notes 锚点源与 chapters 实际锚点同源化（修 #227 完整版）

CHANGELOG.md

@@ -9,6 +9,7 @@
 
 ### Fixed
 
+- 章节笔记：章节目录产物现在记录实际锚定的数据来源（缓存结构化对话 / 原始转写分片），笔记按同一来源精确复现。此前目录标成「原始分片」却实际锚在缓存对话上，笔记必然指纹校验失败且无法生成（生产事故：BV1fcHD6cEw3）。标签声明的来源已不存在时直接报错，不再默默换成另一份数据。
 - 章节目录：模型按主题弧线而非时间序返回章节时（每章时间锚点正确），章节层不再整体失败，改为按时间排序归一化（排序改变顺序时记入日志，含错位数与 LIS 统计）。此前此类输出会导致整个视频无章节目录（生产事故：BV1fcHD6cEw3）。
 - 章节生成失败原因现在落盘到任务状态的 llm_status.json（chapters_error 字段），不再只出现在服务日志里；补跑成功后旧失败原因自动清除。
 

src/video_transcript_api/api/services/llm_ops.py

@@ -732,6 +732,7 @@ def _handle_llm_task(llm_task: dict):
                         speaker_count_hint=llm_task.get("cached_speaker_count"),
                         skip_chapters=skip_chapters_for_coordinator,
                         timeline_segments=timeline_segments_seed,
+                        timeline_segments_kind=chapters_seed_kind,
                     )
 
                 # 适配返回格式

src/video_transcript_api/llm/coordinator.py

@@ -141,6 +141,7 @@ def process(
         contradiction_scan: Optional[bool] = None,
         skip_chapters: bool = False,
         timeline_segments: Optional[List[Dict]] = None,
+        timeline_segments_kind: Optional[str] = None,
     ) -> Dict:
         """处理文本（统一入口）
 
@@ -173,6 +174,10 @@ def process(
             timeline_segments: 章节输入（dialogs 或原始 segments 列表）。由
                 llm_ops 按输入梯度解析后传入，保持 llm/ 包不依赖 cache 读盘。
                 None/空时 processor 会诚实返回 SKIPPED_NO_TIMELINE。
+            timeline_segments_kind: timeline_segments 的真实来源标签（调用方
+                已解析的事实，如 ``cached_dialogs`` / ``segments``）。标签会原样
+                写进 llm_chapters.json 的 source.kind，notes reader 据此精确复现
+                同一份锚点数据，缺失才默认按原始 segments 形态处理。
 
         Returns:
             处理结果字典:
@@ -314,7 +319,14 @@ def process(
         else:
             logger.info("Step 3/3: Chapters generation")
             chapters_input = timeline_segments
-            chapters_source_kind = "segments" if timeline_segments else "none"
+            # notes_anchor_kind_truth: source.kind 必须是「实际喂给 chapters 的
+            # 锚点数据」的真实来源，而不是「入参长什么样」。调用方 llm_ops 按输入
+            # 梯度解析种子时已经知道它是缓存 dialogs 还是原始 segments，这里直接
+            # 采用调用方的事实；本轮 structured dialogs 覆盖种子时改写为 dialogs。
+            # 两者都是真实锚点，notes reader 才能据此精确复现（#227）。
+            chapters_source_kind = (
+                (timeline_segments_kind or "segments") if timeline_segments else "none"
+            )
             structured_for_chapters = calibration_result.get("structured_data")
             if isinstance(structured_for_chapters, dict):
                 dialogs = structured_for_chapters.get("dialogs")

src/video_transcript_api/llm/processors/notes_processor.py

@@ -117,28 +117,39 @@ def load_notes_source_segments(
     ``dialogs``/``cached_dialogs`` sources use ``llm_processed.json``.  A
     ``segments`` source uses the shared ``load_segments`` adapter, preserving
     the plain-text path's normalized list and original index order.
+
+    A declared ``source_kind`` is reproduced exactly: if the cache no longer
+    holds that anchor, the lookup fails as ``(None, None)`` instead of silently
+    substituting the other source, which would make notes hash a different
+    anchor than chapters did (#227).  Only a product carrying no label at all
+    falls back to probing ``dialogs`` then ``segments``.
     """
     cache_path = Path(cache_dir)
     normalized_kind = (source_kind or "").lower()
-    wants_dialogs = normalized_kind in {
+
+    processed = _read_json_file(cache_path / "llm_processed.json")
+    dialogs = processed.get("dialogs") if processed else None
+    dialogs_source = dialogs if isinstance(dialogs, list) and dialogs else None
+
+    if normalized_kind in {
         "dialogs",
         "cached_dialogs",
         "structured",
         "plain_structured",
-    }
-    wants_segments = normalized_kind in {"segments", "timeline"}
+    }:
+        return (dialogs_source, "dialogs") if dialogs_source is not None else (None, None)
 
-    processed = _read_json_file(cache_path / "llm_processed.json")
-    dialogs = processed.get("dialogs") if processed else None
-    if (wants_dialogs or not wants_segments) and isinstance(dialogs, list) and dialogs:
-        return dialogs, "dialogs"
+    if normalized_kind in {"segments", "timeline"}:
+        segments = load_segments(cache_path)
+        if isinstance(segments, list) and segments:
+            return segments, "segments"
+        return None, None
 
+    if dialogs_source is not None:
+        return dialogs_source, "dialogs"
     segments = load_segments(cache_path)
     if isinstance(segments, list) and segments:
         return segments, "segments"
-
-    if not wants_segments and isinstance(dialogs, list) and dialogs:
-        return dialogs, "dialogs"
     return None, None