fix: enforce forecast attribution and allergen advice

docs/design.md

@@ -54,7 +54,7 @@ service 将最终解析得到的完整地点名作为 `location_scope.full_name`
 
 QWeather 认证由独立的 `QWeatherJWTAuthenticator` 负责。它从运行环境读取 Base64 编码的 Ed25519 PKCS#8 私钥 PEM，解码后通过 EdDSA 签发 JWT：Header 只加入凭据 ID `kid`，Payload 只加入项目 ID `sub`、提前 30 秒的 `iat` 和可配置的短期 `exp`。provider 只依赖认证协议生成 Bearer header，不接触私钥字段，也不支持把长期 API KEY 混入同一请求。
 
-上下文快照被转换为带稳定 source ID 和验证 URL 的文档，与 RSS 输入使用同一套引用校验。08:00 的模型提示要求同时输出 AQI 标准，禁止换算或混用标准，并参考健康提示和生活指数生成运动、穿衣及口罩建议；PM2.5 直接按“PM2.5 数值 单位”表达，过敏指数和花粉只进入生活建议。天气与空气质量文档使用同一个 provider 展示名称。
+上下文快照被转换为带稳定 source ID 和验证 URL 的文档，与 RSS 输入使用同一套引用校验。08:00 的模型提示要求同时输出 AQI 标准，禁止换算或混用标准，并参考健康提示和生活指数生成运动、穿衣及口罩建议；PM2.5 直接按“PM2.5 数值 单位”表达，过敏指数和花粉只进入生活建议。建议使用结构化 topic，程序强制每日预报覆盖穿衣、除湿、运动和口罩；当前输入包含综合过敏指数或花粉过敏原时，同时强制覆盖过敏建议。QWeather adapter 根据供应商的过敏指数代码标记天气文档，Open-Meteo 花粉文档直接携带相同标记，核心编排不依赖供应商文案识别过敏数据。天气与空气质量文档使用同一个 provider 展示名称。
 
 ## 状态
 
@@ -86,7 +86,9 @@ SQLite 没有原生日期时间类型，状态存储需要直接对 TEXT 做范
 
 ## 来源引用
 
-输入文章和辅助 API 响应都带唯一 source ID、展示名称与 URL。模型输出 JSON 中每条结论包含 `source_ids`；程序验证所有 ID 都存在，再生成平台无关的 `BriefingResult`。各投递 provider 的 renderer 从相同结构化结果生成自己的链接与排版，并始终以 API provider 名称或 RSS 配置中的 `name` 作为链接文字；RSS 配置应使用公众号、微博账号或发布机构等公开名称。引用未知来源会使任务失败，禁止模型编造链接。
+输入文章和辅助 API 响应都带唯一 source ID、展示名称与 URL。模型输出 JSON 中标题、概述和每条结论都包含 `source_ids`；程序验证引用非空且所有 ID 都存在，再生成平台无关的 `BriefingResult`。各投递 provider 的 renderer 从相同结构化结果生成自己的链接与排版，在标题、概述和各分项后明确标注“来源”，并始终以 API provider 名称或 RSS 配置中的 `name` 作为链接文字；RSS 配置应使用公众号、微博账号或发布机构等公开名称。引用未知来源或遗漏标题、概述来源会使任务进入 LLM 修复重试，禁止模型编造链接。
+
+同一时段的天气现象在来源间冲突时，模型不得把片段拼成无争议的单一结论。契约要求明确呈现差异，并在能够识别当地权威气象机构时优先采用其最新信息，同时保留冲突来源引用；核心编排不硬编码地区机构名称或供应商优先级。
 
 ## 平台模板与投递
 

tests/test_allergen.py

@@ -68,6 +68,7 @@ def test_allergen_to_document_format() -> None:
 
     assert document.id == "allergen:test"
     assert document.name == "Test 花粉"
+    assert document.has_allergen_information
     assert "桦木：15 粒/m³（中）" in document.content
     assert "禾本：3 粒/m³（低）" in document.content
     assert "总体等级：中" in document.content

tests/test_llm.py

@@ -14,7 +14,9 @@
 def test_rejects_model_invented_source() -> None:
     payload = {
         "headline": "Briefing",
+        "headline_source_ids": ["source"],
         "overview": "Overview",
+        "overview_source_ids": ["source"],
         "conclusions": [{"text": "Claim", "source_ids": ["invented"]}],
         "active_warnings": [],
         "resolved_warning_ids": [],
@@ -32,7 +34,9 @@ def test_rejects_model_invented_source() -> None:
 def test_accepts_suppressed_message_with_unchanged_active_warning() -> None:
     payload = {
         "headline": "Briefing",
+        "headline_source_ids": ["source"],
         "overview": "Overview",
+        "overview_source_ids": ["source"],
         "conclusions": [],
         "active_warnings": [
             {
@@ -131,6 +135,51 @@ def test_rejects_conclusion_without_source_ids() -> None:
         )
 
 
+@pytest.mark.parametrize("source_ids", [None, "source", [None], [""]])
+def test_rejects_malformed_source_ids(source_ids) -> None:
+    payload = {
+        "headline": "Briefing",
+        "overview": "Overview",
+        "conclusions": [{"text": "Claim", "source_ids": source_ids}],
+        "active_warnings": [],
+        "resolved_warning_ids": [],
+        "advice": [],
+        "disaster_tracking": [],
+    }
+
+    with pytest.raises(LLMError, match="source_ids must"):
+        parse_result(
+            payload,
+            pendulum.datetime(2026, 7, 13, 9, tz="Asia/Shanghai"),
+            {"source"},
+        )
+
+
+@pytest.mark.parametrize("key", ["conclusions", "disaster_tracking", "advice"])
+@pytest.mark.parametrize("text", [None, "", "   ", 42])
+def test_rejects_sourced_item_without_non_empty_text(key: str, text) -> None:
+    item = {"text": text, "source_ids": ["source"]}
+    if key == "advice":
+        item["topic"] = "clothing"
+    payload = {
+        "headline": "Briefing",
+        "overview": "Overview",
+        "conclusions": [],
+        "active_warnings": [],
+        "resolved_warning_ids": [],
+        "advice": [],
+        "disaster_tracking": [],
+        key: [item],
+    }
+
+    with pytest.raises(LLMError, match=f"{key} entries must contain non-empty text"):
+        parse_result(
+            payload,
+            pendulum.datetime(2026, 7, 13, 9, tz="Asia/Shanghai"),
+            {"source"},
+        )
+
+
 def test_rejects_active_warnings_not_an_array() -> None:
     payload = {
         "headline": "Briefing",
@@ -227,6 +276,75 @@ def test_rejects_non_boolean_should_publish() -> None:
         )
 
 
+@pytest.mark.parametrize("field", ("headline_source_ids", "overview_source_ids"))
+def test_rejects_summary_without_source_ids(field: str) -> None:
+    payload = {
+        "headline": "Briefing",
+        "headline_source_ids": ["source"],
+        "overview": "Overview",
+        "overview_source_ids": ["source"],
+        "conclusions": [],
+        "active_warnings": [],
+        "resolved_warning_ids": [],
+        "advice": [],
+        "disaster_tracking": [],
+    }
+    del payload[field]
+
+    with pytest.raises(LLMError, match=rf"{field} must cite"):
+        parse_result(
+            payload,
+            pendulum.datetime(2026, 7, 13, 9, tz="Asia/Shanghai"),
+            {"source"},
+        )
+
+
+def test_rejects_advice_without_a_valid_topic() -> None:
+    payload = {
+        "headline": "Briefing",
+        "headline_source_ids": ["source"],
+        "overview": "Overview",
+        "overview_source_ids": ["source"],
+        "conclusions": [],
+        "active_warnings": [],
+        "resolved_warning_ids": [],
+        "advice": [{"text": "Advice", "source_ids": ["source"]}],
+        "disaster_tracking": [],
+    }
+
+    with pytest.raises(LLMError, match="must use a valid topic"):
+        parse_result(
+            payload,
+            pendulum.datetime(2026, 7, 13, 9, tz="Asia/Shanghai"),
+            {"source"},
+        )
+
+
+@pytest.mark.parametrize(
+    ("advice", "message"),
+    (("not-an-array", "advice must be an array"), (["not-an-object"], "advice entries must be objects")),
+)
+def test_rejects_invalid_advice_structure(advice: object, message: str) -> None:
+    payload = {
+        "headline": "Briefing",
+        "headline_source_ids": ["source"],
+        "overview": "Overview",
+        "overview_source_ids": ["source"],
+        "conclusions": [],
+        "active_warnings": [],
+        "resolved_warning_ids": [],
+        "advice": advice,
+        "disaster_tracking": [],
+    }
+
+    with pytest.raises(LLMError, match=message):
+        parse_result(
+            payload,
+            pendulum.datetime(2026, 7, 13, 9, tz="Asia/Shanghai"),
+            {"source"},
+        )
+
+
 async def test_openai_provider_rejects_empty_json_content() -> None:
     async with httpx.AsyncClient(
         transport=httpx.MockTransport(

