feat: llm-filter 节点级断点存档

- 每条 LLM 判定成功立即追加 filter_partial.jsonl（多线程加锁串行化）
- 失败/中断后重跑只重判未判定条目，已判定结果复用，与 OCR 存档同机制
- 附带上下文净化回归重跑脚本（run_011d01f19999）

tests/test_llm_filter.py

@@ -709,3 +709,111 @@ def test_real_run_rules_and_dialogue_regression(monkeypatch, tmp_path) -> None:
     })
     assert len(fake.bodies) == unique_llm
     assert len(fake.bodies) < len(entries)  # 去重确实省调用。
+
+
+class _FailOnTarget429:
+    """模拟持续限流：目标字幕含指定词时恒抛 429，其余正常返回 dialogue。
+
+    用于构造"部分条目成功、个别条目持续 429"的断点重跑场景：第一次 invoke
+    整体失败但成功条目已写存档；解除限流后第二次 invoke 只重判失败条目。
+    """
+
+    def __init__(self, marker: str) -> None:
+        self.marker = marker
+        self.enabled = True
+        self.bodies: list[dict] = []
+
+    def __call__(self, request, timeout=None):
+        body = json.loads(request.data.decode("utf-8"))
+        self.bodies.append(body)
+        target = next(
+            line for line in body["messages"][1]["content"].splitlines()
+            if line.startswith("【目标】")
+        )
+        if self.enabled and self.marker in target:
+            raise urllib.error.HTTPError(request.full_url, 429, "rate limited", {}, None)
+        payload = json.dumps({"choices": [{"message": {"content": "dialogue"}}]}).encode()
+        return FakeResponse(payload)
+
+
+def _llm_invoke(srt_text: str, out: Path) -> tuple[object, Path]:
+    """用给定 SRT 文本构造并执行一次 llm-filter invoke，返回 (响应, 输入文件)。"""
+    srt = out.parent / "in.srt"
+    srt.write_text(srt_text, encoding="utf-8")
+    return (
+        invoke(
+            InvokeRequest(
+                run_id="r", node_instance_id="",
+                inputs={"srt_uri": str(srt)},
+                params={},
+                output_dir=str(out),
+            )
+        ),
+        srt,
+    )
+
+
+def test_load_and_append_partial(tmp_path) -> None:
+    """判定存档读写：无存档/损坏行跳过，追加后可读回。"""
+    from nodes.llm_filter import _PARTIAL_NAME, _append_partial, _load_partial
+
+    out = tmp_path / "out"
+    assert _load_partial(out) == {}  # 目录不存在 → 空。
+    out.mkdir()
+    assert _load_partial(out) == {}  # 无存档 → 空。
+    # 损坏行（半行写入）跳过，正常行读回。
+    (out / _PARTIAL_NAME).write_text(
+        '{"index": 0, "category": "dialogue"}\n\n{broken\n{"index": 3, "category": "garbage"}\n',
+        encoding="utf-8",
+    )
+    assert _load_partial(out) == {0: "dialogue", 3: "garbage"}
+    # 追加一条后读回。
+    _append_partial(out, 5, "overlay")
+    assert _load_partial(out)[5] == "overlay"
+
+
+def test_invoke_resume_skips_archived_judgments(monkeypatch, tmp_path) -> None:
+    """断点存档：已判定条目重跑时不重复调用 LLM（去重后只补判未判定）。"""
+    fake = _patch_llm(monkeypatch, contents=["dialogue"] * 4)
+    out = tmp_path / "out"
+    out.mkdir()
+    # 预置存档：索引 0、1 已判定（模拟上次失败前已完成的部分）。
+    from nodes.llm_filter import _append_partial
+
+    _append_partial(out, 0, "dialogue")
+    _append_partial(out, 1, "dialogue")
+    response, _srt = _llm_invoke(_SRT, out)
+    assert response.status == "completed", response.error
+    # 4 条唯一文本中 2 条已存档，只调用剩余 2 条。
+    assert len(fake.bodies) == 2
+    # 输出与全量判定一致：全部 dialogue → 4 条都保留。
+    assert response.outputs["kept"] == 4
+    assert Path(response.outputs["srt_uri"]).read_text(encoding="utf-8").count("-->") == 4
+
+
+def test_invoke_partial_failure_then_resume(monkeypatch, tmp_path) -> None:
+    """真实断点重跑：个别条目持续 429 → 整体失败但成功判定已写存档，
+    解除限流后重跑只补判失败条目，最终产物与一次跑完一致。"""
+    fake = _FailOnTarget429(marker="无意义")
+    monkeypatch.setattr("nodes.llm_filter.urllib.request.urlopen", fake)
+    monkeypatch.setattr("nodes.llm_filter.time.sleep", lambda s: None)
+    out = tmp_path / "out"
+    first, _srt = _llm_invoke(_SRT, out)
+    assert first.status == "failed"
+    assert "429" in (first.error or "")
+    calls_first = len(fake.bodies)
+    # 成功判定的 3 条已写存档；持续 429 的"答：无意义杂项"（索引 2）不在存档。
+    from nodes.llm_filter import _load_partial
+
+    partial = _load_partial(out)
+    assert 2 not in partial
+    assert len(partial) == 3
+    # 解除限流后重跑：只补判失败条目（1 次调用），其余复用存档。
+    fake.enabled = False
+    second, _srt = _llm_invoke(_SRT, out)
+    assert second.status == "completed", second.error
+    assert len(fake.bodies) == calls_first + 1
+    # 输出：解除限流后全部判定为 dialogue → 4 条都保留（与"一次跑完"一致）。
+    out_text = Path(second.outputs["srt_uri"]).read_text(encoding="utf-8")
+    assert out_text.count("-->") == 4
+    assert "00:00:09,000" in out_text