fix: Isolate tracking failures and log failed judge evaluations

Wrap tracker.track_judge_result() in try/except inside _run_and_track
so a tracking failure (e.g., shutdown LD client) does not propagate
through the wrapper task and destroy the evaluation results that were
successfully computed. Restores the resilience of the previous
add_done_callback approach, which asyncio handled with isolation.

Also log a warning when a judge evaluation fails (r.success is False)
so failures are visible rather than silently skipped.

Author: jsonbailey

packages/sdk/server-ai/src/ldai/managed_model.py

@@ -1,6 +1,7 @@
 import asyncio
 from typing import List, Optional
 
+from ldai import log
 from ldai.models import AICompletionConfig, LDMessage
 from ldai.providers.model_runner import ModelRunner
 from ldai.providers.types import JudgeResult, ModelResponse
@@ -68,7 +69,16 @@ async def _run_and_track(eval_task: asyncio.Task) -> List[JudgeResult]:
             results = await eval_task
             for r in results:
                 if r.success:
-                    tracker.track_judge_result(r)
+                    try:
+                        tracker.track_judge_result(r)
+                    except Exception:
+                        pass
+                else:
+                    log.warning(
+                        'Judge evaluation failed for %r: %s',
+                        r.judge_config_key,
+                        r.error_message,
+                    )
             return results
 
         return asyncio.create_task(_run_and_track(evaluator_task))