diff --git a/graphify/build.py b/graphify/build.py
index a8e69899f..64fd6c556 100644
--- a/graphify/build.py
+++ b/graphify/build.py
@@ -450,10 +450,11 @@ def _normalize_hyperedge_members(he: object) -> None:
 def _fold_node_aliases(node: dict) -> None:
     """Fold legacy node field aliases onto canonical keys, in place (#2194).
 
-    ``name`` -> ``label`` and ``path`` -> ``source_file``. Uses an empty-check
-    (not mere key presence) so a node carrying ``label: ""``/``None`` next to a
-    real ``name`` is healed too. When the canonical field already holds a value
-    it wins and the alias key is left untouched. Without this fold an alias-only
+    ``name`` -> ``label`` and ``file``/``path``/``filepath`` -> ``source_file``
+    (first non-blank alias in that order wins). Uses an empty-check (not mere
+    key presence) so a node carrying ``label: ""``/``None`` next to a real
+    ``name`` is healed too. When the canonical field already holds a value it
+    wins and the alias key is left untouched. Without this fold an alias-only
     node enters the graph with no label/source_file: it fails validation, gets
     ``norm_label == ""`` (invisible to query/explain), and is excluded from every
     label-keyed merge/dedup — a permanent ghost that ``graphify update``
@@ -472,9 +473,10 @@ def _fold_node_aliases(node: dict) -> None:
 def _fold_edge_aliases(edge: dict) -> None:
     """Fold legacy edge field aliases onto canonical keys, in place (#2194).
 
-    ``type`` -> ``relation``. A ``confidence_score`` float with no ``confidence``
-    enum backfills ``confidence: "INFERRED"`` — never EXTRACTED (alias recovery
-    is not provenance) and never a threshold mapping of the float. The
+    ``type``/``relationship`` -> ``relation`` (``type`` checked first). A
+    ``confidence_score`` float with no ``confidence`` enum backfills
+    ``confidence: "INFERRED"`` — never EXTRACTED (alias recovery is not
+    provenance) and never a threshold mapping of the float. The
     ``confidence_score`` key itself is NOT popped: it is a legitimate companion
     field that the edge loop sanitizes and to_json round-trips.
 
@@ -1830,7 +1832,7 @@ def build(
         # coerces too, but that runs after dedup — too late for this path.
         _coerce_non_string_ids(combined)
         # Fold legacy node field aliases before dedup (#2194): dedup runs BEFORE
-        # build_from_json and keys on `label`, so a `name`/`path` alias node
+        # build_from_json and keys on `label`, so a `name`/`file`/`path` alias node
         # would be invisible to it and only label-dedup one build later, after
         # build_from_json's own fold has healed the persisted graph.json.
         for n in combined["nodes"]:
diff --git a/tests/test_lost_saved_semantic_nodes.py b/tests/test_lost_saved_semantic_nodes.py
index 8312b7bab..3926074a4 100644
--- a/tests/test_lost_saved_semantic_nodes.py
+++ b/tests/test_lost_saved_semantic_nodes.py
@@ -2,7 +2,12 @@
 (#1145 extended) is preserved, not lost."""
 from __future__ import annotations
 
-from graphify.build import _lost_saved_semantic_nodes
+import json
+from pathlib import Path
+
+from graphify.build import _lost_saved_semantic_nodes, build_from_json
+from graphify.export import to_json
+from graphify.watch import _check_shrink
 
 
 def _sem(node_id: str, label: str, sf: str) -> dict:
@@ -99,3 +104,39 @@ def test_malformed_non_string_source_file_does_not_crash():
     new = [{"id": "n1", "label": "N1", "source_file": True}]
     lost = _lost_saved_semantic_nodes(saved, new)
     assert [n["id"] for n in lost] == ["s1"]
+
+
+# --- end-to-end: the write paths that call the guard (#4389) ---
+
+
+def _saved_graph(tmp_path: Path, sem: dict) -> tuple[Path, dict]:
+    out = tmp_path / "graph.json"
+    assert to_json(build_from_json({"nodes": [sem], "edges": []}), {}, str(out), force=True)
+    return out, json.loads(out.read_text(encoding="utf-8"))
+
+
+def _node_data(G) -> dict:
+    return {"nodes": [{"id": nid, **attrs} for nid, attrs in G.nodes(data=True)], "links": []}
+
+
+def test_to_json_writes_after_ghost_dedup(tmp_path):
+    sem = _sem("src_models_user_ts_user", "User", "src/models/user.ts")
+    out, _ = _saved_graph(tmp_path, sem)
+    G = build_from_json({"nodes": [sem, _ast("user_user", "User", "src/models/user.ts")], "edges": []})
+    assert list(G.nodes) == ["user_user"]  # ghost id folded into its AST twin
+    assert to_json(G, {}, str(out)) is True
+
+
+def test_to_json_still_refuses_genuine_loss(tmp_path, capsys):
+    sem = _sem("src_models_user_ts_user", "User", "src/models/user.ts")
+    out, _ = _saved_graph(tmp_path, sem)
+    G = build_from_json({"nodes": [_ast("user_account", "Account", "src/models/user.ts")], "edges": []})
+    assert to_json(G, {}, str(out)) is False
+    assert "src_models_user_ts_user" in capsys.readouterr().err
+
+
+def test_check_shrink_allows_ghost_dedup(tmp_path):
+    sem = _sem("src_models_user_ts_user", "User", "src/models/user.ts")
+    _, existing = _saved_graph(tmp_path, sem)
+    G = build_from_json({"nodes": [sem, _ast("user_user", "User", "src/models/user.ts")], "edges": []})
+    assert _check_shrink(force=False, existing_data=existing, new_data=_node_data(G)) is True