nx_atlas_gapscan.nx

nx_atlas_gapscan.nx -- F225/F225a STRUCTURAL GAP SCAN over the REAL from-god graph: under-connected node pairs scored on the contract axes ([[project-nishi-atlas-loop-integration-2026-07-18]]: "walk the graph for under-connected node pairs x maturity-gap x momentum"). Sibling of nx_atlas_discover (recombines the ~28-row commontask CATALOG -> 12 mcp-live units, silently capped at 48 organs / 40 first-past-the-post proposals) and of nx_atlas_recombine (dmkt PRODUCT recombination, one-best-partner-per-anchor by capability richness). This organ is the MATURITY-GAP axis: it scans all candidate pairs and ranks by true top-K. Four defects of the catalog-mode sibling are fixed BY CONSTRUCTION here: (1) SILENT CAP -> every bound is DECLARED in the envelope line (scale law: no silent truncation). (2) O(n^2 x E) exclusion rescan -> per-node wing/ca/ce are HOISTED once (O(n)); the pair body is pure arithmetic and the O(deg) edge probe runs ONLY for pairs that already clear minscore. (3) "top proposals" that were actually FIRST-40 -> a real bounded top-K insertion, descending by score. (4) toy substrate -> candidates come from the graph, islands and test scaffolding excluded. nx_atlas_gapscan [store-prefix] [out-prefix] [minscore] [topk] defaults: knowledge/store/ecograph_full knowledge/store/gapscan- 5 40 out row: DSC<n> <A>-><B> <score> proposed atlas-graph <evidence> try: plan 10 <A> then 20 <B> NOVELTY = absence of an import edge in EITHER direction (the graph's own ground truth), so an already-composed pair can never be proposed -- the F225 neg-control holds structurally, not by a filter. HUB EXCLUSION (measured, not assumed): the first live run over 16651 nodes returned a top-K saturated by nx_syscalls.nx (Ca=13539) -- a leaf imported by ~80% of the tree carries NO under-connection signal, so pairing it with anything is noise, not novelty. Nodes at or above AR_HUB_CA are therefore not candidates, and an ANCHOR CAP bounds how many emitted rows may share one node so a single anchor cannot crowd the set. Both bounds are reported in the envelope. Wing identity is an FNV hash of the name prefix through the 2nd underscore; a hash collision can only COST a cross-wing bonus (never invent one), so the failure mode is a missed proposal, never a false one. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0