fix: pin model revisions and quarantine MariaDB CI routing

tools/summarizer/app.py

@@ -94,6 +94,8 @@ async def require_service_token(request: Request, call_next):
         _external_ai_allowed.reset(allowed_token)
 
 MODEL_NAME = "sshleifer/distilbart-cnn-12-6"
+# Keep tokenizer and weights on the same reviewed immutable upstream snapshot.
+MODEL_REVISION = "a4f8f3ea906ed274767e9906dbaede7531d660ff"
 MAX_INPUT_CHARS = 20000
 MAX_CONTEXT_CHARS = 2200
 MAX_EXTRACT_FILE_BYTES = 5 * 1024 * 1024
@@ -159,8 +161,8 @@ MODEL_DISABLED = SKIP_MODEL_LOAD
 
 
 def _load_runtime():
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
-    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, revision=MODEL_REVISION, trust_remote_code=False)
+    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME, revision=MODEL_REVISION, trust_remote_code=False, weights_only=True)
     model.eval()
     has_cuda = torch.cuda.is_available()
     device = torch.device("cuda" if has_cuda else "cpu")

tools/summarizer/tests/test_app.py

@@ -84,6 +84,37 @@ def test_health_reports_runtime_without_ollama_and_without_forcing_model_load(mo
     assert payload["ollama_loaded_models"] == []
 
 
+def test_model_load_pins_both_artifacts_and_denies_remote_code_and_unsafe_weights(monkeypatch):
+    module = load_app_module(monkeypatch)
+    calls = []
+
+    class LoadedModel:
+        def eval(self):
+            return self
+
+        def to(self, device):
+            return self
+
+    def tokenizer_loader(name, **options):
+        calls.append((name, options))
+        return object()
+
+    def model_loader(name, **options):
+        calls.append((name, options))
+        return LoadedModel()
+
+    monkeypatch.setattr(module.AutoTokenizer, "from_pretrained", tokenizer_loader)
+    monkeypatch.setattr(module.AutoModelForSeq2SeqLM, "from_pretrained", model_loader)
+    monkeypatch.setattr(module.torch.cuda, "is_available", lambda: False)
+    module._load_runtime()
+
+    assert [name for name, _ in calls] == ["sshleifer/distilbart-cnn-12-6"] * 2
+    for _, options in calls:
+        assert options.get("revision") == "a4f8f3ea906ed274767e9906dbaede7531d660ff"
+        assert options.get("trust_remote_code") is False
+    assert calls[1][1].get("weights_only") is True
+
+
 def test_summarize_returns_503_with_explicit_reason_when_model_loading_is_disabled(monkeypatch):
     module = load_app_module(monkeypatch)
     client = TestClient(module.app)