feat: AI 模型按场景独立配置（llm_scenes）+ 去重多来源合并

- system.yaml 新增 llm_scenes（translation / daily_report，含用途/方法/模型要求说明）
- load_llm_config(scene=...) 场景覆盖；日报摘要 temperature 0.3 硬编码 → 配置
- M3 去重：唯一篇记录 source_ids（跨源重复合并，首个来源为 source_id）
- source_ids 经翻译透传至 events，日报事件 source 多来源拼接展示（≤3 个）
- 已部署 pi5：merge 实证 10 源合并；日报 report_id=224 正常入库

tests/test_dedup.py

@@ -1,5 +1,6 @@
 """M3 三层去重模块单元测试。"""
 
+import json
 from pathlib import Path
 
 import pytest
@@ -17,6 +18,7 @@ from dedup import (
     normalize_content,
     simhash64,
 )
+from dedup.pipeline import dedup_source
 from extractor.models import ProcessedArticle
 
 # --------------------------------------------------------------------------- #
@@ -531,3 +533,100 @@ class TestDedupResult:
         summary = r.short_summary()
         assert "[DUP/simhash]" in summary
         assert "hd=2" in summary
+
+
+# --------------------------------------------------------------------------- #
+# dedup_source 跨源来源合并（M9.2：最终显示新闻记录多个来源）
+# --------------------------------------------------------------------------- #
+
+
+class TestMergeSources:
+    """dedup_source 在跨源重复时把来源合并进唯一篇。"""
+
+    DATE_STR = "20260805"
+
+    def _write_processed(
+        self,
+        base: Path,
+        source_id: str,
+        article: ProcessedArticle,
+    ) -> None:
+        """写入 data/processed/{source_id}/{date}/{url_hash}.json。"""
+        d = base / "data" / "processed" / source_id / self.DATE_STR
+        d.mkdir(parents=True, exist_ok=True)
+        (d / f"{article.url_hash}.json").write_text(
+            article.model_dump_json(indent=2, ensure_ascii=False), encoding="utf-8"
+        )
+
+    def test_cross_source_merge(self, tmp_path, monkeypatch):
+        """同内容两源报道 → 唯一篇 source_ids 记录两个来源，主来源不变。"""
+        monkeypatch.chdir(tmp_path)  # 隔离 data/ 相对路径与默认指纹库
+
+        content = ("The Federal Reserve kept interest rates unchanged on Wednesday. "
+                   "Markets rallied in response.")
+        art_a = _make_article(
+            url="https://www.reuters.com/business/1", url_hash="aaaa111111111111",
+            source_id="reuters", content=content,
+        )
+        art_b = _make_article(
+            url="https://www.cnbc.com/2026/1", url_hash="bbbb222222222222",
+            source_id="cnbc", source_name="CNBC", content=content,
+        )
+        self._write_processed(tmp_path, "reuters", art_a)
+        self._write_processed(tmp_path, "cnbc", art_b)
+
+        with Deduper() as deduper:
+            dedup_source("reuters", deduper, self.DATE_STR)
+            dedup_source("cnbc", deduper, self.DATE_STR)
+
+        # 唯一篇 = reuters（先处理），跨源重复后 source_ids 合并
+        uniq = tmp_path / "data" / "deduped" / self.DATE_STR / "uniques" / "aaaa111111111111.json"
+        assert uniq.exists()
+        data = json.loads(uniq.read_text(encoding="utf-8"))
+        assert data["source_id"] == "reuters"          # 主来源不变
+        assert data["source_ids"] == ["reuters", "cnbc"]
+        assert (tmp_path / "data" / "deduped" / self.DATE_STR / "uniques"
+                / "bbbb222222222222.json").exists() is False  # 重复篇不单独落盘
+
+    def test_unique_initializes_source_ids(self, tmp_path, monkeypatch):
+        """无重复时唯一篇 source_ids 初始化为 [source_id]。"""
+        monkeypatch.chdir(tmp_path)
+        art = _make_article(url="https://x.com/1", url_hash="cccc333333333333",
+                            source_id="ft", source_name="Financial Times")
+        self._write_processed(tmp_path, "ft", art)
+
+        with Deduper() as deduper:
+            dedup_source("ft", deduper, self.DATE_STR)
+
+        uniq = tmp_path / "data" / "deduped" / self.DATE_STR / "uniques" / "cccc333333333333.json"
+        data = json.loads(uniq.read_text(encoding="utf-8"))
+        assert data["source_ids"] == ["ft"]
+
+    def test_merge_idempotent(self, tmp_path, monkeypatch):
+        """同一来源重复出现多次合并时去重（不产生重复来源）。"""
+        monkeypatch.chdir(tmp_path)
+        content = "Identical content across sources for idempotent test."
+        art_a = _make_article(
+            url="https://www.reuters.com/business/2", url_hash="dddd444444444444",
+            source_id="reuters", content=content,
+        )
+        art_b = _make_article(
+            url="https://www.cnbc.com/2026/2", url_hash="eeee555555555555",
+            source_id="cnbc", source_name="CNBC", content=content,
+        )
+        art_c = _make_article(
+            url="https://www.marketwatch.com/2", url_hash="ffff666666666666",
+            source_id="marketwatch", source_name="MarketWatch", content=content,
+        )
+        self._write_processed(tmp_path, "reuters", art_a)
+        self._write_processed(tmp_path, "cnbc", art_b)
+        self._write_processed(tmp_path, "marketwatch", art_c)
+
+        with Deduper() as deduper:
+            dedup_source("reuters", deduper, self.DATE_STR)
+            dedup_source("cnbc", deduper, self.DATE_STR)
+            dedup_source("marketwatch", deduper, self.DATE_STR)
+
+        uniq = tmp_path / "data" / "deduped" / self.DATE_STR / "uniques" / "dddd444444444444.json"
+        data = json.loads(uniq.read_text(encoding="utf-8"))
+        assert data["source_ids"] == ["reuters", "cnbc", "marketwatch"]

