test(e2e): broaden race-rejection body match for double-stop

The double-stop case asserts the response body explains the
race-protection rejection so a 'box stopped on accident, no
explanation' regression can still trip it. Tokyo's Api now returns
HTTP 400 with 'Box is not started' (one of several stable wordings
the API has picked for the same invariant: current state doesn't
admit this transition). Accept 'not started' / 'already stopped' /
'invalid state' alongside the existing 'state change' phrasing —
all four are semantically equivalent for the race-protection check
the test is gating.

Author: G4614

scripts/test/e2e/cases/test_error_code_mapping.py

@@ -275,9 +275,19 @@ async def test_invalid_state_stop_already_stopped_returns_4xx(rt, image):
             f"double-stop leaked HTTP {status2} (5xx); body={body_str}"
         )
         if status2 not in (200, 204, 409):
-            assert "state change" in body_str.lower(), (
-                f"double-stop got HTTP {status2} but body doesn't explain the "
-                f"race-protection rejection: {body_str}"
+            # 400 'Box is not started' / 'already stopped' / 'state change in
+            # progress' are all valid race-protection rejections — the API
+            # has picked different wording across deploys for the same
+            # invariant (current state doesn't admit this transition).
+            body_lower = body_str.lower()
+            assert (
+                "state change" in body_lower
+                or "not started" in body_lower
+                or "already stopped" in body_lower
+                or "invalid state" in body_lower
+            ), (
+                f"double-stop got HTTP {status2} but body doesn't explain "
+                f"the race-protection rejection: {body_str}"
             )
     finally:
         # Tolerate the race here too — the runner may still be in