feat: organize news packages into verified event timelines

Add optional events, evidence-backed timelines, versioned package corrections and bounded historical reprocessing. Keep independent packages readable and isolate retries by failed stage. Configure reasoning effort per model stage and refresh default news Skill drafts without publishing.

backend/tests/test_news_pipeline.py

@@ -12,24 +12,24 @@ def pipeline(trading):
 
 
 def match(pipeline, **changes):
-    return dict(event_id=None, group_key='event', title='公告', category='macro', entities=['央行'],
+    return dict(package_id=None, group_key='event', title='公告', category='macro', entities=['央行'],
         fact='公告', evidence='公告', urgent=False, tags=['央行'], content_kind='report', **changes)
 
 
-@pytest.mark.parametrize('event_id,group_key', [(None, None), (1, 'event')])
-def test_grouping_target_is_exclusive(pipeline, event_id, group_key):
+@pytest.mark.parametrize('package_id,group_key', [(None, None), (1, 'event')])
+def test_grouping_target_is_exclusive(pipeline, package_id, group_key):
     data = match(pipeline)
-    data.update(event_id=event_id, group_key=group_key)
+    data.update(package_id=package_id, group_key=group_key)
     with pytest.raises(ValueError): pipeline.Match(**data)
 
 
 def test_all_domains_require_nonempty_distinct_packages(pipeline):
     item = pipeline.Match(**match(pipeline))
     with pytest.raises(ValueError, match='重复关联'):
-        pipeline.Grouping(reason='政治公告', events=[item, item])
+        pipeline.Grouping(reason='政治公告', packages=[item, item])
     with pytest.raises(ValueError):
-        pipeline.Grouping(reason='说明', events=[])
-    assert pipeline.Grouping(reason='文化说明', events=[item]).events
+        pipeline.Grouping(reason='说明', packages=[])
+    assert pipeline.Grouping(reason='文化说明', packages=[item]).packages
 
 
 @pytest.mark.parametrize('scope,enabled,published', [('shared', True, 1), ('news', False, 1), ('news', True, None)])
@@ -61,7 +61,7 @@ def test_old_backfill_is_not_selected_by_ingestion_time(pipeline):
 def event_result(pipeline, refs):
     return pipeline.EventResult(id=1, title='Synthetic event', summary='Synthetic summary', category='macro', entities=[],
         risk_stars=1, importance_stars=4, direction='unknown', verification='insufficient', confidence='low', reasoning='Synthetic', uncertainty='Synthetic',
-        facts='Synthetic', timeline='Synthetic', asset_impacts='Unknown', opinions='Unknown', counter_evidence='Unknown', watchpoints='Unknown', claims=[pipeline.Claim(statement='Synthetic', verdict='insufficient', evidence_refs=[])], related_packages=[], excluded_article_ids=[], evidence_refs=refs)
+        timeline_points=[dict(occurred_at=None, time_precision='unknown', time_text='Unknown', reported_at=None, description='Synthetic', verification='insufficient', evidence_refs=[pipeline.Evidence.model_validate(r).model_dump() for r in refs])], facts='Synthetic', timeline='Synthetic', asset_impacts='Unknown', opinions='Unknown', counter_evidence='Unknown', watchpoints='Unknown', claims=[pipeline.Claim(statement='Synthetic', verdict='insufficient', evidence_refs=[])], related_packages=[], excluded_article_ids=[], evidence_refs=refs)
 
 
 def test_long_event_uses_bounded_layers_and_retains_all_sources(pipeline, monkeypatch):
@@ -152,8 +152,8 @@ def test_package_scope_survives_direct_and_extracted_roundup(pipeline, monkeypat
     observed = []
     async def complete(provider, system, payload, schema):
         observed.append(payload)
-        if schema is pipeline.Extraction:
-            return pipeline.Extraction(facts=payload['package']['fact'], evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
+        if payload['task'] == 'extract':
+            return schema(facts=payload['package']['fact'], evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
         return event_result(pipeline, [pipeline.Evidence(article_id=1, quote='b')]).model_copy(update={'id': payload['id']}), {}
     monkeypatch.setattr(pipeline, 'complete', complete)
     for identifier, fact in [(1, '央行公布利率'), (2, '公司发布新产品')]:
@@ -188,3 +188,53 @@ def test_long_event_reserves_real_research_tools_before_extraction(pipeline, mon
     monkeypatch.setattr(research.llm, 'stream', stream)
     result, _ = asyncio.run(pipeline.summarize_articles(provider, 'rules', [article], package, review))
     assert result.id == 1 and extracted
+
+
+@pytest.mark.parametrize('has_related_fact', [True, False])
+def test_package_roundup_skips_empty_slices_but_requires_related_evidence(pipeline, monkeypatch, has_related_fact):
+    provider = SimpleNamespace(max_tokens=1000, context_window=18000)
+    article = {'id': 1, 'title': 'Roundup', 'summary': '',
+        'content': ('RATE decision. ' if has_related_fact else '')+'Unrelated product announcement. '*1500}
+    package = {'id': 1, 'fact': 'RATE decision'}
+    system = pipeline.BASE+'\n'+pipeline.TEMPLATES['analysis'][1]
+    extracted, submitted = [], []
+
+    async def complete(provider, system, payload, schema):
+        if payload['task'] == 'extract':
+            related = 'RATE' in payload['content']
+            extracted.append(related)
+            return schema.model_validate({'facts': 'RATE decision' if related else '',
+                'evidence_refs': [{'article_id': 1, 'quote': 'RATE'}] if related else []}), {'input_tokens': 1}
+        assert payload['task'] == 'event'
+        submitted.append(payload)
+        assert all(e['facts'] and e['evidence_refs'] for e in payload['extractions'])
+        return event_result(pipeline, [pipeline.Evidence(article_id=1, quote='RATE')]), {}
+
+    monkeypatch.setattr(pipeline, 'complete', complete)
+    if has_related_fact:
+        result, usage = asyncio.run(pipeline.summarize_articles(provider, system, [article], package))
+        assert result.evidence_refs[0].quote == 'RATE' and len(submitted) == 1
+        assert len(usage['parts']) == len(extracted)
+        assert len(submitted[0]['extractions']) == sum(extracted)
+    else:
+        with pytest.raises(ValueError, match='未提取到当前包相关证据'):
+            asyncio.run(pipeline.summarize_articles(provider, system, [article], package))
+        assert not submitted
+    assert len(extracted) > 1 and False in extracted
+
+
+@pytest.mark.parametrize('facts,refs', [('Unsupported fact', []), ('', [{'article_id': 1, 'quote': 'Original'}])])
+def test_package_extraction_requires_facts_and_evidence_together(pipeline, facts, refs):
+    with pytest.raises(ValueError, match='分段事实与证据必须同时为空或同时提供'):
+        pipeline.PackageExtraction(facts=facts, evidence_refs=refs)
+
+
+def test_grouping_extraction_and_final_summary_still_require_evidence(pipeline, monkeypatch):
+    async def complete(provider, system, payload, schema):
+        return schema.model_validate({'facts': '', 'evidence_refs': []}), {}
+    monkeypatch.setattr(pipeline, 'complete', complete)
+    with pytest.raises(ValueError):
+        asyncio.run(pipeline.extract_articles(SimpleNamespace(max_tokens=100, context_window=12000), 'rules',
+            [{'id': 1, 'title': 'Original', 'content': '', 'summary': ''}]))
+    with pytest.raises(ValueError):
+        event_result(pipeline, [])