strata_tokenizer: the approved wording for an unknown pre-tokenizer

serve/test_tool_format.py

@@ -154,7 +154,9 @@ class UnknownPreTokenizerAtStart(unittest.TestCase):
                 server.main(["--engine", "mock", "--tokenizer", str(t), "--port", "0"])
             self.assertEqual(cm.exception.code, 2)
             self.assertIn("'qwen4'", err.getvalue())
-            self.assertIn("is not loaded", err.getvalue())
+            self.assertIn("so it isn't loaded", err.getvalue())
+            self.assertIn("Strata doesn't have yet ('qwen4')", err.getvalue())
+            self.assertIn("the model's tokenizer (", err.getvalue())
             self.assertNotIn("Traceback", err.getvalue())
 
 

tools/strata_tokenizer.py

@@ -120,9 +120,8 @@ class Tokenizer:
         if pre_tokenizer_kind(pre) is None:
             # Not a fallback: another model's pre-tokenizer splits text differently, so encoding with the qwen35 rule would
             # change every prompt without any error (llama.cpp has one rule per `tokenizer.ggml.pre` type for the same reason).
-            raise UnknownPreTokenizer("this model's tokenizer uses the text-splitting rule %r, which this version of Strata does not have "
-                                      "(it has qwen35 and deepseek). Encoding with another rule would quietly change what the model reads, "
-                                      "so the model is not loaded." % (pre,))
+            raise UnknownPreTokenizer("This model splits text with a rule Strata doesn't have yet (%r), so it isn't loaded: "
+                                      "running it with the wrong rule would change what it reads." % (pre,))
         self.token_types = token_types
         self.special_ids = special_ids or {}
         self.ids = {t: i for i, t in enumerate(tokens)}