fix: eliminate double LLM judge call and eval buffer pollution

evaluate() was calling _llm_judge twice per item (once via
compute_reward, once directly) — double the API cost for no benefit.
Now extracts correctness from compute_reward's buffer instead.

Also: compute_reward appends to training metric buffers during eval,
which would pollute wandb training charts. Now rolls back buffer
entries added during eval so training metrics stay clean.

environments/web_research_env.py

@@ -475,14 +475,7 @@ class WebResearchEnv(HermesAgentBaseEnv):
                 )
                 result = await agent.run(messages)
 
-                # Extract final response and compute reward
+                # Extract final response and tool usage from messages
-                ctx = ToolContext(task_id)
-                try:
-                    reward = await self.compute_reward(item, result, ctx)
-                finally:
-                    ctx.cleanup()
-
-                # Extract final response for logging
                 final_response = ""
                 tool_call_count = 0
                 for msg in reversed(result.messages):
@@ -491,12 +484,32 @@ class WebResearchEnv(HermesAgentBaseEnv):
                     if msg.get("role") == "assistant" and msg.get("tool_calls"):
                         tool_call_count += len(msg["tool_calls"])
 
-                # Score correctness separately for the metric
+                # Compute reward (includes LLM judge for correctness)
-                correctness = await self._llm_judge(
+                # Temporarily save buffer lengths so we can extract the
-                    question=item["question"],
+                # correctness score without calling judge twice, and avoid
-                    expected=item["answer"],
+                # polluting training metric buffers with eval data.
-                    model_answer=final_response,
+                buf_len = len(self._correctness_buffer)
+                ctx = ToolContext(task_id)
+                try:
+                    reward = await self.compute_reward(item, result, ctx)
+                finally:
+                    ctx.cleanup()
+
+                # Extract correctness from the buffer (compute_reward appended it)
+                # then remove eval entries from training buffers
+                correctness = (
+                    self._correctness_buffer[buf_len]
+                    if len(self._correctness_buffer) > buf_len
+                    else 0.0
                 )
+                # Roll back buffers to avoid polluting training metrics
+                for buf in (
+                    self._reward_buffer, self._correctness_buffer,
+                    self._tool_usage_buffer, self._efficiency_buffer,
+                    self._diversity_buffer,
+                ):
+                    if len(buf) > buf_len:
+                        buf.pop()
 
                 samples.append({
                     "prompt": item["question"],