nx_firstbyte.nx

nx_firstbyte.nx -- FIRST-BYTE-UP ANALYSIS FOR /compare: what does a rung REST ON, where does its chain stop being sound, and WHICH unsound link is the one you can actually act on? WHY THIS EXISTS (measured 2026-08-23). Five foundation defects were each found LATE, by a lane trying to build the rung ABOVE them: nx_skeleton capped at 32 bones / 2048 verts against a corpus of 104-370 joints and 14,164-423,919 verts; nx_gsplat pinned to a compile-time 512x384 while every inner kernel was ALREADY viewport-agnostic; PF_NB=48 binning an oracle holding 4096 points/section; the LANGUAGE having no vector types; and nothing able to bind a generated mesh to a generated skeleton. In EVERY case /compare showed the rung OPEN and never showed WHAT IT RESTS ON. *** NOT A SECOND GRAPH. *** nx_eco_graph_build already walks and persists the import graph, and eg_ancestors is "roots-to-god (transitive imports, down to leaves)" -- that IS the first-byte-up chain, and it had never been pointed at a compare board. This organ derives NO edges: it is the JOIN between <domain>.matrix rows and that closure, plus a per-link soundness query. *** THE CAP AXIS IS MEASURED. *** `nx_capcensus file <path>` answers per module IN ITS EXIT CODE -- 0 CLEAN / 1 CAPPED / 3 UNREADABLE -- so a foundation check BRANCHES WITHOUT PARSING. We compose that verdict; we never re-implement its classifier. Because the answer is the exit code and not the text, a truncated capture cannot change it. ABSTENTION IS INHERITED: the census refuses to report CLEAN over a module it never read, so this organ refuses READY over a chain containing such a module. *** RANKING: "NEAREST" IS NOT "MOST ACTIONABLE". *** v1 named the FIRST capped link found. With only 2.56% of estate bounds DERIVED, most chains carry SEVERAL capped links -- so v1 reported `first_unsound=nx_syscalls.nx` on rungs whose real problem was elsewhere: true, and completely non-actionable. A CAP IN A LEAF EVERYONE SHARES IS A DIFFERENT FINDING FROM A CAP IN YOUR OWN MODULE, AND REPORTING THE FIRST HIDES THE SECOND. Every capped link is now reported, classified and ranked: OWN the rung's own source -- the lane that owns this rung owns the fix. Ranked first. RARE few dependents -- a targeted fix. Ranked by ASCENDING fan-in, rarest first. UBIQUITOUS appears in EVERY resolved chain on this board -- an estate-wide concern, never a rung's blocker. Ranked last so it can never mask an actionable link again. BOTH CLASS BOUNDARIES ARE DERIVED, NOT PICKED. "Ubiquitous" is defined by the board itself (present in all N resolved chains, N counted this run), and "rare" is ordered by the graph's own fan-in via eg_ca -- the closure is already computed, so the dependent count is free. There is no threshold constant anywhere in this classification, because a picked one would be the seventh phony cap. A rung with ONE capped link in its own module and a rung with NINE scattered through shared leaves are different situations, and the per-rung own/rare/ubiq counts now say which is which. KNOWN ERROR TERM ON EVERY CAPPED VERDICT, disclosed by the census itself: it flagged CE_C_SHOWMAX=12 -- a COUNTER INDEX, not a capacity bound -- because the name contains "MAX".