diagnose_extraction counts external-package import refs (ref_*) as dangling edges — false "graph may be incomplete/corrupt" warning on JS/TS projects

[jj954081-dotcom]

Environment

• graphifyy 0.9.26, Python 3.12.10, Windows 11
• Project: pnpm monorepo — Next.js 16 storefront + Strapi 5 CMS, 230 code files (TS/TSX), node_modules excluded (default)

What happens

Step 4.5 of the skill pipeline (graphify.diagnostics.diagnose_extraction) reports:

dangling_endpoint_edges: 217
GRAPH HEALTH WARNING: 217 dangling-endpoint edges - graph may be incomplete/corrupt.

Breaking those 217 edges down by missing endpoint shows they are all AST import references to external npm packages, for which no nodes can exist because node_modules is (correctly) excluded:

 58  ref_strapi_strapi      (@strapi/strapi)
 39  ref_next
 38  ref_next_link
 33  ref_react
 13  ref_next_navigation
  4  ref_next_headers
  3  ref_path
  3  ref_strapi_strapi_admin
 ...

So the warning fires on what is arguably every ordinary JS/TS project, and re-extraction cannot fix it — the count stays identical run to run. In our case the LLM-extracted semantic edges were 100% valid; all "dangling" edges came from the deterministic AST layer's external-import refs. We spent a full forced re-extraction chasing this before isolating the cause.

The build layer already handles this correctly (these edges are excluded from graph.json), so the graph itself is fine — only the diagnostic message is misleading.

Steps to reproduce

1. Any Next.js/React/Strapi repo, run the standard /graphify . pipeline.
2. Run diagnose_extraction(extraction, directed=False, root='.') on the merged extraction.
3. Observe dangling_endpoint_edges ≈ (number of external-module import edges), plus the corrupt-graph warning.

Expected / suggested

• Classify edges whose missing endpoint matches the external-ref scheme (ref_*, or "endpoint not under scan_root") into a separate counter, e.g. external_reference_edges, excluded from the GRAPH HEALTH WARNING; or
• resolve them to synthetic "external package" nodes so they are not dangling at all;
• keep the warning for genuinely dangling endpoints (typo'd IDs, cross-chunk drift) — that signal is valuable, it is just drowned out today.