fix: SRT 按 cue 解析并按 ID 回填译文，OCR 空帧分段与失败重试

tests/test_ocr_recovery.py

@@ -0,0 +1,66 @@
+"""R06：真实图片清单上的临时网络故障、空帧与跨空白字幕段回归。"""
+
+import json
+from pathlib import Path
+
+import pytest
+
+from nodes import subtitle_ocr
+from wov_sdk.models import InvokeRequest, InvokeResponse
+
+
+def test_identical_text_separated_by_blank_is_two_cues():
+    """A、空白、A 不能合并，否则字幕会覆盖原本无文字的时段。"""
+    manifest = [{"time": i * 0.5} for i in range(4)]
+    assert subtitle_ocr._merge_kept(manifest, ["你好", "你好", "", "你好"]) == [
+        (0.0, 0.5, "你好"), (1.5, 1.5, "你好")]
+
+
+@pytest.mark.parametrize("raises", [False, True])
+def test_failed_frame_retries_and_resume_preserves_success(monkeypatch, tmp_path, raises):
+    """失败帧单独重试，持续故障不存为空；下次只补失败帧，成功空帧不重做。"""
+    assets = Path(__file__).resolve().parent.parent / "testdata"
+    image = assets / "ocr_text.png"
+    empty = assets / "ocr_notext.png"
+    if not image.is_file() or not empty.is_file():
+        pytest.skip("缺少真实 OCR 图片")
+    manifest = tmp_path / "frames.json"
+    manifest.write_text(json.dumps([{"time": 0, "image_uri": str(empty)},
+                                    {"time": 0.5, "image_uri": str(image)}]))
+    out = tmp_path / "out"
+    request = InvokeRequest(run_id="r", node_instance_id="", inputs={"frames_manifest": str(manifest)}, output_dir=str(out))
+    calls = []
+    broken = True
+
+    def invoke(node_id, req):
+        uri = req.inputs["image_uri"]
+        calls.append(uri)
+        if uri == str(image) and broken:
+            if raises:
+                raise TimeoutError("timeout")
+            return InvokeResponse(status="failed", error="timeout")
+        return InvokeResponse(status="completed", outputs={"text": "" if uri == str(empty) else "你好"})
+
+    monkeypatch.setattr("wov_app.registry.invoke", invoke)
+    response = subtitle_ocr.invoke(request)
+    assert response.status == "failed"
+    assert calls.count(str(image)) == 2
+    assert calls.count(str(empty)) == 1
+    partial = [json.loads(line) for line in (out / "ocr_partial.jsonl").read_text().splitlines()]
+    assert partial == [{"frame": 0, "text": "", "status": "completed"}]
+    assert not (out / "subtitle.srt").exists()
+    broken = False
+    calls.clear()
+    response = subtitle_ocr.invoke(request)
+    assert response.status == "completed"
+    assert calls == [str(image)]
+    assert "你好" in Path(response.outputs["srt_uri"]).read_text()
+
+
+def test_legacy_empty_checkpoint_is_rechecked(tmp_path):
+    """旧版空串可能来自超时，不能当成确认无文字；旧版非空成功结果可复用。"""
+    (tmp_path / "ocr_partial.jsonl").write_text(
+        json.dumps({"frame": 0, "text": ""}) + "\n" +
+        json.dumps({"frame": 1, "text": "你好"}, ensure_ascii=False) + "\n" +
+        json.dumps({"frame": 2, "text": "", "status": "completed"}) + "\n")
+    assert subtitle_ocr._load_partial(tmp_path) == {1: "你好", 2: ""}