nx_imgbench.nx

nx_imgbench.nx -- THE REVERSE-IMAGE RULER. Answers, by measurement rather than opinion, the only question that matters on the road to TinEye/Google parity: WHICH REAL-WORLD IMAGE MODIFICATIONS DOES OUR ENGINE ACTUALLY SURVIVE, AND WHICH DOES IT NOT? TinEye's whole product claim is finding MODIFIED copies -- resized, cropped, edited, colour-adjusted, watermarked. Google Lens adds semantic "looks like". Our engine's near-duplicate tier is a 64-bit GLOBAL perceptual hash (nx_phash dHash) and its looks-alike tier is an 80-dim integer edge-orientation descriptor (nx_visdesc, MPEG-7 EHD / GIST lineage). Both are GLOBAL descriptors. Global descriptors have a known, structural blind spot -- geometry. This ruler MEASURES exactly how big that blind spot is instead of asserting it, and emits a GAP QUEUE ranked worst-first that IS the build order for the next rung (the same discipline the gamebench ruler uses for the game lane). METHOD (liar-killed by construction): * a deterministic corpus of NCORP procedurally-generated multi-scale images (coarse blob field + a structural layer + fine detail) -- distinct, reproducible, no network, no fixtures on disk. * NQ of them are QUERIES. Each query is put through all 16 nx_imgxform classes. * for every (query, class): rank the transformed image against ALL NCORP originals and check whether the true source ranks #1. That is recall@1 -- the metric a reverse-image engine lives on. * ALSO report det@THRESH: would the LIVE client (threshold 10 Hamming) have returned it at all? recall@1 can be 1000 while det@10 is 0 -- "nearest but not near enough" is a silent miss in prod. * dHash (Hamming) and visdesc (L1) are measured HEAD-TO-HEAD on identical inputs, so the decision to wire the unwired descriptor tier is made on data, not taste. HONESTY BOUND, stated in the output: recall@1 against NCORP distractors is an UPPER BOUND on real-web performance. A production corpus has 10^8..10^10 distractors; a class scoring below ~900 here is already unusable, and a class scoring 1000 here is merely "not yet disproven". Composes nx_imgxform (deform) + nx_phash (hash) + nx_visdesc (describe) -- ZERO new ranking math. license_tier: ORIGINAL