tests/test_llm.py

@@ -463,6 +463,29 @@ class TestLoadLLMConfig:
         config = load_llm_config(provider="deepseek")
         assert config.max_attempts == 3
 
+    def test_scene_translation(self, monkeypatch):
+        """translation 场景覆盖 llm 默认段（模型/温度）。"""
+        monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-deepseek-test-key")
+        config = load_llm_config(provider="deepseek", scene="translation")
+        assert config.model == "deepseek-v4-flash"
+        assert config.temperature == 0.1
+        assert config.max_attempts == 3
+
+    def test_scene_daily_report(self, monkeypatch):
+        """daily_report 场景独立配置（温度 0.3 / max_tokens 1500）。"""
+        monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-deepseek-test-key")
+        config = load_llm_config(scene="daily_report")
+        assert config.provider == "deepseek"
+        assert config.model == "deepseek-v4-flash"
+        assert config.temperature == 0.3
+        assert config.max_tokens == 1500
+
+    def test_scene_unknown_falls_back_to_default(self, monkeypatch):
+        """未定义的场景名回退 llm 默认段，不报错。"""
+        monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-deepseek-test-key")
+        config = load_llm_config(provider="deepseek", scene="not_exists")
+        assert config.temperature == 0.1
+
 
 # --------------------------------------------------------------------------- #
 # translate_and_extract（mock LLM）

tests/test_report_db.py

@@ -177,3 +177,25 @@ class TestBuildReportData:
                                Counter(), "")
         # "?" 不写入 DB，留 None
         assert r.events[0].sentiment is None
+
+    def test_multi_source_label(self) -> None:
+        """去重合并后的多来源 → source 拼接展示（Reuters, CNBC）。"""
+        now = datetime(2026, 8, 4, 8, 0, 0)
+        ev = _fake_high_event("多来源事件", 5, source_id="reuters",
+                              url="https://reuters.com/news/9")
+        # 模拟 M3 去重合并：source_ids 含两个来源
+        ev["article"]["source_ids"] = ["reuters", "cnbc"]
+        r = _build_report_data(now, {}, [ev], Counter(), Counter(), Counter(),
+                               Counter(), "")
+        assert r.events[0].source == "Reuters, CNBC"
+        assert r.events[0].url == "https://reuters.com/news/9"
+
+    def test_single_source_falls_back(self) -> None:
+        """source_ids 为空/单一时回退单源逻辑（不拼接）。"""
+        now = datetime(2026, 8, 4, 8, 0, 0)
+        ev = _fake_high_event("单来源事件", 4, source_id="investinglive",
+                              url="https://investinglive.com/news/3")
+        ev["article"]["source_ids"] = ["investinglive"]
+        r = _build_report_data(now, {}, [ev], Counter(), Counter(), Counter(),
+                               Counter(), "")
+        assert r.events[0].source == "InvestingLive"