diff --git a/cortex_viz/server/graph_build_run.py b/cortex_viz/server/graph_build_run.py
index 7fbb757..d8600b8 100644
--- a/cortex_viz/server/graph_build_run.py
+++ b/cortex_viz/server/graph_build_run.py
@@ -542,13 +542,22 @@ def _on_batch(label: str, nodes_objs, edges_objs) -> None:
         # early return, or error. Subscribers drain whatever was
         # emitted, receive ``done``, and close. close() is
         # idempotent; the buffer survives for late-subscriber
-        # replay until the next build's reset().
+        # replay until the next build's reset(). A failure here must
+        # not mask the build outcome being unwound, so it is never
+        # re-raised — but not-masking and not-reporting are
+        # different things (#135): silently swallowing it is exactly
+        # why #134's AttributeError on this same call went unnoticed
+        # for a release. Report it on stderr instead; subscribers
+        # still never receive ``done`` and are left on an open
+        # connection, but now there is a line naming why.
         try:
             from cortex_viz.server import graph_event_stream as _ev
 
             _ev.close()
-        except Exception:
-            # close() is idempotent and the buffer survives for late-subscriber replay;
-            # a failure here must not mask the build outcome being unwound.
-            pass
+        except Exception as _close_exc:
+            print(
+                f"[cortex] end-of-stream terminator failed: "
+                f"{type(_close_exc).__name__}: {_close_exc}",
+                file=sys.stderr,
+            )
         state._graph_build_lock.release()
diff --git a/tests/test_graph_build_run_terminator.py b/tests/test_graph_build_run_terminator.py
new file mode 100644
index 0000000..c7cd571
--- /dev/null
+++ b/tests/test_graph_build_run_terminator.py
@@ -0,0 +1,69 @@
+"""#135: the end-of-build SSE terminator must report a failure, never mask it.
+
+``run_build``'s ``finally`` sends the single ``close()`` end-of-stream call.
+#134 made ``close()`` resolve again; this file exercises what happens the
+*next* time it raises — the handler must not re-raise (a failure here must
+not mask the build outcome being unwound) but must no longer swallow the
+failure silently either, since that silence is exactly why #134's
+``AttributeError`` on this call survived a release unnoticed.
+
+``run_build`` is exercised directly (not through a fake), forcing the very
+first line of its ``try`` to fail so the test reaches ``finally`` without
+depending on a real store or database.
+
+``graph_event_stream.close`` is patched with ``raising=False``: this file is
+independent of #134 (a separate PR against the same base), and whether the
+module-level ``close`` forwarder exists yet depends on merge order, not on
+this fix. Both scenarios must behave identically here — the finally block
+must report a failing terminator regardless of *why* it failed.
+"""
+
+from __future__ import annotations
+
+from cortex_viz.server import graph_build_run
+from cortex_viz.server import graph_cache_state as state
+from cortex_viz.server import graph_event_stream as ges
+
+
+def _run_with_forced_early_failure(monkeypatch):
+    """Force ``run_build``'s try block to fail on its first statement, so
+    the finally block is reached without a store, a DB, or a real build."""
+
+    def _boom():
+        raise RuntimeError("forced early build failure")
+
+    monkeypatch.setattr(graph_build_run, "_roster_fingerprint", _boom)
+    state._graph_build_lock.acquire()
+    graph_build_run.run_build(store=object(), domain_filter=None)
+
+
+def test_terminator_failure_is_reported_on_stderr(monkeypatch, capsys):
+    monkeypatch.setattr(
+        ges,
+        "close",
+        lambda: (_ for _ in ()).throw(AttributeError("no attribute 'close'")),
+        raising=False,
+    )
+
+    _run_with_forced_early_failure(monkeypatch)
+
+    err = capsys.readouterr().err
+    assert "end-of-stream terminator failed" in err
+    assert "AttributeError" in err
+    assert "no attribute 'close'" in err
+    # Not re-raised: the build's own error report is still the thing that
+    # reaches stderr as the build outcome, and the lock is still released.
+    assert "background build error" in err
+    assert not state._graph_build_lock.locked()
+
+
+def test_nominal_terminator_stays_quiet(monkeypatch, capsys):
+    closed = []
+    monkeypatch.setattr(ges, "close", lambda: closed.append(True), raising=False)
+
+    _run_with_forced_early_failure(monkeypatch)
+
+    err = capsys.readouterr().err
+    assert closed == [True]
+    assert "end-of-stream terminator failed" not in err
+    assert not state._graph_build_lock.locked()