bug(eval): a regression is explained as a noise-floor problem — the note has no branch for gain <= 0 #1091

Problem

_noise_refusal_note (trust_evidence.py:335-354) explains a refusal with three causes and says so in its docstring (trust_evidence.py:338), but distinguishable (stats_eval.py:368-369) has four ways to be false: fewer than 2 baseline repeats, gain <= 0.0, and gain <= bar + 1e-12. gain <= 0.0 has no branch of its own, so it falls through to whichever other cause happens to be true and a regression is explained as noise:

• 3 noisy repeats, gain −0.300 → gain -0.300 is not distinguishable from noise (gain does not exceed the noise floor: baseline mean +0.500 ± 0.200) — the gain is negative, so no floor could make it distinguishable;
• 3 zero-spread repeats, gain −0.050 → the same sentence with ± 0.000, an absolute band printed against a loss;
• 1 repeat, gain −0.300 → blames the repeat count, although gain > 0.0 fails at any count.

The docstring/code drift is the bug: the promise lists three causes, the predicate has four, and the one that is not printed is the one that describes a regression.

Evidence

Probe run-1115-0104/probe-negative-gain-note.py (PYTHONPATH=<clone> python3 probe-negative-gain-note.py), live main c94e9d67:

calibrate_noise on E reports distinguishable=False, gain=-0.300, gain_over_noise=0.0 — only the sentence is wrong.

Scope

• trust_evidence.py:343: branch first on noise.gain <= 0.0 and return (the candidate did not improve on the baseline: the gain must be positive). The head already prints the gain (gain -0.300 is not distinguishable from noise), so the branch must not print it a second time.
• trust_evidence.py:338: the docstring lists the four causes in code order — gain ≤ 0 → fewer than 2 repeats → min_effect → floor.
• Every branch keeps the phrase not distinguishable from noise verbatim; three test files assert it.
• Do not touch distinguishable, the verdict strings or gain_over_noise (that was fix(eval): name the real cause of a noise refusal and stop faking gain_over_noise (#1089) #1090).

Tests to add

• tests/test_trust_evidence_noise.py: a negative gain (E) prints did not improve on the baseline and no ± band; an exact gain == 0.0 row (NoiseCalibration(gain=0.0, ...)) takes the same branch; a gain=+8.327e-16 row stays in the floor branch (the guard is <= 0.0, not a tolerance, so float noise is not swallowed).
• One changelog fragment.

Existing expectations

None change. The one-repeat test (tests/test_trust_evidence_noise.py:62) uses gain +0.3, so the repeats branch still fires there; the three noise test files give 41 passed before and after with this patch (re-run at c94e9d67).

Acceptance: E, G and H flip to the new sentence, I and the passing control unchanged; pytest -q green offline; every branch still contains the pinned phrase; mkdocs build --strict, python -m tools.gen_api_docs --check and python -m tools.changelog_fragments check pass; one changelog fragment.

Bot: AgentDescent Product