runtime: inject system prompt via rkllm_set_chat_template (not inlined in user msg)

Lab smoke test after the role-based refactor showed the model's first
thought event regurgitating the entire SYSTEM_PROMPT_TEMPLATE — because
the prior fix inlined the system rules into the user message as
'System: [rules]\\n\\nUser request: [prompt]'. The model treated the
rules as additional context to acknowledge / restate.

Proper fix: configure session-level chat template via
rkllm_set_chat_template(handle, system, prefix, postfix) so the
runtime injects the system in the canonical <|im_start|>system\\n...
<|im_end|> slot. The model recognises it as instructions, not as
something to repeat back.

Trade-off documented earlier — calling set_chat_template disables
the runtime's automatic enable_thinking handling. We work around by
including '<think>\\n' in the postfix, so every assistant turn starts
inside a think block:
    postfix = '<|im_end|>\\n<|im_start|>assistant\\n<think>\\n'

This still keeps <|im_end|> as the stop-token signal (the first
token of the postfix), so the runtime correctly terminates each
turn at <|im_end|> — no more fictitious continuations.

55/55 unit tests pass.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: ehsan6sha

src/runtime/rkllm_runtime.py

@@ -1093,21 +1093,37 @@ async def run_troubleshoot(
         original_user_prompt = prompt
 
         # Per-turn role-based content for v1.2.3.
-        # Turn 0: role="user", content = SYSTEM_PROMPT_TEMPLATE concatenated
-        # with the user's actual prompt. v1.2.3's RKLLMInput role enum
-        # doesn't document a "system" role, and rkllm_set_chat_template
-        # would disable thinking-mode handling, so we inline the system
-        # rules into the first user message. The model was trained with
-        # the full system prompt visible in every example, so it still
-        # recognises the rules even via this less-structured channel.
+        # Turn 0: role="user", content = user's actual prompt only. The
+        # SYSTEM prompt is configured ONCE via rkllm_set_chat_template
+        # below so the runtime injects it in the proper <|im_start|>
+        # system\n...<|im_end|> slot — the model sees it as
+        # instructions, not as part of the user's request, and doesn't
+        # regurgitate it back in its first thought event.
         # Turn 1+: role="tool" with the JSON tool response. The runtime
         # appends to the existing KV cache via keep_history=1.
-        first_turn_content = (
-            f"{system_prompt}\n\n"
-            f"User request: {prompt}"
-        )
+        # Configure session-specific chat template — uses our system
+        # prompt + Qwen 3 markers + `<think>\n` postfix to force
+        # thinking-mode (the auto-thinking flag is disabled when
+        # set_chat_template is called per runtime warning).
+        try:
+            self._runtime._lib.rkllm_set_chat_template.restype = ctypes.c_int
+            self._runtime._lib.rkllm_set_chat_template.argtypes = [
+                ctypes.c_void_p, ctypes.c_char_p, ctypes.c_char_p, ctypes.c_char_p,
+            ]
+            system_wrapped = (
+                f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
+            ).encode("utf-8")
+            prefix = b"<|im_start|>user\n"
+            # postfix closes user turn + opens assistant + forces think
+            postfix = b"<|im_end|>\n<|im_start|>assistant\n<think>\n"
+            self._runtime._lib.rkllm_set_chat_template(
+                self._runtime._handle, system_wrapped, prefix, postfix,
+            )
+        except Exception as e:  # noqa: BLE001
+            logger.warning("rkllm_set_chat_template failed: %s", e)
+
         next_role: str = "user"
-        next_content: str = first_turn_content
+        next_content: str = prompt
         next_keep_history: int = 0   # 0 on first turn; 1 thereafter
 
         for turn in range(MAX_TURNS):