diff --git a/graphify/dedup.py b/graphify/dedup.py
index e0ddd2e3b..4dc099e85 100644
--- a/graphify/dedup.py
+++ b/graphify/dedup.py
@@ -9,8 +9,10 @@
 import sys
 import unicodedata
 from collections import defaultdict
+from pathlib import Path
 
 from graphify._minhash import MinHash, MinHashLSH
+from graphify.ids import make_id, normalize_id
 from rapidfuzz.distance import Jaro, JaroWinkler
 
 
@@ -187,6 +189,42 @@ def _is_code(node: dict) -> bool:
     return node.get("file_type") == "code"
 
 
+def _node_defines_id(node: dict, nid: str) -> bool:
+    """True when ``node``'s own ``source_file`` is the file encoded in ``nid``.
+
+    Non-AST node ids are ``make_id(_file_stem(source_file), <label>)``, so the
+    entity's *defining* node has an id equal to its file stem (the file node) or
+    beginning with ``stem + "_"`` (a symbol in that file). A node that merely
+    *references* the entity was extracted from a different file, so its own
+    source_file yields a different stem and this returns False — which is how we
+    tell the authoritative node from a cross-reference that collided onto the
+    same id (#1851)."""
+    sf = node.get("source_file")
+    if not sf:
+        return False
+    from graphify.extractors.base import _file_stem  # local: avoid import cost at module load
+
+    try:
+        stem = make_id(_file_stem(Path(str(sf))))
+    except ValueError:
+        return False
+    if not stem:
+        return False
+    norm_nid = normalize_id(nid)
+    return norm_nid == stem or norm_nid.startswith(stem + "_")
+
+
+def _materially_differs(a: dict, b: dict) -> bool:
+    """True when two same-id nodes carry different content worth surfacing.
+
+    Compares the normalized label and the description: if either differs, keeping
+    one and dropping the other silently discards information. Same-entity copies
+    (identical label and description) return False so the merge stays quiet."""
+    if _norm(a.get("label")) != _norm(b.get("label")):
+        return True
+    return (a.get("description") or "").strip() != (b.get("description") or "").strip()
+
+
 # ── main entry point ──────────────────────────────────────────────────────────
 
 def deduplicate_entities(
@@ -220,30 +258,63 @@ def deduplicate_entities(
     if len(nodes) <= 1:
         return nodes, edges
 
-    # Pre-deduplicate: keep first occurrence of each id.
-    # Warn when two nodes share an ID but originate from different source files —
-    # this indicates a cross-chunk ID collision (#1504) where silent data loss occurs.
-    seen_ids: dict[str, dict] = {}
+    # Pre-deduplicate: collapse the nodes sharing each id down to one survivor.
+    #
+    # Non-AST node ids are '<slug(source_file_stem)>_<slug(label)>', so a doc that
+    # *references* an entity can mint the same id as the entity's own node, and the
+    # two collide by construction (#1504). Two things follow:
+    #   * Survivor selection prefers the node whose own source_file defines the id
+    #     (see _node_defines_id), falling back to first-seen order — deterministic,
+    #     and independent of the chunk-merge order (#1851).
+    #   * We warn only when a *dropped* node would actually lose information — a
+    #     different label/description, or a genuine cross-file collision — instead
+    #     of on every source_file mismatch. Same-file duplicates that differ are
+    #     now surfaced (previously silent), and same-entity copies stay quiet.
+    id_groups: dict[str, list[dict]] = {}
     for node in nodes:
         nid = node.get("id", "")
         if not nid:
             continue
-        if nid not in seen_ids:
-            seen_ids[nid] = node
-        else:
-            existing_sf = seen_ids[nid].get("source_file") or ""
-            new_sf = node.get("source_file") or ""
-            if existing_sf != new_sf:
+        id_groups.setdefault(nid, []).append(node)
+
+    unique_nodes: list[dict] = []
+    for nid, group in id_groups.items():
+        survivor = next((n for n in group if _node_defines_id(n, nid)), group[0])
+        unique_nodes.append(survivor)
+        if len(group) == 1:
+            continue
+        survivor_sf = survivor.get("source_file") or ""
+        for dropped in group:
+            if dropped is survivor:
+                continue
+            dropped_sf = dropped.get("source_file") or ""
+            cross_file = dropped_sf != survivor_sf
+            if not cross_file and not _materially_differs(survivor, dropped):
+                continue  # redundant same-file copy — nothing lost, stay quiet
+            drop_label = dropped.get("label") or nid
+            keep_label = survivor.get("label") or nid
+            if cross_file:
+                print(
+                    f"[graphify] WARNING: node id '{nid}' is emitted from more than "
+                    f"one file — keeping '{keep_label}' (from '{survivor_sf or '?'}') "
+                    f"and dropping '{drop_label}' (from '{dropped_sf or '?'}'). Node "
+                    f"ids are '<source-path>_<label>', so a doc that references an "
+                    f"entity reuses the id of the entity's own node; when it is the "
+                    f"same entity, edges are preserved and only the dropped node's "
+                    f"label/description is lost. If these are genuinely different "
+                    f"entities sharing a path stem, extract each subfolder separately "
+                    f"and combine with 'graphify merge-graphs'.",
+                    file=sys.stderr,
+                )
+            else:
                 print(
-                    f"[graphify] WARNING: node '{nid}' from '{new_sf}' collides with "
-                    f"node from '{existing_sf}' — the second node will be dropped. "
-                    f"This is a cross-chunk ID collision caused by two files with the "
-                    f"same name in different directories. To avoid data loss, run "
-                    f"'graphify extract' per subfolder and merge with "
-                    f"'graphify merge-graphs'.",
+                    f"[graphify] WARNING: '{survivor_sf or '?'}' produced multiple "
+                    f"nodes with id '{nid}' but different content — keeping "
+                    f"'{keep_label}' and dropping '{drop_label}'. Edges are preserved "
+                    f"(they key on the id); the dropped node's label/description is "
+                    f"discarded.",
                     file=sys.stderr,
                 )
-    unique_nodes = list(seen_ids.values())
 
     if len(unique_nodes) <= 1:
         return unique_nodes, edges
diff --git a/tests/test_dedup.py b/tests/test_dedup.py
index 267388e57..845866358 100644
--- a/tests/test_dedup.py
+++ b/tests/test_dedup.py
@@ -366,12 +366,14 @@ def test_dedup_still_merges_crossfile_true_duplicates():
     assert len(result_nodes) == 1
 
 
-# ── #1504: cross-chunk node ID collision warning ──────────────────────────────
+# ── #1504 / #1851: cross-chunk node ID collision warning ──────────────────────
 
 def test_cross_chunk_id_collision_emits_warning(capsys):
-    """When two nodes share the same ID but come from different source files
-    (a cross-chunk LLM ID collision), a WARNING must be printed to stderr
-    and only the first node survives (#1504)."""
+    """Two nodes sharing an ID but coming from different source files
+    (a cross-chunk collision, #1504) still warn, and only one survives. Neither
+    file defines the id, so first-seen order is the tiebreak. The reworded
+    message must state the id-scheme cause rather than the old — frequently
+    wrong — 'same name in different directories' claim (#1851)."""
     nodes = [
         {"id": "readme_booking_service", "label": "Booking Service",
          "file_type": "concept", "source_file": "module-a/README.md"},
@@ -388,15 +390,18 @@ def test_cross_chunk_id_collision_emits_warning(capsys):
     assert "readme_booking_service" in captured.err
     assert "module-b/README.md" in captured.err
     assert "module-a/README.md" in captured.err
+    # #1851: drop the misdiagnosis; a genuine cross-file drop still offers the workaround.
+    assert "same name in different directories" not in captured.err
+    assert "merge-graphs" in captured.err
 
 
-def test_same_id_same_source_file_no_warning(capsys):
-    """When two nodes share both ID and source_file (same-file dedup),
-    no collision warning should be emitted."""
+def test_same_id_same_source_file_identical_no_warning(capsys):
+    """Same ID, same source_file, identical content — a redundant copy with
+    nothing to lose — stays quiet."""
     nodes = [
         {"id": "readme_booking_service", "label": "Booking Service",
          "file_type": "concept", "source_file": "module-a/README.md"},
-        {"id": "readme_booking_service", "label": "Booking Service (dupe)",
+        {"id": "readme_booking_service", "label": "Booking Service",
          "file_type": "concept", "source_file": "module-a/README.md"},
     ]
     result_nodes, _ = deduplicate_entities(nodes, [], communities={})
@@ -404,3 +409,43 @@ def test_same_id_same_source_file_no_warning(capsys):
     assert len(result_nodes) == 1
     captured = capsys.readouterr()
     assert "WARNING" not in captured.err
+
+
+def test_same_id_same_source_file_different_content_warns(capsys):
+    """#1851: two same-ID nodes from the *same* file but with different
+    labels/descriptions were previously dropped silently, discarding a
+    description unseen. This must now warn — without pushing the per-subfolder
+    re-extraction workaround, which is irrelevant within a single file."""
+    nodes = [
+        {"id": "readme_booking_service", "label": "Booking Service",
+         "description": "Handles reservations.",
+         "file_type": "concept", "source_file": "module-a/README.md"},
+        {"id": "readme_booking_service", "label": "Booking Service helper",
+         "description": "A different summary entirely.",
+         "file_type": "concept", "source_file": "module-a/README.md"},
+    ]
+    result_nodes, _ = deduplicate_entities(nodes, [], communities={})
+
+    assert len(result_nodes) == 1
+    captured = capsys.readouterr()
+    assert "WARNING" in captured.err
+    assert "readme_booking_service" in captured.err
+    assert "merge-graphs" not in captured.err
+
+
+def test_collision_survivor_prefers_defining_file():
+    """#1851: when a cross-reference node collides with the entity's own node,
+    the node defined by the file encoded in the id survives regardless of which
+    one was seen first (removing the old insertion-order dependence)."""
+    defining = {"id": "agents_make_batch_fixtures_make_batch_fixtures",
+                "label": "make-batch-fixtures agent", "file_type": "concept",
+                "source_file": "agents/make-batch-fixtures.md"}
+    reference = {"id": "agents_make_batch_fixtures_make_batch_fixtures",
+                 "label": "make-batch-fixtures", "file_type": "concept",
+                 "source_file": "available/diagnose-issue/SKILL.md"}
+
+    # Reference chunk merged first — the defining node must still win.
+    result_nodes, _ = deduplicate_entities([reference, defining], [], communities={})
+    assert len(result_nodes) == 1
+    assert result_nodes[0]["source_file"] == "agents/make-batch-fixtures.md"
+    assert result_nodes[0]["label"] == "make-batch-fixtures agent"