fix: test_hallucination_detection missing qualitative marker and tight tolerance

[planetf1]

Problem

test_hallucination_detection in test/stdlib/components/intrinsic/test_rag.py:152 has two issues:

1. Missing @pytest.mark.qualitative — every other LLM-output-quality test in the file is marked qualitative, but this one isn't. It runs in fast test loops where it shouldn't.

2. Tolerance too tight — asserts pytest.approx(r, abs=3e-2) on a generative model score. Observed drift of 0.036 causes spurious failures (reported in fix: evict Ollama models between test modules to prevent memory starvation #804).

Fix

• Add @pytest.mark.qualitative decorator
• Widen tolerance from abs=3e-2 to abs=5e-2

Related

• Bug: Flaky test in langchain_messages.py - qualitative assertion too strict #384 — same class of issue (qualitative assertion too strict)
• test: semantic assertions and recording for qualitative tests #735 / Proposal: Semantic assertions for qualitative test validation #692 — semantic assertion infrastructure
• Reported in fix: evict Ollama models between test modules to prevent memory starvation #804 (comment by @planetf1)

[planetf1]

Investigation

• @pytest.mark.qualitative was already present on test_hallucination_detection — no marker change needed
• Tolerance widened from abs=3e-2 → abs=5e-2 on both assertions (lines 161, 166)

Justification for 5e-2

• Score is a weighted expected value over token logprobs (not a raw LLM string or embedding distance)
• Even at temperature=0.0, logprobs shift slightly across runs (inference non-determinism: batching, quantization, kernel ordering)
• Observed drift of 0.036 on a score of ~0.728 is ~5% relative — categorical signal (hallucinated vs not) remains stable
• 5e-2 catches real regressions (e.g. score flipping from 0.7→0.2) while absorbing normal logprob jitter