nx_imgsearch_tier.nx

nx_imgsearch_tier.nx -- THE REVERSE-IMAGE TIER BASE CLASS (vtable OOP). A production reverse-image engine is never ONE matcher. TinEye is a copy-detector; Google Lens is a semantic recogniser; Yandex fuses both plus faces and text. Our own ruler (nx_imgbench) proves why: no single descriptor covers the modification space -- dHash owns photometric edits and is blind to geometry, the edge-orientation descriptor owns crops and is blind to mirroring, and neither knows what the picture is OF. So the engine is defined as a FUSION OF TIERS, and a tier is an object with a vtable. Adding the next rung -- local keypoints + geometric verification, or a quantised semantic embedding -- is implemented as a NEW SUBTYPE with its own constructor. It is never an edit to the query engine. That is the whole point of the base class: the engine below it must not grow an `if` per rung. ASYMMETRIC DESCRIPTORS, deliberately. A tier declares TWO widths: idx_dim -- i64 slots persisted per corpus image (storage cost, multiplied by 10^8 images) qry_dim -- i64 slots computed per query (compute cost, paid once per search) They differ for good reasons. The dihedral tier stores ONE fingerprint but computes EIGHT at query time, buying full mirror/rotation invariance for zero storage. A future local-feature tier will store a compact bag-of-visual-words vector but compute a full keypoint set per query. Forcing both sides to one width would have taxed every stored image to pay for a query-side trick. Distances are INTEGER and each tier publishes max_dist, so scores are comparable across tiers without floats -- the same integer-quantisation discipline the text engine rides. license_tier: ORIGINAL