chat : fix reasoning leak with force-opened bare <think> templates (ggml-org#24674)

* chat : fix reasoning leak with force-opened bare <think> templates

The reasoning start tag inferred from prior turns can carry trailing
whitespace (e.g. <think>\n) while a force-open template prefills a bare
<think>. Trim the tag used for the prefix split so the bare prefill is
matched instead of being swallowed into content.

* chat : fix Nemotron Nano v2 regression

---------

Co-authored-by: Alde Rojas <hello@alde.dev>

Author: newjordan

common/chat-auto-parser-generator.cpp

@@ -147,7 +147,8 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs, cons
         } else {
             parser = content.build_parser(ctx);
         }
-        return pure_content ? p.prefix(generation_prompt, reasoning.start) + parser : p.prefix(generation_prompt, reasoning.start) << parser;
+        const std::string reasoning_start = trim_whitespace(reasoning.start);
+        return pure_content ? p.prefix(generation_prompt, reasoning_start) + parser : p.prefix(generation_prompt, reasoning_start) << parser;
     });
 }
 

common/chat-diff-analyzer.cpp

@@ -124,16 +124,16 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
               analysis.tools.format.section_end    = "";
               analysis.tools.format.per_call_start = "<TOOLCALL>";
               analysis.tools.format.per_call_end   = "</TOOLCALL>";
+              analysis.tools.format.tools_array_wrapped = true;
               analysis.content.mode                = content_mode::PLAIN;
               analysis.content.start               = "";
               analysis.content.end                 = "";
               analysis.reasoning.mode              = reasoning_mode::TAG_BASED;
-              analysis.reasoning.start             = "<think>\n\n";
+              analysis.reasoning.start             = "<think>\n";
               analysis.reasoning.end               = "</think>";
               analysis.assistant_start             = "<SPECIAL_11>Assistant";
               analysis.user_start                  = "<SPECIAL_11>User";
               analysis.preserved_tokens.clear();
-              analysis.preserved_tokens.push_back("<SPECIAL_12>");
               analysis.preserved_tokens.push_back("<SPECIAL_11>");
               analysis.preserved_tokens.push_back("</think>");
               analysis.preserved_tokens.push_back("<TOOLCALL>");

tests/test-chat.cpp

@@ -4706,9 +4706,16 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
     // Format: <TOOLCALL>[{"name": "func", "arguments": {...}}]</TOOLCALL>
     {
         auto tst = peg_tester("models/templates/NVIDIA-Nemotron-Nano-v2.jinja", detailed_debug);
-        tst.test("<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>")
+        tst.test("I'm\nthinking\n</think>\n<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
             .tools({ special_function_tool })
-            .expect(message_assist_call)
+            .expect(message_assist_call_thoughts)
+            .run();
+
+        tst.test("I'm\nthinking\n</think>\n\n<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>\n")
+            .reasoning_format(COMMON_REASONING_FORMAT_AUTO)
+            .tools({ special_function_tool })
+            .expect(message_assist_call_thoughts)
             .run();
 
         // Continuation tests