fix(ревью#2): db_docs→_db + classify_raw в MemRepo + try/except в build_classify_prompt

deploy/llm_prompt.py

@@ -226,12 +226,14 @@ def build_prompt(current_spec: list, doc_text: str) -> tuple:
 
 def build_classify_prompt(header_text):
     """Build classify prompt. Returns (prompt, prompt_id)."""
-    from db import prompts as db_prompts
-    prompt = db_prompts.get_active("classify")
-    if prompt:
-        body = prompt["body"].replace("{header_text}", header_text)
-        return body, prompt.get("id", "")
-    # Fallback
+    try:
+        from db import prompts as db_prompts
+        prompt = db_prompts.get_active("classify")
+        if prompt:
+            body = prompt["body"].replace("{header_text}", header_text)
+            return body, prompt.get("id", "")
+    except Exception:
+        pass  # DB unavailable — use fallback
     body = """Ты — классификатор договорных документов облачного провайдера НУБЕС.
 
 Ниже фрагмент текста документа. Определи:

deploy/repository.py

@@ -31,7 +31,8 @@ class Repository(Protocol):
 
     def set_classification(self, doc_id: str, doc_type: str, own_number: str = None,
                            parent_number: str = None, doc_date: str = None,
-                           counterparty: str = None) -> None:
+                           counterparty: str = None,
+                           classify_raw: str = None, classify_input: str = None) -> None:
         """Сохранить результат классификации."""
         ...
 
@@ -107,10 +108,12 @@ class PgRepository:
         documents.set_error(doc_id, error)
 
     def set_classification(self, doc_id, doc_type, own_number=None, parent_number=None,
-                           doc_date=None, counterparty=None):
+                           doc_date=None, counterparty=None,
+                           classify_raw=None, classify_input=None):
         from db import documents
         documents.set_classification(doc_id, doc_type, own_number, parent_number,
-                                     doc_date, counterparty)
+                                     doc_date, counterparty,
+                                     classify_raw=classify_raw, classify_input=classify_input)
 
     def set_classify_garbage(self, doc_id, reason=""):
         from db import documents
@@ -199,15 +202,14 @@ class MemRepository:
             self.documents[doc_id]["error_message"] = error
 
     def set_classification(self, doc_id, doc_type, own_number=None, parent_number=None,
-                           doc_date=None, counterparty=None):
+                           doc_date=None, counterparty=None,
+                           classify_raw=None, classify_input=None):
         if doc_id in self.documents:
             d = self.documents[doc_id]
-            d["doc_type"] = doc_type
-            d["own_number"] = own_number
-            d["parent_number"] = parent_number
-            d["doc_date"] = doc_date
-            d["counterparty"] = counterparty
-            d["classify_status"] = "classified"
+            d.update({"doc_type": doc_type, "own_number": own_number,
+                       "parent_number": parent_number, "doc_date": doc_date,
+                       "counterparty": counterparty, "classify_raw": classify_raw,
+                       "classify_input": classify_input, "classify_status": "classified"})
 
     def set_classify_garbage(self, doc_id, reason=""):
         if doc_id in self.documents:

deploy/services/classify.py

@@ -92,8 +92,8 @@ def classify_batch(batch_id, llm_client=None, repo=None):
     _llm = llm_client if llm_client else _get_classify_client()
 
     # Сбросить статус — allow re-classify after file changes
-    db_docs.reset_classify_status(batch_id)
-    pending = db_docs.list_pending(batch_id)
+    _db.reset_classify_status(batch_id)
+    pending = _db.list_pending(batch_id)
     if not pending:
         return {"ok": False, "error": "no pending documents"}
 
@@ -110,24 +110,24 @@ def classify_batch(batch_id, llm_client=None, repo=None):
         try:
             # Stage 1: garbage by filename (0 tokens)
             if _is_garbage_by_filename(doc["filename"]):
-                db_docs.set_classify_garbage(doc["id"], "filename_regex")
+                _db.set_classify_garbage(doc["id"], "filename_regex")
                 garbage += 1
                 return True
 
