fix(ml-backend): 推理分批，避免 Retrieve Predictions 撑爆显存；新增批量推送脚本

Retrieve Predictions 会把选中的任务一次发给后端，而后端把整个列表交给
model.predict(source=paths)——ultralytics 不理会 batch 参数，会当成一个 batch，
于是显存爆掉。这与之前 visualize_detections 踩的是同一个坑。后端现在按
YOLO_BATCH（默认 16）自己分批。

另加 scripts/push_predictions.py：重训完在本地分批推理，再用批量导入接口
POST /api/projects/<id>/import/predictions 提交，不经过后端服务，显存与请求体都可控。
支持 --replace 先删同版本旧预测（预测是追加的，重复跑会堆两份）、--limit 与 --dry-run。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_push_predictions.py

@@ -0,0 +1,88 @@
+"""批量给 Label Studio 推预测：URL 还原与分页拉取。"""
+
+import sys
+
+import pytest
+
+pytest.importorskip("requests")
+
+sys.path.insert(0, "scripts")
+import push_predictions as pp  # noqa: E402
+
+
+def test_local_path_from_local_files_url(tmp_path):
+    url = "/data/local-files/?d=boxes_shap_label_studio/images/blackdot_0489.png"
+    assert pp.local_path(url, tmp_path) == (
+        tmp_path / "boxes_shap_label_studio/images/blackdot_0489.png"
+    )
+
+
+def test_local_path_unquotes_escaped_names(tmp_path):
+    """文件名里有中文或空格时 URL 是转义过的，得还原回来。"""
+    url = "/data/local-files/?d=proj/images/%E5%8F%8D%E9%9D%A2%20a.png"
+    assert pp.local_path(url, tmp_path) == tmp_path / "proj/images/反面 a.png"
+
+
+def test_local_path_rejects_non_local_urls(tmp_path):
+    with pytest.raises(ValueError, match="本地存储"):
+        pp.local_path("https://example.com/a.png", tmp_path)
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self._payload = payload
+        self.status_code = status
+
+    def json(self):
+        return self._payload
+
+    def raise_for_status(self):
+        if self.status_code >= 400:
+            raise AssertionError(f"HTTP {self.status_code}")
+
+
+class _FakeSession:
+    """按页返回任务，记录请求参数。"""
+
+    def __init__(self, pages):
+        self.pages = pages
+        self.headers = {}
+        self.calls = []
+
+    def get(self, url, params=None, timeout=None):
+        self.calls.append(params)
+        page = params["page"]
+        return _FakeResponse(self.pages[page - 1] if page <= len(self.pages) else {"tasks": []})
+
+
+@pytest.fixture
+def client(monkeypatch):
+    monkeypatch.setattr(pp, "LabelStudio", pp.LabelStudio)
+    instance = pp.LabelStudio.__new__(pp.LabelStudio)
+    instance.url = "http://ls"
+    return instance
+
+
+def test_tasks_pagination_stops_at_total(client):
+    """分页要拉完再停，不能只拿第一页。"""
+    client.session = _FakeSession([
+        {"tasks": [{"id": 1}, {"id": 2}], "total": 3},
+        {"tasks": [{"id": 3}], "total": 3},
+    ])
+    assert [t["id"] for t in client.tasks(39, page_size=2)] == [1, 2, 3]
+
+
+def test_tasks_handles_drf_style_pagination(client):
+    """有的版本返回 results/count 而不是 tasks/total。"""
+    client.session = _FakeSession([
+        {"results": [{"id": 1}], "count": 2},
+        {"results": [{"id": 2}], "count": 2},
+    ])
+    assert [t["id"] for t in client.tasks(39, page_size=1)] == [1, 2]
+
+
+def test_tasks_stops_on_empty_page(client):
+    """没有 total 字段时，靠空页收尾，不能无限翻。"""
+    client.session = _FakeSession([{"tasks": [{"id": 1}]}, {"tasks": []}])
+    assert [t["id"] for t in client.tasks(39)] == [1]
+    assert len(client.session.calls) == 2