diff --git a/buckaroo/server/handlers.py b/buckaroo/server/handlers.py
index bbbe93f67..55bdf538b 100644
--- a/buckaroo/server/handlers.py
+++ b/buckaroo/server/handlers.py
@@ -233,13 +233,28 @@ async def post(self):
         if session_id is None:
             return
 
+        # Optional backend="xorq" routes /load through the same push-down
+        # path as /load_expr but sourced from a parquet rather than a
+        # build_dir. Default "pandas" keeps existing callers unaffected.
+        backend_arg = body.get("backend", "pandas")
+        if backend_arg not in ("pandas", "xorq"):
+            self.set_status(400)
+            self.write({"error_code": "invalid_backend",
+                "message": f"backend must be 'pandas' or 'xorq' (got {backend_arg!r})"})
+            return
+        if backend_arg == "xorq" and mode != "buckaroo":
+            self.set_status(400)
+            self.write({"error_code": "invalid_mode_for_backend",
+                "message": "backend='xorq' requires mode='buckaroo' (XorqInfiniteBuckaroo)"})
+            return
+
         sessions = self.application.settings["sessions"]
         session = sessions.get_or_create(session_id, path)
         session.mode = mode
-        # Loading via /load is always pandas — clear any xorq state left
-        # by a prior /load_expr on the same session so WS dispatch routes
-        # to the new pandas dataflow rather than a stale xorq one.
-        session.backend = "pandas"
+        session.backend = backend_arg
+        # Reset any xorq/pandas state left by a prior load on the same
+        # session so WS dispatch routes to the new dataflow rather than a
+        # stale one. The branches below repopulate the relevant fields.
         session.xorq_dataflow = None
         session.expr = None
         # Reset the live-typed row-fetch filter so a search term carried
@@ -251,6 +266,81 @@ async def post(self):
         if component_config:
             session.component_config = component_config
 
+        if backend_arg == "xorq":
+            # XorqInfiniteBuckaroo over a materialised parquet. Mirrors
+            # LoadExprHandler's xorq-branch session setup but sourced from
+            # a file path instead of a build_dir.
+            try:
+                from buckaroo.server import xorq_loading  # noqa: PLC0415
+            except ImportError:
+                self.set_status(501)
+                self.write({"error_code": "xorq_not_installed",
+                    "message": "xorq is not installed on this server. "
+                    "Install with `pip install buckaroo[xorq]`."})
+                return
+
+            if not os.path.exists(path):
+                self.set_status(404)
+                self.write({"error_code": "file_not_found",
+                    "message": f"File not found: {path}"})
+                return
+
+            try:
+                expr = xorq_loading.load_expr_parquet_path(path)
+                xorq_dataflow = xorq_loading.XorqServerDataflow(
+                    expr, skip_main_serial=True)
+                metadata = xorq_loading.get_xorq_metadata(xorq_dataflow, path)
+            except Exception:
+                tb = traceback.format_exc()
+                log.error("load (xorq) error path=%s: %s", path, tb)
+                resp: dict = {"error_code": "load_error",
+                    "message": "Failed to load parquet via xorq backend"}
+                if _BUCKAROO_DEBUG:
+                    resp["details"] = tb
+                self.set_status(500)
+                self.write(resp)
+                return
+
+            session.expr = expr
+            session.xorq_dataflow = xorq_dataflow
+            session.df = None
+            session.dataflow = None
+            session.ldf = None
+            session.metadata = metadata
+            session.df_display_args = xorq_dataflow.df_display_args
+            session.df_data_dict = xorq_dataflow.df_data_dict
+            session.df_meta = xorq_dataflow.df_meta
+            session.buckaroo_state = {
+                "cleaning_method": "", "post_processing": "", "sampled": False,
+                "show_commands": False, "df_display": "main",
+                "search_string": "", "quick_command_args": {}}
+            session.buckaroo_options = xorq_dataflow.buckaroo_options
+            session.command_config = xorq_dataflow.command_config
+            session.operation_results = {
+                "transformed_df": {"schema": {"fields": []}, "data": []},
+                "generated_py_code": "# server mode (xorq backend via /load)"}
+            session.operations = []
+
+            if component_config and session.df_display_args:
+                for key in session.df_display_args:
+                    dvc = session.df_display_args[key].get("df_viewer_config")
+                    if dvc is not None:
+                        dvc["component_config"] = {
+                            **dvc.get("component_config", {}),
+                            **component_config,
+                        }
+
+            self._push_state_to_clients(session, metadata)
+            browser_action = "skipped" if no_browser else self._handle_browser_window(session_id)
+
+            log.info("load session=%s path=%s rows=%d backend=xorq browser=%s",
+                session_id, path, metadata["rows"], browser_action)
+            self.write({"session": session_id, "server_pid": os.getpid(),
+                "browser_action": browser_action, **metadata})
+            return
+
+        # Pandas-default / lazy-polars path. Identical to before — the
+        # session.backend assignment above already set "pandas".
