test(#1743): 修复 smart-match 随机噪声引发的确定性排序断言

CI Unit Tests 暴露（沙箱 httpx2 无法 collect test_smart_match*.py 未本地复现）：
- test_smart_match_fallback test_limit_backfills: second(quality=40)/third(30)
  质量项仅差 4 分 < 噪声 20，third 可翻到首位 → second 质量提至 90（差 24>20），
  200 seeds 验证零翻转
- test_smart_match test_returns_sorted_by_score_descending / test_image_assets:
  smart_select_assets 直接调用处注入零噪声 rng（_ZeroNoiseRandom），验证纯评分排序
- 验证生产噪声行为：大分差素材排名稳定、同分素材组合随机

tests/unit/test_smart_match.py

@@ -1,5 +1,6 @@
 """Tests for packages/domain/smart_match.py — 统一智能选素材算法。"""
 
+import random
 from dataclasses import dataclass, field
 from datetime import datetime, timedelta, timezone
 from typing import Any
@@ -7,6 +8,7 @@ from typing import Any
 import pytest
 
 from packages.domain.smart_match import (
+    SCORE_RANDOM_NOISE_MAX,
     SmartMatchResult,
     _diversity_select,
     _duration_bucket,
@@ -46,6 +48,22 @@ class FakeAsset:
 NOW = datetime(2026, 8, 5, 12, 0, 0, tzinfo=timezone.utc)
 
 
+class _ZeroNoiseRandom(random.Random):
+    """零噪声随机源：uniform(0, NOISE_MAX) 恒返回 0，使「按评分降序」类断言确定可复现。
+
+    smart_select_assets 生产环境注入随机噪声（同分素材每次选出不同组合，#1743）；
+    验证纯评分排序的单测用本随机源消除排序随机性。
+    """
+
+    def uniform(self, a, b):
+        if a == 0.0 and b == SCORE_RANDOM_NOISE_MAX:
+            return 0.0
+        return super().uniform(a, b)
+
+
+_ZERO_NOISE = _ZeroNoiseRandom(0)
+
+
 # ── score_asset tests ────────────────────────────────────────────────────────
 
 
@@ -185,7 +203,7 @@ class TestSmartSelectAssets:
             FakeAsset(id="high", quality_score=95, duration=15),
             FakeAsset(id="mid", quality_score=60, duration=15),
         ]
-        results = smart_select_assets(assets)
+        results = smart_select_assets(assets, rng=_ZERO_NOISE)
         scores = [r.score for r in results]
         assert scores == sorted(scores, reverse=True)
         assert results[0].asset.id == "high"
@@ -234,7 +252,7 @@ class TestSmartSelectAssets:
             FakeAsset(id="img1", mime_type="image/jpeg", quality_score=90, duration=None),
             FakeAsset(id="img2", mime_type="image/png", quality_score=70, duration=None),
         ]
-        results = smart_select_assets(assets, kind="image")
+        results = smart_select_assets(assets, kind="image", rng=_ZERO_NOISE)
         assert len(results) == 2
         assert results[0].asset.id == "img1"
 

tests/unit/test_smart_match_fallback.py

@@ -155,7 +155,9 @@ class TestSmartMatchAvailabilityFallback:
         project = Project(id="proj-1", name="Test", owner_user_id="user-1")
         assets = [
             _video_asset("top-exhausted.mp4", quality=100, used_ranges=_exhausted_ranges(15)),
-            _video_asset("second-fresh.mp4", quality=40, used_ranges=None),
+            # second 质量分显著高于 third（质量项差 (90-30)*0.4=24 > 噪声上限 20），
+            # 排除耗尽素材后 second 稳定排首位回补（噪声不影响大分差排名）
+            _video_asset("second-fresh.mp4", quality=90, used_ranges=None),
             _video_asset("third-fresh.mp4", quality=30, used_ranges=None),
         ]
         repo = _StubAssetRepo(assets)