fix(mcp): drain KG cache on tool_reconnect

tool_reconnect cleared ChromaDB caches but left _kg_by_path entries
intact. After an external replacement of knowledge_graph.sqlite3 the
server kept serving the old open sqlite3.Connection, returning stale
results.

Now iterate _kg_by_path under _kg_cache_lock, call close() best-effort,
and clear the dict so the next tool call reopens the KG from disk.
Two new tests in TestKGLazyCache verify cache invalidation and that a
failing close() does not block the clear.

mempalace/mcp_server.py

@@ -1418,10 +1418,11 @@ def tool_memories_filed_away():
 
 
 def tool_reconnect():
-    """Force the MCP server to drop the cached ChromaDB collection and reconnect.
+    """Force the MCP server to drop cached ChromaDB + KnowledgeGraph state.
 
     Use after external scripts or CLI commands modify the palace database
-    directly, which can leave the in-memory HNSW index stale.
+    or replace ``knowledge_graph.sqlite3`` directly, which can leave the
+    in-memory HNSW index stale or pin a closed-on-disk SQLite connection.
     """
     global \
         _client_cache, \
@@ -1439,6 +1440,15 @@ def tool_reconnect():
     # still applies after the reconnect.
     _vector_disabled = False
     _vector_disabled_reason = ""
+    # Drain the per-path KnowledgeGraph cache so a replaced sqlite file is
+    # reopened on the next tool call rather than served from a stale handle.
+    with _kg_cache_lock:
+        for kg in _kg_by_path.values():
+            try:
+                kg.close()
+            except Exception:
+                pass
+        _kg_by_path.clear()
     try:
         col = _get_collection()
         if col is None:

tests/test_mcp_server.py

@@ -1253,3 +1253,45 @@ class TestKGLazyCache:
         ids = {id(kg) for kg in results}
         assert len(ids) == 1, f"expected 1 unique instance, got {len(ids)}"
         assert len(mcp_server._kg_by_path) == 1
+
+    def test_tool_reconnect_drains_kg_cache(self, monkeypatch):
+        """``tool_reconnect`` must close cached KG instances and clear the dict.
+
+        Without this, an external replacement of ``knowledge_graph.sqlite3``
+        leaves the server pinned to a stale ``sqlite3.Connection``.
+        """
+        from mempalace import mcp_server
+
+        class _FakeKG:
+            def __init__(self):
+                self.closed = False
+
+            def close(self):
+                self.closed = True
+
+        fake_a = _FakeKG()
+        fake_b = _FakeKG()
+        monkeypatch.setattr(mcp_server, "_kg_by_path", {"/a": fake_a, "/b": fake_b})
+        # Bypass real ChromaDB so the test isolates KG-cache behaviour.
+        monkeypatch.setattr(mcp_server, "_get_collection", lambda: None)
+
+        mcp_server.tool_reconnect()
+
+        assert fake_a.closed is True
+        assert fake_b.closed is True
+        assert mcp_server._kg_by_path == {}
+
+    def test_tool_reconnect_swallows_kg_close_errors(self, monkeypatch):
+        """A failing ``close()`` on one cached KG must not block cache clearing."""
+        from mempalace import mcp_server
+
+        class _BoomKG:
+            def close(self):
+                raise RuntimeError("boom")
+
+        monkeypatch.setattr(mcp_server, "_kg_by_path", {"/a": _BoomKG()})
+        monkeypatch.setattr(mcp_server, "_get_collection", lambda: None)
+
+        mcp_server.tool_reconnect()
+
+        assert mcp_server._kg_by_path == {}