nx_knn.nx

nx_knn.nx -- k-Nearest-Neighbors classifier. The cardinal feedback-loras-and-negatives-are-patches-not-systems names "classifiers" as a load-bearing piece: inputs from classifiers lead to known outputs with statistical measurement. This file ships the most bits-up classifier in the literature. kNN (Cover & Hart 1967): - No training phase. - No model weights to load. - Just a labeled point cloud + a distance metric. - Given query: find k closest labeled points, return majority class + confidence. Why kNN as the FIRST substrate classifier: 1. Bits-up minimum. No FPU dep, no matrix factorisation, no training optimiser, no hyperparameter tuning beyond k. 2. Theoretical guarantee. Cover & Hart proved kNN error is bounded by twice the Bayes error rate. No deep-learning magic. 3. Composes naturally with the closed-loop + Monte Carlo primitives. Each kNN prediction is a measurable verdict; confidence becomes the variance input to MC. 4. Pluggable into ANY downstream domain. Age, pose, gender, style -- all become kNN problems if the caller provides labeled feature vectors. Distance metric: Euclidean squared (sum of (a-b)^2 across features). We avoid sqrt because ranking-by-distance is the same with or without it; the substrate caller can sqrt if they want a true distance value via nx_isqrt_q10. genealogy_id: cover_hart_1967_kNN + fix_hodges_1951_discriminatory_analysis + duda_hart_2001_pattern_classification lineage_id: substrate_knn_v1 nx_safety_envelope: intended_use: "k-Nearest Neighbors classifier (Cover & Hart 1967) -- substrate baseline classifier; also