         # Load data in appropriate mode
         file_obj, metadata = self._load_file_with_error_handling(path, is_lazy=(mode == "lazy"))
         if file_obj is None:
diff --git a/buckaroo/server/xorq_loading.py b/buckaroo/server/xorq_loading.py
index 5765920d2..9fada704d 100644
--- a/buckaroo/server/xorq_loading.py
+++ b/buckaroo/server/xorq_loading.py
@@ -82,6 +82,30 @@ def get_xorq_metadata(xorq_dataflow: XorqServerDataflow, build_dir: str) -> dict
     return {"path": build_dir, "rows": _expr_count(expr), "columns": columns}
 
 
+def load_expr_parquet_path(path: str):
+    """Wrap a local parquet file in a xorq deferred-read expression.
+
+    Counterpart to ``load_expr_build_dir`` for the case where the caller
+    holds a materialised parquet (e.g. a host that materialised a
+    catalog entry to a snapshot file via ``xorq catalog run``) and wants
+    XorqInfiniteBuckaroo's push-down query behaviour rather than the
+    eager pandas/polars load that ``data_loading.load_file`` does.
+
+    ``deferred_read_parquet`` wires the file into xorq's datafusion
+    backend without materialising it; ``XorqServerDataflow`` then takes
+    that expression and answers every ``infinite_request`` via
+    ``handle_infinite_request_xorq`` (the same code path the
+    build-dir-driven xorq loader uses).
+    """
+    from xorq.api import connect, deferred_read_parquet  # noqa: PLC0415
+    from xorq.vendor import ibis  # noqa: PLC0415
+
+    if ibis.options.default_backend is None:
+        ibis.options.default_backend = connect()
+    table_name = Path(path).stem.replace("-", "_") or "parquet"
+    return deferred_read_parquet(path, table_name=table_name)
+
+
 # ---------------------------------------------------------------------------
 # project-authored summary stats (loaded from <project_root>/stats/*.py)
 # ---------------------------------------------------------------------------
diff --git a/tests/unit/server/test_load_expr.py b/tests/unit/server/test_load_expr.py
index 73d32f7b9..a2e06923d 100644
--- a/tests/unit/server/test_load_expr.py
+++ b/tests/unit/server/test_load_expr.py
@@ -373,3 +373,140 @@ async def test_session_reuse_xorq_then_pandas(self):
         finally:
             shutil.rmtree(builds_root, ignore_errors=True)
             os.unlink(csv_path)
+
+
+class TestLoadBackendXorq(tornado.testing.AsyncHTTPTestCase):
+    """POST /load with backend="xorq" — wraps a parquet path in a
+    deferred-read xorq expression and serves it through the same
+    push-down path as /load_expr."""
+
+    def get_app(self):
+        return make_app()
+
+    def test_invalid_backend_returns_400(self):
+        resp = self.fetch("/load", method="POST",
+            body=json.dumps({"session": "lb-bad", "path": "/tmp/x.parquet",
+                "backend": "bogus"}),
+            headers={"Content-Type": "application/json"})
+        self.assertEqual(resp.code, 400)
+        self.assertEqual(json.loads(resp.body)["error_code"], "invalid_backend")
+
+    def test_xorq_backend_requires_buckaroo_mode(self):
+        resp = self.fetch("/load", method="POST",
+            body=json.dumps({"session": "lb-mode", "path": "/tmp/x.parquet",
+                "backend": "xorq", "mode": "lazy"}),
+            headers={"Content-Type": "application/json"})
+        self.assertEqual(resp.code, 400)
+        self.assertEqual(json.loads(resp.body)["error_code"],
+            "invalid_mode_for_backend")
+
+    @tornado.testing.gen_test
+    async def test_xorq_load_failure_preserves_session_backend(self):
+        """Codex P1 on #840: a failed xorq /load must not flip
+        session.backend to 'xorq' before the load is known to succeed.
+        Otherwise a session previously serving pandas ends up with
+        backend='xorq' + xorq_dataflow=None, and the WS state-change
+        handler silently drops further updates (websocket_handler.py:59)."""
+        d = tempfile.mkdtemp()
+        good_path = os.path.join(d, "good.parquet")
+        bogus_path = os.path.join(d, "nonexistent.parquet")
+        try:
+            pd.DataFrame({"idx": list(range(3))}).to_parquet(good_path)
+
+            # 1. Successful pandas load establishes prior session state.
