fix(#2873): stub-mint undeclared edge endpoints, prune manifest deps, fix merge repos-union clobbering

graphify/build.py

@@ -1608,6 +1608,8 @@ def _dropped(item: dict) -> bool:
         # re-extracted its source. Hyperedges are semantic-tier (no _origin,
         # null source_location), so an AST-only re-extract carries them.
         # Deletion pruning below stays tier-blind.
+        if item.get("external"):
+            return False  # external stub nodes/edges (#2873): never carried-forward-drop, never replace-drop
         own = new_ast_sources if _is_ast_tier(item) else new_sem_sources
         if sf in own or _norm_source_file(sf, _eff_root) in own:
             return True  # re-extracted this run — replaced by the new chunk
@@ -2014,7 +2016,11 @@ def prefix_graph_for_global(
     collide and the aggregated community view fuses unrelated communities
     into one meta-node (#3014). 0 (the default) leaves communities untouched.
     """
-    relabel = {n: f"{repo_tag}::{n}" for n in G.nodes}
+    relabel = {
+        n: f"{repo_tag}::{n}"
+        for n, data in G.nodes(data=True)
+        if not data.get("external")
+    }
     H = nx.relabel_nodes(G, relabel, copy=True)
     for node, data in H.nodes(data=True):
         data["repo"] = repo_tag
@@ -2082,7 +2088,21 @@ def distinct_repo_tags(graph_paths: "list[Path]") -> "list[str]":
 
 
 def prune_repo_from_graph(G: nx.Graph, repo_tag: str) -> int:
-    """Remove all nodes tagged with repo_tag from G in-place. Returns count removed."""
-    to_remove = [n for n, d in G.nodes(data=True) if d.get("repo") == repo_tag]
+    """Remove all nodes tagged with repo_tag from G in-place. Returns count removed.
+
+    External nodes (#2873) can be referenced by multiple repos via their
+    'repos' list -- pruning one repo only removes the tag; the node itself
+    is removed only once no repo still references it.
+    """
+    to_remove = []
+    for n, d in G.nodes(data=True):
+        if d.get("repo") == repo_tag:
+            to_remove.append(n)
+        elif d.get("external") and repo_tag in d.get("repos", ()):
+            remaining = [r for r in d["repos"] if r != repo_tag]
+            if remaining:
+                d["repos"] = remaining
+            else:
+                to_remove.append(n)
     G.remove_nodes_from(to_remove)
     return len(to_remove)

graphify/extract.py

@@ -7009,6 +7009,45 @@ def _canon(nid: str) -> str:
             if e.get("target"):
                 e["target"] = _canon(e["target"])
 
+    # Every resolution/repoint/rewire/disambiguation pass above has now had its
+    # chance to land an edge endpoint on a real node id. Anything still not in
+    # the node set at this point is not a bug in one particular resolver — it
+    # is a genuine reference to something outside the graph (an unimported
+    # stdlib/third-party module, an unresolved dotted path, ...). Leaving it
+    # as a bare edge endpoint is not a neutral encoding of "outside the
+    # graph": every consumer that builds a graph object from nodes+edges
+    # (networkx included) materializes it as an attribute-less phantom node,
+    # so the file's declared node set and its produced node set disagree
+    # (#2873). Mint a minimal, explicitly-marked stub for each so they agree.
+    # `external: True` distinguishes these from ordinary file/symbol nodes —
+    # merge-graphs reads it to merge same-named external references across
+    # repos instead of namespacing them like repo-local ids (#2873).
+    _declared_ids = {n.get("id") for n in all_nodes if isinstance(n, dict)}
+    if resolution_context_nodes:
+        _declared_ids |= {
+            n.get("id") for n in resolution_context_nodes if isinstance(n, dict)
+        }
+    _external_stubs: dict[str, dict] = {}
+    for e in all_edges:
+        if not isinstance(e, dict):
+            continue
+        tgt = e.get("target")
+        if not tgt or tgt in _declared_ids or tgt in _external_stubs:
+            continue
+        if e.get("relation") in ("depends_on", "requires") or str(tgt).startswith("pkg_"):
+            continue  # manifest/package deps: keep prior prune-not-fabricate behavior (#2873)
+        _external_stubs[tgt] = {
+            "id": tgt,
+            "label": tgt,
+            "file_type": "code",
+            "type": "module",
+            "confidence": "INFERRED",
+            "external": True,
+            "source_file": None,
+        }
+    if _external_stubs:
+        all_nodes.extend(_external_stubs.values())
+
     # origin_file is an internal disambiguation hint (#1462): the colliding-id pass
     # above reads it to keep same-named cross-file stubs distinct, after which nothing
     # consumes it. Drop it from the returned nodes so it never ships into graph.json as

tests/test_csharp_type_resolution.py

@@ -252,19 +252,19 @@ def test_csharp_import_edges_resolve_internal_namespace_and_alias(tmp_path: Path
         (kind, fqn, target.get("type") if target else None)
         for kind, fqn, target in imports
     ]
-    assert ("namespace", "UnityEngine", None) in [
+    assert ("namespace", "UnityEngine", "module") in [
         (kind, fqn, target.get("type") if target else None)
         for kind, fqn, target in imports
     ]
     assert ("alias", "Game.Core.Damage", "Damage") in [
         (kind, fqn, target.get("label") if target else None)
         for kind, fqn, target in imports
     ]
-    assert ("alias", "System.Math", None) in [
+    assert ("alias", "System.Math", "system_math") in [
         (kind, fqn, target.get("label") if target else None)
         for kind, fqn, target in imports
     ]
-    assert ("static", "Game.Core.Damage", None) in [
+    assert ("static", "Game.Core.Damage", "game_core_damage") in [
         (kind, fqn, target.get("label") if target else None)
         for kind, fqn, target in imports
     ]

tests/test_global_graph.py

@@ -136,6 +136,22 @@ def test_prune_repo_returns_zero_if_not_present():
     assert G.number_of_nodes() == 1
 
 
+def test_prune_repo_preserves_shared_external_node_until_last_repo():
+    from graphify.build import prune_repo_from_graph
+    G = nx.Graph()
+    G.add_node("repoA::userservice", repo="repoA", label="UserService")
+    G.add_node("typing", external=True, repos=["repoA", "repoB"], label="typing")
+
+    removed = prune_repo_from_graph(G, "repoA")
+    assert removed == 1  # only repoA::userservice; typing still referenced by repoB
+    assert "typing" in G.nodes
+    assert G.nodes["typing"]["repos"] == ["repoB"]
+
+    removed = prune_repo_from_graph(G, "repoB")
+    assert removed == 1
+    assert "typing" not in G.nodes
+
+
 # ── global_graph.py ───────────────────────────────────────────────────────────
 
 def test_global_add_creates_global_graph(tmp_path):