chat : fix LFM2 tool-call parsing double-escaping (#24667)

* Add escape test cases

* chat : fix LFM2 tool-call parsing double-escaping

common/chat-peg-parser.cpp

@@ -540,10 +540,11 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
                 auto arg_name_parser = literal(prop_name);
 
                 common_peg_parser arg_value_parser = eps();
-                auto string_value_parser = choice({
-                    literal("\"") + tool_arg_string_value(string_content('"')) + literal("\""),
-                    literal("'") + tool_arg_string_value(string_content('\'')) + literal("'")
-                });
+                // Quoted literal as a value: normalize_quotes_to_json preserves escapes.
+                auto string_value_parser = tool_arg_value(choice({
+                    literal("\"") + string_content('"') + literal("\""),
+                    literal("'") + string_content('\'') + literal("'")
+                }));
 
                 if (is_string_type) {
                     arg_value_parser = string_value_parser;

tests/test-chat.cpp

@@ -1882,11 +1882,29 @@ static void test_lfm2_parser(const std::string & template_path, bool detailed_de
         .expect(simple_assist_msg("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org)."))
         .run();
 
-    // Python tool with multiline code in string
+    // Python tool with multiline code in string: the \n in the literal decodes to a real
+    // newline, emitted as a JSON \n escape (not a doubled backslash).
     tst.test("<|tool_call_start|>[python(code=\"def hello():\\n    print('hey')\")]<|tool_call_end|>")
         .tools({ python_tool })
         .expect_tool_calls({
-            { "python", R"#({"code": "def hello():\\n    print('hey')"})#", "" }
+            { "python", R"#({"code": "def hello():\n    print('hey')"})#", "" }
+        })
+        .run();
+
+    // String escape sequences decode to their actual characters (newline + tab here),
+    // so a "write a two line file" style call produces real line breaks, not literal "\n".
+    tst.test("<|tool_call_start|>[python(code=\"First line\\nSecond line\\tindented\")]<|tool_call_end|>")
+        .tools({ python_tool })
+        .expect_tool_calls({
+            { "python", R"#({"code": "First line\nSecond line\tindented"})#", "" }
+        })
+        .run();
+
+    // Escaped quotes inside a string argument survive the round-trip.
+    tst.test("<|tool_call_start|>[python(code=\"print(\\\"hi\\\")\")]<|tool_call_end|>")
+        .tools({ python_tool })
+        .expect_tool_calls({
+            { "python", R"#({"code": "print(\"hi\")"})#", "" }
         })
         .run();
 

tools/mtmd/mtmd.cpp

@@ -96,16 +96,15 @@ struct mtmd_image_tokens {
             // [BOI] [row0 tokens + newline] ... [row(ny-1) tokens + newline] [EOI]
             return (nx + 1) * ny + 2;
         }
-        // [QWEN_VIDEO] this logic is quite ugly, it's mostly to make qwen-vl temporal merge work, can be improved in the future
-        if (batch_f32.entries.size() == 1 || n_temporal_merge == 1) {
-            return nx * ny;
-        }
         uint32_t nz = batch_f32.entries.size();
-        // TODO: simplify this by repeating the last frame until it fits the temporal merge
-        if (nz % n_temporal_merge != 0) {
-            nz = nz / n_temporal_merge + 1;
-        } else {
-            nz = nz / n_temporal_merge;
+        if (n_temporal_merge > 1) {
+            // [QWEN_VIDEO] this logic is quite ugly, it's mostly to make qwen-vl temporal merge work, can be improved in the future
+            // TODO: simplify this by repeating the last frame until it fits the temporal merge
+            if (nz % n_temporal_merge != 0) {
+                nz = nz / n_temporal_merge + 1;
+            } else {
+                nz = nz / n_temporal_merge;
+            }
         }
         return nx * ny * nz;
     }