tests/test_prompts.py

@@ -20,3 +20,10 @@ def test_prompt_separates_advice_and_avoids_repetition() -> None:
     assert "过敏原信息只能放入 advice" in SYSTEM_PROMPT
     assert "不得使用“原始浓度”" in SYSTEM_PROMPT
     assert "不得在 conclusions 中重复" in SYSTEM_PROMPT
+
+
+def test_prompt_requires_attribution_and_preserves_source_conflicts() -> None:
+    assert "headline_source_ids、overview_source_ids" in SYSTEM_PROMPT
+    assert "不得拼接成无争议的单一结论" in SYSTEM_PROMPT
+    assert "优先采用可识别的当地权威气象机构" in SYSTEM_PROMPT
+    assert "input.required_advice_topics" in SYSTEM_PROMPT

tests/test_reference_data.py

@@ -4,6 +4,7 @@
 from weather_briefing.reference_data import (
     ReferenceDataError,
     load_reference_data,
+    reference_string,
     reference_string_tuple,
     reference_value,
 )
@@ -17,6 +18,7 @@ def test_packaged_reference_data_is_available() -> None:
     )
     assert reference_string_tuple("content_cleaning.json", "default_remove_selectors")
     assert reference_string_tuple("provider_defaults.json", "qweather_lifestyle_index_types")
+    assert reference_string("provider_defaults.json", "qweather_allergen_index_type") == "7"
 
 
 def test_air_quality_guidance_covers_values_above_last_bounded_band() -> None:
@@ -51,6 +53,14 @@ def test_reference_string_tuple_rejects_non_list_value() -> None:
         reference_string_tuple("geography.json", "mainland_china_service_bounds")
 
 
+@pytest.mark.parametrize("value", [None, "", "   ", 7])
+def test_reference_string_rejects_invalid_value(monkeypatch, value) -> None:
+    monkeypatch.setattr("weather_briefing.reference_data.reference_value", lambda *args: value)
+
+    with pytest.raises(ReferenceDataError, match="non-empty string"):
+        reference_string("provider_defaults.json", "qweather_allergen_index_type")
+
+
 def test_load_reference_data_rejects_non_dict_root(monkeypatch) -> None:
     from weather_briefing.reference_data import load_reference_data
 

tests/test_render.py

@@ -1,6 +1,15 @@
 import pendulum
-
-from weather_briefing.models import Article, BriefingResult, Conclusion, SourceDocument, Warning
+import pytest
+
+from weather_briefing.models import (
+    Advice,
+    AdviceTopic,
+    Article,
+    BriefingResult,
+    Conclusion,
+    SourceDocument,
+    Warning,
+)
 from weather_briefing.render import PlainTextRenderer, TelegramHTMLRenderer
 
 
@@ -18,30 +27,56 @@ def test_render_briefing_uses_safe_telegram_html() -> None:
     result = BriefingResult(
         "Daily <Forecast>",
         "Warm & humid",
+        ("source",),
+        ("source",),
         (Conclusion("Carry an umbrella", ("source",)),),
     )
 
     rendered = TelegramHTMLRenderer().render_briefing(result, (article,), ())
 
