fix: harden extraction batch handling

site/services/extractor.py

@@ -250,12 +250,16 @@ def _save_extraction(document_id, extracted, validation, status, raw_response="{
 
 def extract_document(document, llm_client=None, batch_id=None):
     """Извлечь паспорт одного документа, изолируя ошибки на уровне документа."""
-    document_id = document["id"]
-    batch_id = batch_id if batch_id is not None else document.get("batch_id")
-    elements = _decode_elements(document.get("elements", document.get("elements_json")))
-    structured_text = format_elements_with_structure(elements)
+    document_id = document.get("id") if isinstance(document, dict) else None
     raw_response = "{}"
     try:
+        if not document_id:
+            raise ValueError("document id is required")
+        batch_id = batch_id if batch_id is not None else document.get("batch_id")
+        elements = _decode_elements(document.get("elements", document.get("elements_json")))
+        # Formatting is part of this document's failure boundary: corrupt stored
+        # elements must not escape and stop extraction of later batch members.
+        structured_text = format_elements_with_structure(elements)
         if _is_garbage_by_filename(document.get("filename") or "") or _is_garbage_by_header(structured_text[:2000]):
             extracted = {
                 "doc_class": "ignored_accounting", "intent_mode": "no_spec",
@@ -299,11 +303,19 @@ def extract_document(document, llm_client=None, batch_id=None):
         return {"ok": True, "document_id": document_id, "status": result["status"], "extraction": saved}
     except Exception as error:
         log.exception("Stateless extraction failed for document %s", document_id)
-        _save_extraction(document_id, {
-            "doc_class": "unknown", "intent_mode": "no_spec", "counterparty": {},
-            "referenced_contracts": [], "sections": [],
-        }, {"status": "FAILED", "error": str(error)}, "FAILED", raw_response)
-        return {"ok": False, "document_id": document_id, "error": str(error)}
+        result = {"ok": False, "document_id": document_id, "error": str(error)}
+        if document_id:
+            # Failure persistence is best-effort; a secondary write failure must
+            # not escape this handler and abort processing of remaining documents.
+            try:
+                _save_extraction(document_id, {
+                    "doc_class": "unknown", "intent_mode": "no_spec", "counterparty": {},
+                    "referenced_contracts": [], "sections": [],
+                }, {"status": "FAILED", "error": str(error)}, "FAILED", raw_response)
+            except Exception as persistence_error:
+                log.exception("Failed to persist extraction failure for document %s", document_id)
+                result["persistence_error"] = str(persistence_error)
+        return result
 
 
 def extract_batch(batch_id, llm_client=None, progress_callback=None, documents=None):