Nishi Atlas — Quality & SOTA-Gap Audit

Hard evidence, measured by nx_eco_graph_honesty over the live ecograph_full graph (16,575 nodes / 42,792 edges) · 2026-07-18 · every number is a live audit count, re-run to ratchet

Islands — junk floating in the graph (124)

Organs not even undirected-connected to the main component — overwhelmingly test fixtures and scratch files polluting the atlas: hello.nx · z.nx · x.nx · y.nx · _min42.nx · _tr_ref.nx · _tv_ref.nx · nx_typesafe_bad.nx · nx_ev_bad.nx · nx_spore_syscalls.nx · nx_race_kernel.nx · nx_emit_kat.nx …

Sprawl — abandoned / unexplored research wings

Childless non-test organ clusters = the sprawl, quantified. This is where "it appears bad" is real and measurable.

Over-coupling — organs that import too much (123)

Fan-out outliers = refactor targets: nx_wiki_routes (22) · nx_wiki_main (19) · nx_render_html (19) · nx_tls13_server_session_run (16) · nx_opaque_core (16)

Worklist — ratchet these DOWN

1. 124 islands → remove/connect test-junk floaters (they are scratch fixtures, not capabilities).
2. Sprawl wings → triage the abandoned/duplicate clusters (_offc, nx_connect, nx_builder …).
3. 123 over-coupled → refactor the fan-out outliers.
4. Honest blocker: there is no rm/tombstone MCP tool — the janitor cannot remove atlas junk API-first yet. And the atlas is source-derived (nx_eco_graph_build), so removal = delete the .nx source then rebuild. Filed.