runtime: lower temperature 0.6->0.3 + top_k 20->5 for XML adherence

Lab observation 2026-05-27: the model emitted 'diag/summary' as
plain text instead of the trained '<tool_call>{...}</tool_call>'
XML. At temp=0.6/top_k=20 the sampling is loose enough that the
model drifts into narrative prose mode despite the structured-
output training. Tightening to temp=0.3/top_k=5 favours the most-
likely next-token (which, for the trained pattern, is the XML
tag) without going fully greedy (top_k=1 would risk verbose
deterministic loops).

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

Author: ehsan6sha

src/runtime/rkllm_runtime.py

@@ -363,8 +363,16 @@ def init_model(
         # streaming SSE path that emits tokens as they arrive (current
         # generate() blocks until the full turn completes).
         max_new_tokens: int = 768,
-        temperature: float = 0.6,
-        top_k: int = 20,
+        # Lower temperature + tighter top_k for STRUCTURED-OUTPUT
+        # adherence. Lab observation 2026-05-27: at temp=0.6/top_k=20
+        # the model produced narrative prose ("diag/summary") instead
+        # of the trained XML format ("<tool_call>{...}</tool_call>").
+        # The training distribution heavily favours XML in the relevant
+        # positions; lower entropy = the model follows that mode more
+        # consistently. Trade-off is slightly less variety in reasoning
+        # prose, acceptable for a structured tool-calling task.
+        temperature: float = 0.3,
+        top_k: int = 5,
         top_p: float = 0.8,
     ) -> None:
         p = RKLLMParam()