test(retrieval-gate): row 5 must read the threshold as a number and compare it with the hold-out

Red first, on behaviour: both assertions are about the checks the row
prints, neither is an import or an attribute error.

Measured by PM on 23588e5: `bool(threshold)` was the whole check, so the
threshold `report-only; any number is acceptable for v1` read as `a
threshold is written: yes`. A threshold that is never compared with a
number cannot fell anything.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_retrieval_gate.py

@@ -1200,3 +1200,85 @@ def test_a_wiki_set_of_one_question_is_refused_on_the_command_line(
     with pytest.raises(gate.GateUsage) as refusal:
         gate._real_sets([("wiki", str(path), gate.sha256_of(path), str(bundle))])
     assert "wiki-20" in str(refusal.value)
+
+
+# --- step 0: the threshold is a number, and it is compared with something ----
+
+
+def _registration(
+    tmp_path: Path,
+    *,
+    set_path: Path,
+    bundle: Path,
+    threshold: object,
+    name: str = "registration.json",
+) -> Path:
+    registration = tmp_path / name
+    registration.write_text(
+        json.dumps(
+            {
+                "set": str(set_path),
+                "sha256": gate.sha256_of(set_path),
+                "bundle": str(bundle),
+                "metric": "questions answered over questions asked",
+                "threshold": threshold,
+                "threshold_written_at": "2026-09-19T10:00:00Z",
+                "written_by": "the session that wrote the eval",
+                "readings": [{"at": "2026-09-20T09:00:00Z", "value": "not read yet"}],
+            }
+        ),
+        encoding="utf-8",
+    )
+    return registration
+
+
+def test_a_threshold_that_is_not_a_number_is_refused(tmp_path: Path) -> None:
+    """`bool(threshold)` was the whole check, so `report-only; any number is
+    acceptable for v1` read as `a threshold is written: yes`. A threshold that
+    cannot be compared with a number cannot fell anything."""
+    bundles = _bundles(tmp_path)
+    registration = _registration(
+        tmp_path,
+        set_path=FIXTURES / "set-positive.json",
+        bundle=bundles["positive"],
+        threshold="report-only; any number is acceptable for v1",
+    )
+    row = gate.row_five(registration, provenance=lambda _: _carried_by_git())
+    assert any("the threshold is a number: NO" in detail for detail in row.details)
+    assert row.fails
+
+
+def test_the_threshold_is_compared_with_the_measured_hold_out(tmp_path: Path) -> None:
+    """Both directions, from the same code path: a set the bundle answers
+    clears a threshold under it, and a set it does not answer falls under one
+    over it. The measured share is counted here as well, off the set's own
+    questions, so the row is not the only thing that knows it."""
+    bundles = _bundles(tmp_path)
+    clears = gate.row_five(
+        _registration(
+            tmp_path,
+            set_path=FIXTURES / "set-positive.json",
+            bundle=bundles["positive"],
+            threshold=0.8,
+            name="clears.json",
+        ),
+        provenance=lambda _: _carried_by_git(),
+    )
+    assert any("clears the threshold: yes" in detail for detail in clears.details)
+    assert (clears.k, clears.m, clears.status) == (11, 11, gate.GREEN)
+
+    # The independent count: set-miss carries one question, forced to miss.
+    missing = json.loads((FIXTURES / "set-miss.json").read_text(encoding="utf-8"))
+    assert len(missing["questions"]) == 1
+    falls = gate.row_five(
+        _registration(
+            tmp_path,
+            set_path=FIXTURES / "set-miss.json",
+            bundle=bundles["miss"],
+            threshold=0.5,
+            name="falls.json",
+        ),
+        provenance=lambda _: _carried_by_git(),
+    )
+    assert any("clears the threshold: NO" in detail for detail in falls.details)
+    assert (falls.k, falls.m, falls.status) == (10, 11, gate.RED)