nx_word_sketch.nx

nx_word_sketch.nx -- substrate-native word sketch (Sketch Engine displacement, see feedback-corpus-linguistics-s-class-substrate-...). A "word sketch" is the canonical Sketch Engine feature: given a target word, list its statistically most significant collocates. Sketch Engine groups by grammatical relation (subj_of/obj_of/...), which requires a parser; this v1 primitive groups by POSITIONAL WINDOW instead (collocates within +/- W of the target), which is what corpus linguists call a "raw" word sketch. Composes: nx_ngram.nx -- token-stream + hash-table count primitives nx_collocation.nx -- PMI / log-likelihood / T-score / log-Dice Caller flow: 1. tokenise corpus into i64 token stream (one ID per word) 2. compute global per-token totals via nx_ng_count (1-grams) 3. call nx_word_sketch_count to fill a collocate table 4. call nx_word_sketch_score_one for each collocate to get log_dice_q10 5. (optional) sort by log_dice and pick top-K No memory allocation inside the primitive — caller supplies all buffers, keeping the substrate scale-agnostic from MCU to HPC. genealogy_id: rychly_2008_logDice + church_hanks_1990_pmi + kilgarriff_2014_sketch_engine lineage_id: positional_word_sketch_v1