fix: sanitize numpy/pandas types in submit_solution JSON serialization

Resolves "float object cannot be interpreted as integer" error when
submit_solution tries to serialize pd.Timestamp (from cast_datetime),
numpy.int64 (from df.shape), and numpy.float64 values. Now uses
pandas to_json for cleaned_preview and explicit int/float/bool casts.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

server/cleaning_environment.py

@@ -338,9 +338,9 @@ class FSDSCleaningEnvironment(MCPEnvironment):
             )
             return {
                 "message": message,
-                "reward": round(step_reward, 4),
-                "quality_score": round(after_score, 4),
-                "shape": list(episode.working_df.shape),
+                "reward": round(float(step_reward), 4),
+                "quality_score": round(float(after_score), 4),
+                "shape": [int(x) for x in episode.working_df.shape],
             }
 
         @mcp.tool
@@ -358,6 +358,7 @@ class FSDSCleaningEnvironment(MCPEnvironment):
 
         @mcp.tool
         def submit_solution() -> dict[str, Any]:
+            import json as _json
             episode = self._require_episode()
             report = self._evaluate_quality_gates(episode)
             match_score = self._required_operations_score(episode)
@@ -373,14 +374,18 @@ class FSDSCleaningEnvironment(MCPEnvironment):
             )
             episode.total_reward += final_reward
             episode.submitted = True
+            # Use pandas to_json to safely serialize numpy/pandas types (Timestamps, int64, float64).
+            cleaned_preview = _json.loads(
+                episode.working_df.head(5).to_json(orient="records", date_format="iso")
+            )
             return {
                 "done": True,
-                "passed": report["passed"],
+                "passed": bool(report["passed"]),
                 "final_reward": final_reward,
-                "cumulative_reward": round(episode.total_reward, 4),
+                "cumulative_reward": round(float(episode.total_reward), 4),
                 "quality_report": report,
-                "required_operation_coverage": round(match_score, 4),
-                "cleaned_preview": episode.working_df.head(5).to_dict(orient="records"),
+                "required_operation_coverage": round(float(match_score), 4),
+                "cleaned_preview": cleaned_preview,
             }
 
         super().__init__(mcp)
@@ -566,10 +571,10 @@ class FSDSCleaningEnvironment(MCPEnvironment):
         }
         passed = all(test["passed"] for test in tests.values())
         return {
-            "passed": passed,
+            "passed": bool(passed),
             "tests": tests,
-            "shape": list(df.shape),
-            "retention_ratio": round(retention_ratio, 4),
+            "shape": [int(x) for x in df.shape],
+            "retention_ratio": round(float(retention_ratio), 4),
         }
 
     def _quality_score(self, episode: EpisodeData) -> float: