In the grader persistence path, record_ai_result() is called when either grade or feedback is present, but a missing grade results in score=0.0 being recorded (overwriting any existing draft score). Since the grader contract allows grade to be null, avoid calling record_ai_result unless grade is provided (or preserve the existing draft_score instead of forcing 0.0).

[azapg]

In the grader persistence path, record_ai_result() is called when either grade or feedback is present, but a missing grade results in score=0.0 being recorded (overwriting any existing draft score). Since the grader contract allows grade to be null, avoid calling record_ai_result unless grade is provided (or preserve the existing draft_score instead of forcing 0.0).

                    grade = item.get("grade")
                    feedback = item.get("feedback")
                    row.score = grade
                    row.feedback = feedback
                    row.graded_at = _utc_now()
                    row.grading_meta = {
                        **(row.grading_meta or {}),
                        "grading_metadata": item.get("metadata", {}),
                    }
                    # Only record an AI result when a concrete grade is provided.
                    # This avoids overwriting any existing draft score with a synthetic 0.0
                    # when extensions send feedback without a numeric grade.
                    if grade is not None:
                        manager.record_ai_result(
                            submission_id=submission.id,
                            score=float(grade),
                            feedback=feedback or "",

Originally posted by @Copilot in #176 (comment)