perf: skip fact extraction for fully graded projects on batch retries

backend/tests/test_optimization.py

@@ -402,6 +402,43 @@ def test_failed_model_call_cannot_overwrite_a_concurrent_teacher_review(sqlite_d
     assert db.get(ScoringBatch, batch.id).progress_done == 1
 
 
+@pytest.mark.parametrize("complete", [True, False])
+def test_retry_skips_facts_only_for_fully_graded_projects(sqlite_db, monkeypatch, complete):
+    from wisegrade import scoring
+    db = sqlite_db
+    item = project(db)
+    created = services.create_batch(SimpleNamespace(course_version_id=1, rubric_version_id=1), db)
+    batch = db.get(ScoringBatch, created["id"])
+    criteria = batch.rubric_snapshot["criteria"][:2]
+    batch.rubric_snapshot = {**batch.rubric_snapshot, "criteria": criteria}
+    batch.status, batch.progress_total, batch.progress_done = "running", 2, 0
+    item.facts = {"conclusions": ["已抽取并确认的事实"]}
+    db.add_all([
+        CriterionScore(batch_id=batch.id, project_id=item.id, criterion_id=criteria[0]["criterion_id"], status="scored", model_score=0),
+        CriterionScore(batch_id=batch.id, project_id=item.id, criterion_id=criteria[1]["criterion_id"],
+                       status="reviewed" if complete else "needs_review", teacher_score=0 if complete else None),
+    ])
+    db.commit()
+    calls = []
+    def client_factory(*args):
+        calls.append("client")
+    def facts(*args, **kwargs):
+        calls.append("facts")
+        return {}
+    def scored(session, batch, item, criterion, *args, **kwargs):
+        return session.scalar(select(CriterionScore).where(CriterionScore.batch_id == batch.id,
+            CriterionScore.project_id == item.id, CriterionScore.criterion_id == criterion["criterion_id"]))
+    monkeypatch.setattr(scoring, "_project_facts", facts)
+    monkeypatch.setattr(scoring, "_score_criterion", scored)
+    factory = sessionmaker(bind=db.get_bind(), expire_on_commit=False)
+    failures, reviews = scoring._run_project(batch.id, item.id, factory, client_factory, FLAT, lambda *a, **k: [])
+    assert calls == ([] if complete else ["client", "facts"])
+    assert failures == 0 and reviews == (0 if complete else 1)
+    db.expire_all()
+    assert db.get(ScoringBatch, batch.id).progress_done == 2
+    assert db.get(Project, item.id).facts == {"conclusions": ["已抽取并确认的事实"]}
+
+
 def test_unrebuilt_course_document_does_not_complete_task(sqlite_db):
     db = sqlite_db
     doc = KnowledgeDocument(course_version_id=1, title="无页面资料", document_type="course",

backend/wisegrade/scoring.py

@@ -358,6 +358,18 @@ def _run_project(batch_id, project_id, session_factory, client_factory, embedder
     with session_factory() as db:
         batch = db.get(ScoringBatch, batch_id)
         project = db.get(Project, project_id)
+        existing = db.scalars(select(CriterionScore).where(CriterionScore.batch_id == batch_id,
+                              CriterionScore.project_id == project_id)).all()
+        graded = {score.criterion_id for score in existing
+                  if score.status in (STATUS_SCORED, STATUS_REVIEWED)}
+        criteria = batch.rubric_snapshot["criteria"]
+        if all(criterion["criterion_id"] in graded for criterion in criteria):
+            # A retry in a multi-student batch must not spend two more model calls
+            # extracting facts for a student whose scores will all be skipped.
+            db.execute(update(ScoringBatch).where(ScoringBatch.id == batch_id).values(
+                progress_done=ScoringBatch.progress_done + len(criteria)))
+            db.commit()
+            return 0, 0
         try:
             client = client_factory(db, batch)
         except Exception: