nx_library_exceed.nx

nx_library_exceed.nx -- MEASURED exceed benchmark for the public publication library (nishifamily.com/library). Two halves, both honest: A) DATA-DRIVEN CENSUS -- reads knowledge/registry/library_sclass_capability_target.tsv and counts HAVE / PARTIAL / MISSING (coverage permil = (have*1000 + partial*500)/total). Self-assessment vs the sourced best-practice target, NOT an exceed claim. Re-runs as the TSV evolves. B) PROOF-BY-RUN of the genuine FORMAT-MOAT exceed axes -- the cells where incumbents (Calibre-Web, Kavita, Komga, Internet Archive) have NOTHING and we PROVE the capability by RUNNING the real UXF organs (no cell asserted true without a passing run): P1 LIB02 publication record round-trips through the envelope -> CID byte-stable P2 LIB02 the CID self-describes as the DOC profile (no out-of-band schema) P3 LIB04 tampering one byte CHANGES the CID (tamper-evident integrity by construction) P4 LIB04 same record, fields in different insertion order -> SAME CID (dedupe/determinism) P5 LIB09 tolerant-reader preserves an UNKNOWN field + known key still resolves == Dublin Core's own "Dumb-Down Principle" (lib_dublin.txt:158), proven in OUR format NO-FABRICATION negative controls (must come out NEGATIVE, proving the metric is not rigged): N1 two DIFFERENT publications -> DISTINCT CIDs (a collision would fail) N2 an absent capability stays absent: LIB01 (OPDS) MUST read MISSING in the TSV VERDICT GREEN iff census parsed rows AND every P1..P5 passed AND every neg-control held. This is the measured-exceed discipline (no self-grade, no fabrication); it is NOT a beats-Wikipedia-on-scale claim. No hardware writes, no network, no production side effects (Rule 26). license_tier: ORIGINAL