feat: v0.2.0 sprint — ground truth eval, crossover/mutation, checkpointing, similarity guards, dataset loader, CLI commands, extended test coverage

Aggregates all v0.2.0 sprint work (GARAA-30 through GARAA-40) and fixes
2 integration tests that broke when the codebase went async (DSPyLLMAdapter
and full pipeline tests now properly await coroutines).

277 tests pass (260 unit + 17 integration).

Co-Authored-By: Paperclip <noreply@paperclip.ing>

src/prometheus/infrastructure/judge_adapter.py

@@ -59,6 +59,7 @@ class DSPyJudgeAdapter(JudgePort):
             if self._judge_dimensions
             else {}
         )
+        self.call_count: int = 0
 
     async def judge_batch(
         self,
@@ -104,13 +105,15 @@ class DSPyJudgeAdapter(JudgePort):
 
     def _sync_judge(self, task_description: str, input_text: str, output_text: str):
         with dspy.context(lm=self._lm):
-            return self._judge(
+            result = self._judge(
                 task_description=task_description,
                 input_text=input_text,
                 output_text=output_text,
                 judge_criteria=self._judge_criteria,
                 dimension_names=self._dimension_names,
             )
+        self.call_count += 1
+        return result
 
     def _aggregate_result(self, pred: Any) -> tuple[float, str]:
         """Compute weighted aggregate score from dimension scores if available."""