fix(test): widen hallucination detection tolerance (#809) (#810)

Logprob-derived scores drift ~0.036 across runs due to inference
non-determinism. Widen from abs=3e-2 to abs=5e-2 to absorb jitter
while still catching real regressions.

Author: planetf1

test/stdlib/components/intrinsic/test_rag.py

@@ -158,12 +158,12 @@ def test_hallucination_detection(backend):
     result = rag.flag_hallucinated_content(assistant_response, docs, context, backend)
     # pytest.approx() chokes on lists of records, so we do this complicated dance.
     for r, e in zip(result, expected, strict=True):  # type: ignore
-        assert pytest.approx(r, abs=3e-2) == e
+        assert pytest.approx(r, abs=5e-2) == e
 
     # Second call hits a different code path from the first one
     result = rag.flag_hallucinated_content(assistant_response, docs, context, backend)
     for r, e in zip(result, expected, strict=True):  # type: ignore
-        assert pytest.approx(r, abs=3e-2) == e
+        assert pytest.approx(r, abs=5e-2) == e
 
 
 @pytest.mark.qualitative