feat: 素材选取评分注入随机噪声，降低成片查重率 (#1620)

Co-authored-by: xiaoxia <dev@xiaoxiajianji.com>
Co-committed-by: xiaoxia <dev@xiaoxiajianji.com>

apps/api/app/api/routes/templates_editor/clips.py

@@ -48,7 +48,7 @@ from packages.adapters.sqlalchemy_impl.template_repository import (
     SQLAlchemyTemplateRepository,
 )
 from packages.domain.plan_generator_utils import _calc_random_start_time
-from packages.domain.smart_match import score_asset
+from packages.domain.smart_match import SCORE_RANDOM_NOISE_MAX, score_asset
 from packages.shared.mediakit_client import get_mediakit_client
 
 from .dependencies import get_draft_plan_id, get_editor_services
@@ -744,14 +744,15 @@ def create_clips_from_assets_editor(
         clip_duration = 0.0
         start_time: float | None = None
         # 动态按使用次数排序：优先选使用最少的素材，同次数随机打散
-        asset_use_counts = {
-            aid: len(used_segments.get(aid, []))
-            for aid in asset_ids
-        }
+        asset_use_counts = {aid: len(used_segments.get(aid, [])) for aid in asset_ids}
+        # 排序键：smart_match 评分（注入随机噪声）→ 使用次数 → 纯随机。
+        # 噪声让得分接近的素材排名每次浮动，避免同一批素材反复选出相同组合，
+        # 从素材组合层面降低成片查重率；分差 > SCORE_RANDOM_NOISE_MAX 时排名稳定，
+        # 质量差距显著的素材仍保持优先级。
         sorted_candidates = sorted(
             asset_ids,
             key=lambda aid: (
-                -asset_smart_scores.get(aid, 0.0),
+                -(asset_smart_scores.get(aid, 0.0) + random.uniform(0.0, SCORE_RANDOM_NOISE_MAX)),
                 asset_use_counts.get(aid, 0),
                 random.random(),
             ),
@@ -847,15 +848,15 @@ def create_clips_from_assets_editor(
         duplicate_warning = f"查重率 {dup_rate:.1f}% 超过50%，建议更换素材或模板"
         logger.warning(
             "from-assets 成片查重率超标: plan_id=%s dup_rate=%.1f%%",
-            plan_id, dup_rate,
+            plan_id,
+            dup_rate,
         )
 
     # 7. 素材耗尽提示
     exhaustion_warning = None
     if all_assets_exhausted and created_count < len(segments):
         exhaustion_warning = (
-            "素材可切区间不足，部分片段使用了复用素材。"
-            "建议：1) 补充更多素材到素材库 2) 使用不同的素材组合生成"
+            "素材可切区间不足，部分片段使用了复用素材。" "建议：1) 补充更多素材到素材库 2) 使用不同的素材组合生成"
         )
 
     # 8. 立即返回响应
@@ -1022,9 +1023,11 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
                     if cid != clip_id_inner and cid not in updated_clip_ids:
                         segs.append((c.start_time, c.start_time + c.duration))
                 segs.extend(updated_segments.get(asset_id_inner, []))
+
                 # 并入历史已用区间
                 def _norm(segs_in):
                     return {(round(float(a), 3), round(float(b), 3)) for a, b in segs_in}
+
                 return list(_norm(segs) | _norm(historical_segments.get(asset_id_inner, [])))
 
             # 优先使用 SceneChange 策略
@@ -1035,7 +1038,8 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
                     scene_segments = _build_scene_segments(scene_changes, asset_total)
                     logger.info(
                         "后台任务: 素材场景检测完成: asset_id=%s scenes=%d",
-                        asset_id, len(scene_segments),
+                        asset_id,
+                        len(scene_segments),
                     )
 
             # SceneChange 未获得有效结果 → 尝试 analyze_videos 作为 fallback
@@ -1047,7 +1051,8 @@ def _update_mediakit_recommendations_async(  # pragma: no cover
                     scene_segments = [(rec_start, asset_total)]
                     logger.info(
                         "后台任务: 使用 analyze_videos fallback: asset_id=%s start=%.2f",
-                        asset_id, rec_start,
+                        asset_id,
+                        rec_start,
                     )
 
             if not scene_segments: