negative-result informativeness scorer classifies its own test fixture below the INFORMATIVE threshold #843

Summary

Three tests in tests/evidence/test_negative_result_informativeness.py::TestCLI fail on main:

• test_cli_entry_flag
• test_cli_out_json_creates_file
• test_cli_out_md_creates_file

Each invokes scripts/classify_negative_result_informativeness.py on the _make_entry() fixture and expects an INFORMATIVE classification (exit 0). The scorer classifies the fixture below the informative band, so the script exits non-zero and the assertions fail.

Hidden until now because CI's pytest step aborted at collection (fixed in #841). These fail on unmodified main.

Root cause

scripts/classify_negative_result_informativeness.py scores each entry 0.0–7.0 and bands it:

INFORMATIVE:     6.0–7.0
PARTIAL:         3.5–5.5
NON_INFORMATIVE: 0.0–3.0
INFORMATIVE_SCORE_MIN = 6.0

The _make_entry() fixture used by TestCLI does not reach 6.0 under the current dimension scorers (_identity_score, _context_score, _specificity_score, etc.), so it is not classified INFORMATIVE. Either:

• (a) the scorer under-credits a legitimately informative entry (scoring bug), or
• (b) the fixture is not actually informative enough and the test expectation is wrong (test bug).

Why this needs review (Class C — do not blind-fix)

This is scoring/classification behavior. Per AGENTS.md/CLAUDE.md stop conditions, changes to scoring thresholds, rubric weights, or calibration require human review — they must not be adjusted to make a test pass. Resolving this correctly means deciding whether the rubric or the fixture is wrong, with a decision record if the rubric changes.

Repro

pip install -e ".[dev]"
python -m pytest "tests/evidence/test_negative_result_informativeness.py::TestCLI"

Related to the negative-result informativeness feature (Phase L / Loop area).