feat: 优化学习视频转写与字幕清洗

tests/test_nodes.py

@@ -292,13 +292,13 @@ def test_resolve_incomplete_candidate_skipped(tmp_path) -> None:
     empty = tmp_path / "empty"
     empty.mkdir()
     resolved = resolve_model_path({}, env={}, candidates=[empty])
-    assert resolved == "large-v3"
+    assert resolved == "large-v2"
 
 
 def test_resolve_fallback_remote() -> None:
-    """验证全部本地候选缺失时回退到远端 large-v3 作为最后兜底。"""
+    """验证全部本地候选缺失时回退到远端 large-v2 作为最后兜底。"""
     resolved = resolve_model_path({}, env={}, candidates=[])
-    assert resolved == "large-v3"
+    assert resolved == "large-v2"
 
 
 def test_format_timestamp() -> None:
@@ -1497,5 +1497,68 @@ def test_whisper_decode_full_cleanup_jp_hallucination(tmp_path, monkeypatch) ->
     assert all(t != "-" for t in text_lines)
 
 
+def test_whisper_decode_full_filters_short_moan(tmp_path, monkeypatch) -> None:
+    """decode_full=true 时纯呻吟碎片整条删除，真实短词保留（2026-09 决策）。"""
+    class MoanModel:
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def transcribe(self, path, **kwargs):
+            # 混合：正常句 + 纯呻吟碎片 + 真实短词（都应保留）。
+            return (
+                [
+                    FakeSegment(0, 2, "気持ちいい"),   # 正常
+                    FakeSegment(2, 3, "あ〜"),         # 纯呻吟 → 删
+                    FakeSegment(3, 4, "ん？"),         # 纯呻吟 → 删
+                    FakeSegment(4, 5, "そこ"),         # 真实短词 → 保留
+                    FakeSegment(5, 6, "やばい"),       # 真实短词 → 保留
+                ],
+                None,
+            )
+
+    monkeypatch.setitem(
+        sys.modules, "faster_whisper", types.SimpleNamespace(WhisperModel=lambda *a, **k: MoanModel())
+    )
+    _make_wav(tmp_path / "audio.wav", 5)
+    resp = whisper_invoke(
+        _whisper_request(tmp_path, params={"language": "ja", "decode_full": True})
+    )
+    assert resp.status == "completed"
+    content = Path(resp.outputs["srt_uri"]).read_text(encoding="utf-8")
+    # 纯呻吟碎片被整条删除；真实短词与正常句保留。
+    assert "あ〜" not in content
+    assert "ん？" not in content
+    assert "気持ちいい" in content
+    assert "そこ" in content
+    assert "やばい" in content
+    assert content.count("-->") == 3  # 5 段 - 2 段呻吟 = 3 条
+
+
+def test_whisper_decode_full_short_moan_can_disable(tmp_path, monkeypatch) -> None:
+    """short_moan_max_chars=0 时关闭短呻吟过滤，所有内容原样保留。"""
+    class MoanModel2:
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def transcribe(self, path, **kwargs):
+            return ([FakeSegment(0, 1, "あ〜"), FakeSegment(1, 2, "そこ")], None)
+
+    monkeypatch.setitem(
+        sys.modules, "faster_whisper", types.SimpleNamespace(WhisperModel=lambda *a, **k: MoanModel2())
+    )
+    _make_wav(tmp_path / "audio.wav", 5)
+    resp = whisper_invoke(
+        _whisper_request(
+            tmp_path,
+            params={"language": "ja", "decode_full": True, "short_moan_max_chars": 0},
+        )
+    )
+    assert resp.status == "completed"
+    content = Path(resp.outputs["srt_uri"]).read_text(encoding="utf-8")
+    assert "あ〜" in content
+    assert "そこ" in content
+    assert content.count("-->") == 2
+
+