nx_acquire_lib.nx

nx_acquire_lib.nx -- THE LIBRARIAN'S DOOR: what may we do with an EXTERNAL artifact? WHY THIS EXISTS. The estate already holds every primitive for taking in a third-party research artifact: nx_research_fetch mirrors the bytes and emits a provenance line, nx_canon_cid content- addresses them (cid_of_file), nx_asset_provenance signs the binding and already ships a prov_class_downloaded, nx_mirror_orchestrator verifies a digest FAIL-CLOSED and derives an idempotent CAS key, and nx_licgate_lib decides commercial rights. NOTHING JOINS THEM, so an external work has never had a record. This lib is that record's DECISION half. THE MEASUREMENT THAT SHAPED IT, taken 2026-08-25 against the mirrored listing at https://europe.naverlabs.com/research/code/ (knowledge/fetched/cmp_librarian_naverlabs_code.html, 522716 bytes, sha256 deeb6ab8..5592): of the 26 releases on page 1, exactly 2 carry ANY licence string -- 77 permil -- and BOTH are prose parentheticals inside a description, not a field. The markup has no licence element at all. 24 of 26 would require opening the linked repository. => A LIBRARIAN THAT READS THE LICENCE OFF A LISTING PAGE IS WRONG 923 TIMES IN 1000. => EVIDENCE IS ITS OWN AXIS. "the page asserted Apache-2.0" and "we hold the LICENSE file and read it" are DIFFERENT FACTS and must never collapse into one column. An unread licence ABSTAINS. This is the same law nx_licgate_lib's `verified` column encodes -- and which, until 2026-08-25, it parsed and never consulted, so an admittedly-UNCONFIRMED row returned full SHIP_OK. THE FOUR VERDICTS ARE DELIBERATELY INDEPENDENT, because they fail apart in practice. Collapsing them is exactly how a project gets mirrored for study and then accidentally republished: may_mirror -- may we keep a copy at all. THE ANTI-ROT RIGHT. may_redistribute -- may we serve that copy onward to third parties. may_learn_from -- may we study it, and use it as an ORACLE or a BENCHMARK, taking no code. must_attribute -- an OBLIGATION, not a right: it takes the MAX and DEFAULTS TO YES. REVIEW BLOCKS. THE ONLY VERDICT THAT PERMITS IS YES. A consumer that reads REVIEW as permission has misread this lib -- REVIEW means "a human must settle a term a machine may not", which is the third state, not a soft pass. Encoded in the SAME order as nx_licgate_lib (0 NO < 1 REVIEW < 2 YES) so "most restrictive wins" stays a plain MIN and the two rulers can never disagree on direction. PURE BY CONSTRUCTION: aq_decide touches no file and no clock. Its whole input is its arguments, so its gate needs no fixture and cannot go vacuous through a missing one. license_tier: ORIGINAL