+            resp = await _post(self.get_http_port(), "/load",
+                {"session": "lb-roll", "path": good_path, "mode": "buckaroo"})
+            self.assertEqual(resp.code, 200)
+
+            sessions = self._app.settings["sessions"]
+            session = sessions.get("lb-roll")
+            self.assertEqual(session.backend, "pandas")
+            prior_dataflow = session.dataflow
+            self.assertIsNotNone(prior_dataflow)
+
+            # 2. Failed xorq load (path doesn't exist → 404).
+            resp = await _post(self.get_http_port(), "/load",
+                {"session": "lb-roll", "path": bogus_path,
+                 "mode": "buckaroo", "backend": "xorq"})
+            self.assertEqual(resp.code, 404)
+
+            # 3. Session must not have been half-mutated. backend stays
+            # 'pandas'; the existing pandas dataflow stays reachable so
+            # WS dispatch can still answer buckaroo_state_change.
+            self.assertEqual(session.backend, "pandas",
+                "session.backend must not flip to 'xorq' on failed xorq load")
+            self.assertIs(session.dataflow, prior_dataflow,
+                "session.dataflow must survive a failed xorq load")
+            self.assertIsNone(session.xorq_dataflow)
+        finally:
+            shutil.rmtree(d, ignore_errors=True)
+
+    def test_xorq_not_installed_returns_501(self):
+        """Codex P2 on #840: when xorq.api is not importable the handler
+        must return 501 xorq_not_installed, not a generic 500 load_error.
+        The probe in handlers.py has to be explicit — importing
+        ``buckaroo.server.xorq_loading`` succeeds even without xorq
+        because the transitive ``import xorq.api`` calls in
+        ``xorq_stats_v2`` and ``xorq_stat_pipeline`` are guarded with
+        try/except.
+
+        Uses a real parquet path so the path-exists check passes; that
+        way a 500 result here proves the bug isn't masked as file-not-
+        found and is genuinely the unreachable 501 branch."""
+        from unittest.mock import patch
+        d = tempfile.mkdtemp()
+        parquet_path = os.path.join(d, "p.parquet")
+        try:
+            pd.DataFrame({"idx": [0]}).to_parquet(parquet_path)
+            # sys.modules[name] = None forces a subsequent `import name`
+            # to raise ImportError. patch.dict restores the mapping so
+            # other tests still see the real xorq.api.
+            with patch.dict(sys.modules, {"xorq.api": None}):
+                resp = self.fetch("/load", method="POST",
+                    body=json.dumps({"session": "lb-noxorq",
+                        "path": parquet_path,
+                        "mode": "buckaroo", "backend": "xorq"}),
+                    headers={"Content-Type": "application/json"})
+            self.assertEqual(resp.code, 501)
+            self.assertEqual(json.loads(resp.body)["error_code"],
+                "xorq_not_installed")
+        finally:
+            shutil.rmtree(d, ignore_errors=True)
+
+    @tornado.testing.gen_test
+    async def test_load_parquet_via_xorq_backend(self):
+        """Happy path: POST /load with backend=xorq, then issue an
+        infinite_request — the row count must come from the parquet via
+        XorqServerDataflow, proving routing went through the xorq path."""
+        d = tempfile.mkdtemp()
+        parquet_path = os.path.join(d, "lb_fixture.parquet")
+        try:
+            pd.DataFrame({"idx": list(range(7)), "name": ["a", "b", "c", "d", "e", "f", "g"]}).to_parquet(parquet_path)
+
+            resp = await _post(self.get_http_port(), "/load",
+                {"session": "lb-ok", "path": parquet_path,
+                 "mode": "buckaroo", "backend": "xorq"})
+            self.assertEqual(resp.code, 200)
+            body = json.loads(resp.body)
+            self.assertEqual(body["session"], "lb-ok")
+            self.assertEqual(body["rows"], 7)
+            self.assertEqual({c["name"] for c in body["columns"]}, {"idx", "name"})
+
+            ws = await tornado.websocket.websocket_connect(
+                f"ws://localhost:{self.get_http_port()}/ws/lb-ok")
+            await ws.read_message()  # initial_state
+
+            ws.write_message(json.dumps({
+                "type": "infinite_request",
+                "payload_args": {"start": 0, "end": 10,
+                    "sourceName": "default", "origEnd": 10}}))
+            r = json.loads(await ws.read_message())
+            self.assertEqual(r["type"], "infinite_resp")
+            self.assertEqual(r["length"], 7)
+
+            binary_frame = await ws.read_message()
+            table = pq.read_table(io.BytesIO(binary_frame))
+            self.assertEqual(table.num_rows, 7)
+            ws.close()
+        finally:
+            shutil.rmtree(d, ignore_errors=True)