             # Stage 2: garbage by header keywords (0 tokens)
             text = _smart_extract(doc["elements_json"])
             if _is_garbage_by_header(text):
-                db_docs.set_classify_garbage(doc["id"], "header_keywords")
+                _db.set_classify_garbage(doc["id"], "header_keywords")
                 garbage += 1
                 return True
 
             # Stage 3: LLM classification (only for remaining)
-            db_docs.set_classify_processing(doc["id"])  # crash recovery marker
+            _db.set_classify_processing(doc["id"])  # crash recovery marker
             result, raw, needed_fix = _call_llm_classify(text, _llm)
             json_total += 1
             if needed_fix:
                 json_fixes += 1
-            db_docs.set_classification(
+            _db.set_classification(
                 doc["id"],
                 result.get("doc_type", "other"),
                 result.get("own_number"),
@@ -139,7 +139,7 @@ def classify_batch(batch_id, llm_client=None, repo=None):
             )
             return True
         except Exception as e:
-            db_docs.set_classify_failed(doc["id"], str(e))
+            _db.set_classify_failed(doc["id"], str(e))
             return False
 
     with ThreadPoolExecutor(max_workers=MAX_WORKERS) as pool:

deploy/tests/unit/test_classify_batch.py

@@ -0,0 +1,62 @@
+"""Test classify_batch with MemRepository + FakeLLM."""
+from services.classify import classify_batch
+from services.llm_client import FakeLLMClient
+from repository import MemRepository
+
+
+class TestClassifyBatch:
+    def test_garbage_by_filename(self):
+        """Счёт-фактура отфильтровывается без LLM."""
+        repo = MemRepository()
+        llm = FakeLLMClient({"default": '{"doc_type":"contract"}'})
+        
+        repo.insert_document("счет-фактура №123.docx", "mime", "data", batch_id="b1")
+        
+        result = classify_batch("b1", llm_client=llm, repo=repo)
+        assert result["ok"] is True
+        assert result["garbage"] == 1
+        assert result["done"] == 1
+        assert len(llm.calls) == 0  # LLM не вызывался
+
+    def test_garbage_by_header(self):
+        """Акт сверки в тексте отфильтровывается."""
+        repo = MemRepository()
+        llm = FakeLLMClient({"default": '{"doc_type":"contract"}'})
+        
+        doc = repo.insert_document("документ.docx", "mime", "data", batch_id="b1")
+        repo.set_document_parsed(doc["id"], [
+            {"type": "paragraph", "text": "АКТ СВЕРКИ взаимных расчётов", "style": ""}
+        ])
+        
+        result = classify_batch("b1", llm_client=llm, repo=repo)
+        assert result["garbage"] == 1
+        assert len(llm.calls) == 0
+
+    def test_llm_classify(self):
+        """Договор классифицируется через LLM."""
+        repo = MemRepository()
+        llm = FakeLLMClient({
+            "default": '{"doc_type":"contract","own_number":"03700_1","doc_date":"2026-02-01","counterparty":"ЗАО XXX001"}'
+        })
+        
+        doc = repo.insert_document("договор-XXX001.docx", "mime", "data", batch_id="b1")
+        repo.set_document_parsed(doc["id"], [
+            {"type": "paragraph", "text": "ДОГОВОР № 03700_1", "style": ""}
+        ])
+        
+        result = classify_batch("b1", llm_client=llm, repo=repo)
+        assert result["done"] == 1, f"done={result['done']}, failed={result['failed']}, doc={repo.get_document(doc['id'])}"
+        assert result["garbage"] == 0
+        assert len(llm.calls) == 1
+        
+        updated = repo.get_document(doc["id"])
+        assert updated["doc_type"] == "contract"
+        assert updated["own_number"] == "03700_1"
+        assert updated["counterparty"] == "ЗАО XXX001"
+
+    def test_no_pending(self):
+        """Пустой батч."""
+        repo = MemRepository()
+        result = classify_batch("empty", llm_client=FakeLLMClient(), repo=repo)
+        assert result["ok"] is False
+        assert "no pending" in result["error"]