feat(llm): fake supports per-call response sequences

Add VisionResponses to Fake. Vision returns the element matching the
call index (clamped to last beyond length), falling back to VisionResp
when empty. Enables self-consistency voting tests.

internal/llm/fake_test.go

@@ -19,3 +19,28 @@ func TestFakeClient(t *testing.T) {
 		t.Fatalf("transcribe got %q err %v", tr, err)
 	}
 }
+
+func TestFakeVisionResponsesCycle(t *testing.T) {
+	f := &Fake{VisionResponses: []string{"a", "b", "c"}}
+	want := []string{"a", "b", "c", "c", "c"} // clamps to last beyond slice length
+	for i, w := range want {
+		got, err := f.Vision(context.Background(), "p", nil, Options{})
+		if err != nil {
+			t.Fatalf("call %d err %v", i, err)
+		}
+		if got != w {
+			t.Fatalf("call %d got %q want %q", i, got, w)
+		}
+	}
+	if len(f.VisionCalls) != len(want) {
+		t.Fatalf("recorded %d calls, want %d", len(f.VisionCalls), len(want))
+	}
+}
+
+func TestFakeVisionFallsBackToVisionResp(t *testing.T) {
+	f := &Fake{VisionResp: "single"}
+	got, err := f.Vision(context.Background(), "p", nil, Options{})
+	if err != nil || got != "single" {
+		t.Fatalf("got %q err %v, want single", got, err)
+	}
+}