Pin transformers==4.57.6 and add runtime version guard

transformers 5.x silently produces wrong predictions (all 'unknown') when
loading a 4.x checkpoint — no error is raised. Pinning the exact version
the model was saved with, and adding a startup check, makes the failure
loud and clear instead of invisible.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

requirements.txt

@@ -1,5 +1,5 @@
 datasets>=2.14.0
-transformers>=4.35.0,<5.0.0
+transformers==4.57.6
 torch>=2.0.0
 pandas>=2.0.0
 numpy>=1.24.0

src/__init__.py

@@ -1,5 +1,15 @@
 from __future__ import annotations
 
+import importlib.metadata
+from packaging.version import Version
+
+_tv = Version(importlib.metadata.version("transformers"))
+if _tv >= Version("5.0.0"):
+    raise RuntimeError(
+        f"transformers {_tv} is installed but this project requires <5.0.0. "
+        "Run: pip install transformers==4.57.6"
+    )
+
 from data.schema import Argument, Debate
 
 LABEL2ID = {"claim": 0, "counter_claim": 1, "premise": 2, "unknown": 3}