Fix reward serialization - add reward field to ClaimsObservation

The reward was being calculated but not returned to the client because
ClaimsObservation was missing the reward field. OpenEnv's serialization
looks for observation.reward but it didn't exist.

Changes:
- Add reward field to ClaimsObservation in models.py
- Set observation.reward after step() execution
- Set reward=0.0 on initial reset observation

Tested locally: fraud case +17.40, normal case +13.20

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

models.py

@@ -69,6 +69,9 @@ class ClaimsObservation(Observation):
     is_terminal: bool = Field(default=False, description="Whether episode is done")
     terminal_reason: str = Field(default="", description="Why episode ended")
 
+    # Reward (required for OpenEnv serialization)
+    reward: float = Field(default=0.0, description="Reward from this step")
+
 
 class ClaimsState(State):
     """

server/claims_environment.py

@@ -164,6 +164,7 @@ class ClaimsEnvironment(Environment):
             time_elapsed_minutes=0,
             queries_made=0,
             is_terminal=False,
+            reward=0.0,  # Initial observation has no reward
         )
 
     def step(self, action: ClaimsAction) -> ClaimsObservation:
@@ -196,6 +197,9 @@ class ClaimsEnvironment(Environment):
         self._last_reward = reward
         self._state.total_reward += reward
 
+        # Set reward on observation for client serialization
+        observation.reward = reward
+
         return observation
 
     def _execute_action(self, action: ClaimsAction) -> Tuple[ClaimsObservation, float]: