textconv: recover Docling captions nested under tables/pictures (1016)

app/textconv.py

@@ -77,6 +77,8 @@ def grid_markdown(grid) -> str:
 
 # ---------------------------------------------------------------- Docling (PDF only)
 
+CAPTION_START = re.compile(r"(Fig\.?|Figure|FIGURE|Table|TABLE)\s*S?\d+[A-Za-z]?\b")
+
 def convert_docling(pdf: Path):
     from docling.datamodel.base_models import InputFormat
     from docling.datamodel.pipeline_options import PdfPipelineOptions
@@ -107,8 +109,20 @@ def convert_docling(pdf: Path):
             continue
         label = item.label.value
         if isinstance(item, (TableItem, PictureItem)):
-            for cref in item.captions:
-                cap = cref.resolve(doc)
+            caps = [cref.resolve(doc) for cref in item.captions]
+            # Docling 有时把图注挂在表 / 图的子节点里、不登记进 captions(traverse_pictures=False 会整个跳过), 这里补回(1016)。
+            # 只收 label=caption, 或以 "Fig./Figure/Table + 编号" 开头的子节点文字(会被错标成 footnote / text);
+            # 图内其余文字(坐标轴、图例)不收。
+            seen = {c.self_ref for c in caps}
+            for ch, _ in doc.iterate_items(root=item, traverse_pictures=True):
+                txt = one_line(getattr(ch, "text", ""))
+                if ch is item or ch.self_ref in seen or not txt:
+                    continue
+                if ch.label.value == "caption" or CAPTION_START.match(txt):
+                    caps.append(ch)
+                    seen.add(ch.self_ref)
+                    owned.add(ch.self_ref)         # 主循环之后若再遇到同一节点, 不再输出第二遍
+            for cap in caps:
                 B.add(one_line(cap.text), "caption", boxes(cap))
             if isinstance(item, TableItem):
                 md = grid_markdown([[c.text for c in row] for row in item.data.grid])