Fix research context limits and preserve output correction audits

backend/tests/test_news_risk.py

@@ -36,3 +36,38 @@ def test_context_budget_includes_unicode_and_tool_schema(risk):
     from types import SimpleNamespace
     provider = SimpleNamespace(max_tokens=256, context_window=3500)
     assert not risk.fits_context(provider, '只使用原文证据', {'content': '风'*1000}, IndexResult)
+
+
+@pytest.mark.parametrize('context_window', [250000, 1000000])
+def test_news_context_is_capped_at_250k_including_output(risk, monkeypatch, context_window):
+    from pydantic import BaseModel
+    from types import SimpleNamespace
+    from unittest.mock import AsyncMock
+
+    class Result(BaseModel):
+        facts: str
+
+    provider = SimpleNamespace(max_tokens=32000, context_window=context_window)
+    assert risk.fits_context(provider, '', {'content': 'x' * 200000}, Result)
+    stream = AsyncMock()
+    monkeypatch.setattr(risk.llm, 'stream', stream)
+    with pytest.raises(ValueError, match='超出模型上下文'):
+        asyncio.run(risk.complete(provider, '', {'content': 'x' * 220000}, Result))
+    stream.assert_not_called()
+
+
+def test_research_reserves_tool_room_within_250k_cap(risk):
+    from alphapulse.news_research.tools import Research
+    from pydantic import BaseModel
+    from types import SimpleNamespace
+
+    class Result(BaseModel):
+        facts: str
+
+    reader = Research()
+    payload = {'content': 'x' * 190000}
+    for context_window in (250000, 1000000):
+        provider = SimpleNamespace(max_tokens=32000, context_window=context_window)
+        assert risk.fits_context(provider, '', payload, Result)
+        assert not reader.fits(provider, '', payload, Result)
+        assert provider.context_window == context_window