[Birfy]

Closes #1089

build_trust_evidence printed one "inside the noise floor" sentence for three different refusal causes. This branches the note by cause and stops gain_over_noise from dividing by 1e-9.

• Too few repeats (repeats < 2): the note says the floor cannot be estimated from N baseline repeat(s) instead of printing a one-point "floor".
• Gain inside a measurable floor: the note prints baseline mean +0.500 ± 0.200, not an absolute interval placed next to a delta.
• Floor narrower than min_effect: the note names min_effect and the floor half-width.
• Every branch keeps the pinned phrase not distinguishable from noise.
• gain_over_noise is 0.0 when the floor is undefined (fewer than 2 repeats or zero spread). The docstring says 0.0 there means undefined. distinguishable is unchanged.
• One changelog fragment. No new public symbol.

Tests: tests/test_noise_calibration.py (parametrised undefined-floor cases + measurable-ratio case; the zero-spread > 1.0 assertion becomes == 0.0), tests/test_trust_evidence_noise.py (one-repeat, below-min_effect, ± band, passing zero-spread).

Local: 167 passed / 13 skipped on every test file touching noise/gain_over_noise; tools.changelog_fragments check and tools.gen_api_docs --check OK. tests/test_plugin_host_cli.py::test_present_host_passes fails on the box on plain main as well (environment), so it's unrelated.

[Birfy]

Verdict: PASS

Tip SHA: 6e63c7b385ed2180a4fae9593114573b7ccd5f22
Ran: tests touching noise / gain_over_noise (167 passed, 13 skipped), python -m tools.changelog_fragments check, python -m tools.gen_api_docs --check, risk_gate (medium).
Not run: full pytest -q to completion on the box (stopped at test_plugin_host_cli.py::test_present_host_passes, which also fails on plain main locally; CI covers the full suite), mkdocs build --strict (CI).

Review: distinguishable is unchanged. Each of the three refusal causes gets its own sentence, all keep not distinguishable from noise. gain_over_noise is 0.0 exactly when the floor is undefined.

[Birfy]

Verdict: PASS
Tip SHA: 6e63c7b
Ran: full pytest -q on the tip (fresh clone): all pass except 3 box-environment failures that fail identically on plain main (test_integrations.py::test_the_launcher_install_writes_ignores_a_shadowing_cwd, ::test_the_launcher_survives_a_bare_agentdescent_directory_too, test_plugin_host_cli.py::test_present_host_passes); targeted noise/trust-evidence/diagnose tests 53 passed; python -m tools.changelog_fragments check OK; python -m tools.gen_api_docs --check OK; risk_gate.py from main → T2.
Not run: mkdocs build --strict locally (CI docs / build passed on this tip).

Correctness

• gain_over_noise now returns 0.0 when len(baseline_runs) < 2 or half_width <= 0, instead of dividing by 1e-9. This is exactly the undefined set from bug(eval): noise refusal names the wrong cause and prints an absolute band against a delta #1089, and distinguishable / bar are untouched, so the verdicts stay the same.
• _noise_refusal_note branches in the same order the verdict logic uses: repeats < 2 first (that is the len(baseline_runs) >= 2 guard), then half_width < min_effect (where bar = max(half_width, min_effect) is the effect-size bar), then the measurable-floor case. noise.repeats is len(baseline_runs), and noise.noise_floor is the same (lo, hi) the half-width comes from, so the note can't drift from the calculation. One edge case: at exact equality half_width == min_effect the floor branch is used. The two bars are equal there, so either wording is accurate.
• Every branch keeps the pinned phrase not distinguishable from noise. The ± band around baseline_mean fixes the old mix of an absolute interval and a delta.

Design fit / scope: private helper, no new public symbol, docstring says 0.0 means undefined, one changelog fragment. This matches the issue scope.

Tests: these cover all four probe cases (A–D) from #1089, plus the one existing assertion that changes (> 1.0 → == 0.0), with a comment explaining it. Good.

Nit (non-blocking): 0.0 now means both "no positive gain" and "undefined ratio". The docstring documents this. A later change could expose None/NaN instead if consumers need to tell the two apart, but that would be an interface change and is out of scope here.