (concept signature) nx_concept_embedding centroid + provenance -> manifest <nishi-library writer> manifest -> content-addressed storage nx_cosine_similarity new image vs stored concept -> match nx_render_target_match rendered vs target -> verdict USE CASES (all three of user 2026-05-15's verticals): RETAIL VIRTUAL TRY-ON: name = "lingerie_velvet_red_xs" centroid = mean(features across 5 photos of the product) source_hashes = sha256 of each product photo n_examples = 5 Later prompt: "model wearing [concept:lingerie_velvet_red_xs]" -> blend stored centroid into prompt embedding (IP-adapter path) -> render with the actual product visualized VIDEO-GAME ASSET CATALOG: name = "weapon_runic_greatsword" centroid = mean(concept art frames) -> in-engine asset gen conditioned on the concept STORY-CHARACTER REFERENCE: name = "char_elara" centroid = mean(prior renders the user approved) -> future story renders condition on this concept for identity stability across the arc MANIFEST STRUCTURE (canonical-bytes-serializable, content-addressable): schema_version v1 forever; new version = new struct kind name_buf + name_len concept's human-readable handle centroid NX_IMAGE_FEATURE_LEN nx_int (Q10) cells n_examples how many images contributed to the centroid fidelity_q10 mean fidelity across the N source extracts source_hash_buf 32 * n_examples bytes (sha256 per source img)