Harden OpenAI tool-call/chat-template + reasoning parsing

serving_chat:
- Tolerate a malformed historical tool_call `arguments` string (a valid JSON
  document followed by trailing content) instead of 400-ing the whole
  multi-turn request: salvage the leading JSON document via raw_decode, else
  keep the raw string. (A 112-message tool-history request was rejected with
  orjson "unexpected content after document".)
- Catch TypeError (not only jinja2.TemplateError) from the chat-template
  render so a `tojson` filter on a Jinja Undefined becomes a clean 400 instead
  of a 500 (upstream #20700 / 5e9bd21979).

reasoning_parser:
- Strip only LEADING think-start marker tokens; a global replace would delete a
  `<think>` token that legitimately appears inside reasoning content. Preserve
  model-generated whitespace in reasoning/normal text (drop .strip()/.rstrip())
  (upstream #24251 / dac78768f0).
- Add Glm45 detector tests: leading-only strip, token-inside-content preserved,
  repeated leading markers, streaming trailing whitespace.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

test/registered/unit/parser/test_reasoning_parser.py

@@ -427,6 +427,40 @@ class TestGlm45Detector(CustomTestCase):
         self.assertEqual(result.normal_text, text)
         self.assertEqual(result.reasoning_text, "")
 
+    def test_detect_and_parse_preserves_whitespace(self):
+        """Reasoning + normal text must keep model-generated whitespace (newlines,
+        indentation); the parser must not rewrite the payload."""
+        text = "<think>\n  Let me think\n</think>\n\nThe answer is 42.\n"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "\n  Let me think\n")
+        self.assertEqual(result.normal_text, "\n\nThe answer is 42.\n")
+
+    def test_detect_and_parse_keeps_think_token_inside_content(self):
+        """A think-start token appearing INSIDE the reasoning content is real
+        content, not a marker: only LEADING markers are stripped (a global
+        replace would corrupt the reasoning)."""
+        text = "<think>quote the <think> tag literally</think>answer"
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "quote the <think> tag literally")
+        self.assertEqual(result.normal_text, "answer")
+
+    def test_detect_and_parse_strips_repeated_leading_think_tokens(self):
+        """Repeated leading think-start tokens are markers, not payload."""
+        text = "<think><think>Let me think</think>The answer is 42."
+        result = self.detector.detect_and_parse(text)
+        self.assertEqual(result.reasoning_text, "Let me think")
+        self.assertEqual(result.normal_text, "The answer is 42.")
+
+    def test_streaming_preserves_reasoning_trailing_whitespace(self):
+        """Streaming reasoning text must preserve trailing whitespace before the
+        end token (no rstrip)."""
+        self.detector.parse_streaming_increment("<think>")
+        result = self.detector.parse_streaming_increment(
+            "reasoning  \n</think>answer"
+        )
+        self.assertEqual(result.reasoning_text, "reasoning  \n")
+        self.assertEqual(result.normal_text, "answer")
+
     def test_streaming_normal_flow(self):
         """Test streaming with normal reasoning flow."""
         # Start reasoning