feat: persist document_structure.json, thread structure truth through downstream pipeline

paperforge/worker/ocr.py

@@ -1760,7 +1760,10 @@ def postprocess_ocr_result(vault: Path, key: str, all_results: list[dict]) -> tu
         backfill_span_metadata_from_pdf(all_raw_blocks, source_pdf_path)
         write_raw_blocks_jsonl(artifacts.blocks_raw, all_raw_blocks)
 
-    structured = build_structured_blocks(all_raw_blocks)
+    structured = build_structured_blocks(
+        all_raw_blocks,
+        structure_output_dir=artifacts.blocks_structured.parent,
+    )
     write_structured_blocks_jsonl(artifacts.blocks_structured, structured)
     # Write role-level span profiles
     from paperforge.worker.ocr_profiles import write_role_span_profiles

paperforge/worker/ocr_blocks.py

@@ -7,7 +7,10 @@ from typing import Any
 from paperforge.worker.ocr_roles import assign_block_role
 
 
-def build_structured_blocks(raw_blocks: list[dict]) -> list[dict]:
+def build_structured_blocks(
+    raw_blocks: list[dict],
+    structure_output_dir: str | Path | None = None,
+) -> list[dict]:
     # Group raw blocks by page so assign_block_role can see page-local context
     by_page: dict[int, list[dict]] = {}
     for block in raw_blocks:
@@ -86,9 +89,35 @@ def build_structured_blocks(raw_blocks: list[dict]) -> list[dict]:
 
         rows = rescue_roles_with_document_context(rows, paper_context["role_profiles"], doc_structure)
 
+    # Persist document structure artifact for downstream debugging
+    if doc_structure and structure_output_dir:
+        _write_document_structure_json(doc_structure, structure_output_dir)
+
     return rows
 
 
+def _write_document_structure_json(doc_structure, output_dir: str | Path) -> None:
+    """Serialize DocumentStructure to JSON for downstream inspection."""
+    import dataclasses
+
+    if hasattr(doc_structure, "_asdict"):
+        data = doc_structure._asdict()
+    elif dataclasses.is_dataclass(doc_structure):
+        data = dataclasses.asdict(doc_structure)
+    else:
+        return
+    # Convert non-serializable types
+    for k, v in data.items():
+        if hasattr(v, "_asdict"):
+            data[k] = v._asdict()
+        elif dataclasses.is_dataclass(v):
+            data[k] = dataclasses.asdict(v)
+    output_path = Path(output_dir) / "document_structure.json"
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    with open(output_path, "w", encoding="utf-8") as f:
+        json.dump(data, f, indent=2, ensure_ascii=False)
+
+
 def write_structured_blocks_jsonl(path: Path, rows: list[dict]) -> None:
     path.parent.mkdir(parents=True, exist_ok=True)
     with path.open("w", encoding="utf-8") as f:

paperforge/worker/ocr_rebuild.py

@@ -67,7 +67,10 @@ def run_derived_rebuild_for_keys(vault: Path, keys: list[str]) -> dict:
         # Rebuild structured blocks
         from paperforge.worker.ocr_blocks import build_structured_blocks, write_structured_blocks_jsonl
 
-        structured = build_structured_blocks(all_raw_blocks)
+        structured = build_structured_blocks(
+            all_raw_blocks,
+            structure_output_dir=artifacts.blocks_structured.parent,
+        )
         write_structured_blocks_jsonl(artifacts.blocks_structured, structured)
         # Write role-level span profiles
         from paperforge.worker.ocr_profiles import write_role_span_profiles