nx_semantic_retrieval.nx

nx_semantic_retrieval.nx -- SPEC A built bits-up: match a query BEYOND exact terms via co-occurrence. Terms that appear together across the corpus are related; so a document can be relevant to a query even when it never contains the query word, as long as it contains terms that CO-OCCUR with it. This is the rung where retrieval can exceed BM25 (which scores 0 when the exact term is absent). Built on Layer-8 integer ALU ops only (the spec's primitives): COUNT (term presence), IMUL+IADD (vector dot product), ISQRT (cosine magnitude) -- no floats, no learned embeddings. Honest boundary: co-occurrence over a small corpus is a weak proxy for true semantics; learned embeddings (the EMBED_TRAIN primitive the Researcher flagged) are the LLM rung, not faked here. license_tier: ORIGINAL Pairs with nx_research_extract (re_has/re_count) + nx_bm25 (exact baseline).