fixes

Author: mohammedahmed18

codeflash/models/models.py

@@ -52,7 +52,6 @@ class TestDiffScope(str, Enum):
     RETURN_VALUE = "return_value"
     STDOUT = "stdout"
     DID_PASS = "did_pass"  # noqa: S105
-    TIMED_OUT = "timed_out"
 
 
 @dataclass

codeflash/optimization/function_optimizer.py

@@ -461,6 +461,13 @@ def optimize_function(self) -> Result[BestOptimization, str]:
             return Failure(f"No best optimizations found for function {self.function_to_optimize.qualified_name}")
         return Success(best_optimization)
 
+    def reset_optimization_metrics_for_candidate(
+        self, opt_id: str, speedup_ratios: dict, is_correct: dict, optimized_runtimes: dict
+    ) -> None:
+        speedup_ratios[opt_id] = None
+        is_correct[opt_id] = False
+        optimized_runtimes[opt_id] = None
+
     def was_candidate_tested_before(self, normalized_code: str) -> bool:
         # check if this code has been evaluated before by checking the ast normalized code string
         return normalized_code in self.ast_code_to_id
@@ -602,6 +609,9 @@ def determine_best_candidate(
                     "shorter_source_code": candidate.source_code,
                     "diff_len": diff_length(candidate.source_code.flat, code_context.read_writable_code.flat),
                 }
+                self.reset_optimization_metrics_for_candidate(
+                    candidate.optimization_id, speedup_ratios, is_correct, optimized_runtimes
+                )
                 run_results, new_candidate = self.run_optimized_candidate(
                     optimization_candidate_index=candidate_index,
                     baseline_results=original_code_baseline,
@@ -617,9 +627,9 @@ def determine_best_candidate(
 
                 console.rule()
                 if not is_successful(run_results):
-                    optimized_runtimes[candidate.optimization_id] = None
-                    is_correct[candidate.optimization_id] = False
-                    speedup_ratios[candidate.optimization_id] = None
+                    self.reset_optimization_metrics_for_candidate(
+                        candidate.optimization_id, speedup_ratios, is_correct, optimized_runtimes
+                    )
                 else:
                     candidate_result: OptimizedCandidateResult = run_results.unwrap()
                     best_test_runtime = candidate_result.best_test_runtime

codeflash/verification/equivalence.py

@@ -65,8 +65,8 @@ def compare_test_results(original_results: TestResults, candidate_results: TestR
         test_src_code = original_test_result.id.get_src_code(original_test_result.file_name)
         test_diff = TestDiff(
             scope=TestDiffScope.RETURN_VALUE,
-            original_value=f"{original_test_result.return_value!r}",
-            candidate_value=f"{cdd_test_result.return_value!r}",
+            original_value=repr(original_test_result.return_value),
+            candidate_value=repr(cdd_test_result.return_value),
             test_src_code=test_src_code,
             candidate_pytest_error=cdd_pytest_error,
             original_pass=original_test_result.did_pass,
@@ -88,16 +88,15 @@ def compare_test_results(original_results: TestResults, candidate_results: TestR
                 )
             except Exception as e:
                 logger.error(e)
-            break
-        if (original_test_result.stdout and cdd_test_result.stdout) and not comparator(
+        elif (original_test_result.stdout and cdd_test_result.stdout) and not comparator(
             original_test_result.stdout, cdd_test_result.stdout
         ):
             test_diff.scope = TestDiffScope.STDOUT
             test_diff.original_value = str(original_test_result.stdout)
             test_diff.candidate_value = str(cdd_test_result.stdout)
             test_diffs.append(test_diff)
 
-        if original_test_result.test_type in {
+        elif original_test_result.test_type in {
             TestType.EXISTING_UNIT_TEST,
             TestType.CONCOLIC_COVERAGE_TEST,
             TestType.GENERATED_REGRESSION,