修复hard数据集误判

@@ -450,9 +450,9 @@ python -m lyric_dedup.cli generate-eval-set \
 ```bash
 python scripts/evaluate_postgres.py \
   --dsn postgresql:///lyric_dedup \
-  --csv data/generated_eval/eval_5000.csv \
+  --csv data/generated_eval/eval_hard_5000.csv \
   --base-dir data/generated_eval \
-  --out outputs/results/postgres_eval_5000.csv
+  --out outputs/results/postgres_eval_hard_5000.csv
 ```
 
 它会：

@@ -220,7 +220,11 @@ def _recall_candidates(
 
 
 def _exclude_record_ids_for_eval_row(row: dict[str, str]) -> list[str]:
-    if row.get("sample_type") == "negative_real_holdout_full_song" and row.get("source_record_id"):
+    holdout_sample_types = {
+        "negative_real_holdout_full_song",
+        "negative_near_neighbor_holdout_full_song",
+    }
+    if row.get("sample_type") in holdout_sample_types and row.get("source_record_id"):
         return [row["source_record_id"]]
     return []