fix(evals): add tool_choice=auto + fix traces thread safety (#163)

Two fixes for eval accuracy and stability:

1. TauBench agent: add tool_choice="auto" to match tau2's native
   LLMAgent behavior. Without this, Qwen and GPT-5.4 score 10-14pp
   below leaderboard because the models don't receive explicit
   tool-calling guidance.

2. SystemBuilder: apply self._traces flag to config.traces.enabled.
   Previously builder.traces(False) was a no-op — traces stayed
   enabled, creating SQLite connections in the main thread that
   crashed when accessed from ThreadPoolExecutor worker threads
   in GAIA evals ("SQLite objects created in a thread can only be
   used in that same thread").

Co-authored-by: Jon Saad-Falcon <jonsaadfalcon@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/openjarvis/evals/execution/taubench_env.py

@@ -177,6 +177,7 @@ class JarvisHalfDuplexAgent:
             "temperature": self._temperature,
             "max_tokens": self._max_tokens,
             "tools": openai_tools,
+            "tool_choice": "auto",
         }
         # Disable thinking mode for local models (Qwen3.5 etc.)
         # to avoid <think> tags interfering with tool call parsing

src/openjarvis/system.py

@@ -503,10 +503,15 @@ class SystemBuilder:
         # Resolve model
         model = self._resolve_model(config, engine)
 
-        # Compute telemetry_enabled once
+        # Compute telemetry_enabled and traces_enabled once
         telemetry_enabled = (
             self._telemetry if self._telemetry is not None else config.telemetry.enabled
         )
+        traces_enabled = (
+            self._traces if self._traces is not None else config.traces.enabled
+        )
+        # Apply traces flag to config so downstream code respects it
+        config.traces.enabled = traces_enabled
         gpu_monitor = None
         energy_monitor = None
         if telemetry_enabled and config.telemetry.gpu_metrics: