nx_atlas_graphrec_lib.nx

nx_atlas_graphrec.nx -- NATIVE GRAPH-ADJACENCY RECOMBINATION over the eco graph (F225a). WHY (measured 2026-07-30, debt seq1489): nx_atlas_discover sources candidates from the commontask catalog -- ~48 task rows yielding 27 classified organs -- while the eco graph holds 16,580 nodes. A ~614x substrate gap, and the saturation shows in discover's OWN output: 40 proposals with declared=1, trusted=7, GUESSED=20. Half of every proposal is a kind INFERRED from the organ name because the real catalog pairs are exhausted. A recombinator emitting 50pct guesses is not discovering, it is confabulating. Corroborated externally (July 2026): agent skill-selection accuracy undergoes a PHASE TRANSITION as library size grows -- routing over a FLAT catalog degrades combinatorially. The fix is the SUBSTRATE, not the volume. THIS ORGAN DOES NOT REPLACE nx_atlas_discover. Its header rationale still holds: "the eco graph sees STATIC import edges; plans compose tools at RUNTIME -- edges the graph cannot see." So this is the SIBLING mechanism: discover mines RUNTIME composability from the catalog, graphrec mines STRUCTURAL adjacency from the graph. Distinct ids (GRC vs DSC). METHOD -- classic link prediction (common neighbours) on the undirected import graph: candidate (a,b) iff NO direct edge a<->b AND |N(a) INTERSECT N(b)| >= minscore. A missing edge between two modules with many shared neighbours is the standard signal that the edge SHOULD exist -- i.e. a real composition opportunity, DERIVED not guessed. HUB SUPPRESSION: a witness whose degree exceeds GR_HUB_THR is skipped. Hubs (nx_syscalls and friends) are imported by nearly everything, so they are evidence of nothing; counting them would make every pair look related and rediscover the same saturation in a new costume. This is also what bounds the cost: sum over witnesses of deg^2 instead of all-pairs. PROVENANCE: every row is source=graph. There is no name-guessing path in this organ AT ALL, so its guess rate is 0 BY CONSTRUCTION -- that is the metric seq1489 says must fall. Proposals stay status=proposed: promotion still requires a plan run to PROVE the combo. REFUSAL, NOT A REMINDER: an unloadable or empty graph EXITS 2 rather than printing "0 proposals", which would read as "nothing to recombine" -- the exact lie to avoid. NO SILENT CAPS: witnesses skipped as hubs and pair-table overflow are both REPORTED. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0