-    assert rendered.body.startswith("<b>Daily &lt;Forecast&gt;</b>")
+    assert rendered.body.startswith("<b>Daily &lt;Forecast&gt;</b> （来源：")
     assert "Warm &amp; humid" in rendered.body
     assert '<a href="https://example.invalid/?a=1&amp;b=2">Feed</a>' in rendered.body
-    assert rendered.visible_length == len("Daily <Forecast>\n\nWarm & humid\n\n天气信息\n\n• Carry an umbrella Feed")
+    assert rendered.visible_length == len(
+        "Daily <Forecast> （来源：Feed）\n\n"
+        "Warm & humid （来源：Feed）\n\n"
+        "天气信息\n\n• Carry an umbrella （来源：Feed）"
+    )
 
 
 def test_plain_text_renderer_uses_the_same_structured_briefing() -> None:
-    result = BriefingResult("Daily", "Overview", ())
+    context = SourceDocument("source", "Source", "https://example.invalid/source", "")
+    result = BriefingResult("Daily", "Overview", ("source",), ("source",), ())
 
-    rendered = PlainTextRenderer().render_briefing(result, (), ())
+    rendered = PlainTextRenderer().render_briefing(result, (), (context,))
 
-    assert rendered.body == "Daily\n\nOverview"
+    assert rendered.body == (
+        "Daily （来源：Source: https://example.invalid/source）\n\n"
+        "Overview （来源：Source: https://example.invalid/source）"
+    )
     assert "<b>" not in rendered.body
 
 
+@pytest.mark.parametrize("renderer", (TelegramHTMLRenderer(), PlainTextRenderer()))
+def test_renderers_fail_when_a_source_reference_is_missing(
+    renderer: TelegramHTMLRenderer | PlainTextRenderer,
+) -> None:
+    result = BriefingResult("Daily", "Overview", ("missing",), ("missing",), ())
+
+    with pytest.raises(KeyError, match="missing"):
+        renderer.render_briefing(result, (), ())
+
+
 def test_renderers_fall_back_to_source_id_for_legacy_blank_name() -> None:
     now = pendulum.datetime(2026, 7, 11, 8, tz="Asia/Shanghai")
     article = Article("source", "feed", "  ", "Title", "https://example.invalid/a", now, "Body")
-    result = BriefingResult("Daily", "Overview", (Conclusion("Update", ("source",)),))
+    result = BriefingResult(
+        "Daily",
+        "Overview",
+        ("source",),
+        ("source",),
+        (Conclusion("Update", ("source",)),),
+    )
 
     html = TelegramHTMLRenderer().render_briefing(result, (article,), ())
     plain = PlainTextRenderer().render_briefing(result, (article,), ())
@@ -60,8 +95,10 @@ def test_telegram_html_renderer_uses_context_source_name_as_attribution() -> Non
     result = BriefingResult(
         "Daily",
         "Overview",
+        ("allergen:open-meteo",),
+        ("allergen:open-meteo",),
         (),
-        advice=(Conclusion("花粉浓度较高", ("allergen:open-meteo",)),),
+        advice=(Advice(AdviceTopic.ALLERGEN, "花粉浓度较高", ("allergen:open-meteo",)),),
     )
 
     rendered = TelegramHTMLRenderer().render_briefing(result, (), (context,))
@@ -81,8 +118,10 @@ def test_plain_text_renderer_uses_context_source_name_as_attribution() -> None:
     result = BriefingResult(
         "Daily",
         "Overview",
+        ("allergen:open-meteo",),
+        ("allergen:open-meteo",),
         (),
-        advice=(Conclusion("花粉浓度较高", ("allergen:open-meteo",)),),
+        advice=(Advice(AdviceTopic.ALLERGEN, "花粉浓度较高", ("allergen:open-meteo",)),),
     )
 
     rendered = PlainTextRenderer().render_briefing(result, (), (context,))
@@ -97,6 +136,8 @@ def test_telegram_html_renders_active_warnings_section() -> None:
     result = BriefingResult(
         "Daily",
         "Overview",
+        ("source",),
+        ("source",),
         (),
         active_warnings=(warning,),
     )
@@ -115,6 +156,8 @@ def test_plain_text_renders_active_warnings_section() -> None:
     result = BriefingResult(
         "Daily",
         "Overview",
+        ("source",),
+        ("source",),
         (),
         active_warnings=(warning,),
     )