test(consume, retrieval-gate): the payload must say what of the question it covered, and row 4 must see it

Red first, on behaviour: three asserts about what the payload carries and
what row 4 scores, none on an import or an attribute.

Measured 2026-09-19: N3, N4 and N5 come back with 8, 8 and 1 excerpts and
nothing that says they are weak, so `marked = nothing delivered` reads
three uncovered questions as answered ones. The second gate test is the
known-negative that a marking firing on everything would fail: the 10
questions the three synthetic sets DO answer must stay unmarked.

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

tests/test_retrieval_gate.py

@@ -1287,3 +1287,49 @@ def test_the_threshold_is_compared_with_the_measured_hold_out(tmp_path: Path) ->
     )
     assert any("clears the threshold: NO" in detail for detail in falls.details)
     assert (falls.k, falls.m, falls.status) == (10, 11, gate.RED)
+
+
+# --- step 1: row 4, the marking a consumer can act on -------------------------
+
+
+def test_row_four_marks_every_uncovered_control_and_leaves_the_covered_one(
+    tmp_path: Path,
+) -> None:
+    """Six controls, five uncovered and one covered, and the row must get all
+    six right.
+
+    Three of the five were measured red on 2026-09-19: N3, N4 and N5 came back
+    with 8, 8 and 1 excerpts and nothing in the payload that said they were
+    weak, so `marked = nothing delivered` could not see them.
+    """
+    case = _case(tmp_path, "set-controls.json")
+    # The denominator, counted from the pinned file rather than from the run.
+    declared = json.loads((FIXTURES / "set-controls.json").read_text(encoding="utf-8"))
+    assert len(declared["controls"]) == 6
+    assert sum(1 for control in declared["controls"] if control.get("covered")) == 1
+    row = gate.row_four([case])
+    assert (row.k, row.m, row.status) == (6, 6, gate.GREEN)
+    assert not row.details
+
+
+def test_the_marking_stays_off_every_question_the_synthetic_sets_do_answer(
+    tmp_path: Path,
+) -> None:
+    """The known-negative, and it is wider than row 4's own denominator.
+
+    A marking that fires on the questions the bundles DO answer would take row
+    4 to 6 of 6 and be worthless. Every question of the three sets whose fasit
+    row 1 finds is measured here, and none of them may come back marked.
+    """
+    bundles = _bundles(tmp_path)
+    answered = 0
+    for name in ("set-positive.json", "set-signals.json", "set-quota.json"):
+        question_set = gate.load_set(FIXTURES / name, gate.SYNTHETIC_SETS[name])
+        for question in question_set.questions:
+            bundle = bundles[question.bundle or question_set.bundle]
+            payload = consume.build_payload(
+                bundle, question=question.question, k=question.k, limit=question.limit
+            )
+            assert not gate.marked(payload), f"{question_set.set_id}/{question.id}"
+            answered += 1
+    assert answered == 10, answered