feat(news): index news packages and research persistent events

Extract reusable literal keywords once per source revision, assign complete news packages to long-lived events, and research all event members with immutable evidence and references.

Add additive research storage, vocabulary snapshots, the two-stage workspace, isolated acceptance coverage, and a paused rollout with an initial fixed 50-news production trial.

scripts/reset_news_ai.py

@@ -0,0 +1,61 @@
+"""Inspect or reset only news AI derivatives after isolated sample acceptance."""
+import argparse
+import asyncio
+import hashlib
+import json
+from pathlib import Path
+from sqlalchemy import delete, func, select, text
+from alphapulse.core.db import SessionLocal, engine
+from alphapulse.core.redis import redis
+from alphapulse.models.risk import (RiskSettings, ReprocessTarget, ReprocessItem, ReprocessJob,
+    EventEvidence, EventCorrectionRecord, EventLink, EventVersion, EventMember, NewsEvent,
+    PackageCorrection, PackageEvidence, SummaryVersion, GroupingVersion, SummaryArticle,
+    NewsSummary, ArticleAnalysis, RiskSnapshot, RiskAnalysis, RiskEvent)
+from alphapulse.news_risk.scheduler import GATE
+
+DERIVED = (ReprocessTarget, ReprocessItem, ReprocessJob, EventEvidence, EventCorrectionRecord,
+    EventLink, EventVersion, EventMember, NewsEvent, PackageCorrection, PackageEvidence,
+    SummaryVersion, GroupingVersion, SummaryArticle, NewsSummary, ArticleAnalysis,
+    RiskSnapshot, RiskAnalysis, RiskEvent)
+
+
+def accepted(report, baseline):
+    return (report.get('real_models') is True and report.get('quality_accepted') is True
+        and report.get('baseline_sha256') == hashlib.sha256(baseline).hexdigest()
+        and report.get('baseline_packages') == 20 and report.get('originals') == 27
+        and report.get('failed') == 0 and report.get('unassigned') == 0)
+
+
+async def reset(execute=False, report=None, baseline=None):
+    if execute and (not report or not baseline or not accepted(report,baseline)):
+        raise ValueError('缺少通过质量验收的固定20包真实模型报告；禁止清理')
+    try:
+        async with engine.connect() as gate:
+            if not await gate.scalar(text('SELECT pg_try_advisory_lock(:gate,1)'),{'gate':GATE}):
+                raise ValueError('仍有AI任务运行，不能清理')
+            try:
+                async with SessionLocal() as session:
+                    settings=await session.get(RiskSettings,1,with_for_update=True)
+                    if settings and settings.enabled: raise ValueError('必须先暂停新闻AI')
+                    counts={model.__tablename__:await session.scalar(select(func.count()).select_from(model)) for model in DERIVED}
+                    if execute:
+                        for model in DERIVED: await session.execute(delete(model))
+                        await session.commit()
+                        async for key in redis.scan_iter(match='ap:risk:*',count=100): await redis.delete(key)
+                    return {'executed':execute,'counts':counts}
+            finally:
+                await gate.execute(text('SELECT pg_advisory_unlock(:gate,1)'),{'gate':GATE})
+                await gate.commit()
+    finally:
+        await engine.dispose()
+
+
+if __name__=='__main__':
+    parser=argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--execute',action='store_true')
+    parser.add_argument('--acceptance',type=Path)
+    parser.add_argument('--baseline',type=Path)
+    args=parser.parse_args()
+    print(json.dumps(asyncio.run(reset(args.execute,
+        json.loads(args.acceptance.read_text()) if args.acceptance else None,
+        args.baseline.read_bytes() if args.baseline else None)),ensure_ascii=False))