fix(test): match the current scan_layers mismatch error

The restore path now reports an unmatched checkpoint with 'Checkpoint does not
match the model', listing each weight the checkpoint didn't carry. The test
still expected the older wording. 'Failed to restore checkpoint' is not raised
anywhere, and 'Configuration mismatch' comes from the load_parameters_path
metadata check, which this test doesn't exercise.

Also assert the missing per-layer weight path, so the check is tied to the
scan/unscan mismatch rather than the generic hint line, which any architecture
mismatch would print.

Author: ecnal-cienet

tests/integration/checkpointing_test.py

@@ -228,5 +228,9 @@ def get_cmd(steps, metrics_file):
   with pytest.raises(ValueError) as excinfo:
     train_main(mismatch_command)
 
-  assert "Configuration mismatch" in str(excinfo.value) or "Failed to restore checkpoint" in str(excinfo.value)
-  assert "scan_layers" in str(excinfo.value)
+  # The scanned checkpoint doesn't carry the unscanned model's per-layer weights, so the restore
+  # reports them as missing rather than leaving them at their init values.
+  message = str(excinfo.value)
+  assert "Checkpoint does not match the model" in message
+  assert "decoder/layers/0/" in message
+  assert "scan_layers" in message