nx_bench_intent.nx

nx_bench_intent.nx -- intent/facet-aware retrieval metrics, bits-up. module: nishi-core.search.bench.intent depends: fx.nx (Q16.16 + fx_log2), syscalls.nx (scratch) capability: CORE_COMPUTE wired_status: FULLY_WIRED WHY (operator directive 2026-05-29): rankings must not be "ours" -- they must be grounded in standard, up-to-today IR research and be AGNOSTIC (a person / place / thing-in-time entity model), while still satisfying the DIFFERENT intents a person query carries at once: her WORKS from her own SOCIALS, what she's been IN (filmography), and WHERE/WHEN (timeline, age/date). One ranking cannot over-serve one intent at the others' cost. Method (verified from NIST TREC Dynamic Domain Overview + primary papers): * Graded relevance 1-4 (1 marginal .. 4 key result) [TREC-DD 2017] * Subtopic decomposition == FACETS/INTENTS [TREC-DD] * S-recall: fraction of intents covered in top-k [Zhai-Cohen-Lafferty, "Beyond independent relevance", SIGIR 2003] * Cube Test water-filling with per-subtopic MaxHeight cap: rewards covering many intents, caps redundant hits on one [Luo, Wing, Yang & Hearst, "Water filling: the cube test", CIKM 2013] * Derivative commentary capped at grade <= 2 so it can never outrank a primary source at equal relevance (encodes first-hand > commentary inside the standard graded framework). Agnostic by construction: the facet ids below are a person decomposition, but the kernels take (subtopic_id, grade) arrays -- swap the facet table for a place or product and the same math applies. Subtopic WEIGHTS belong in a config table, never hardcoded (Rule #11); these kernels take the caller's grades as given.