diff --git a/test/registered/unit/server_args/test_resolution_is_reproducible.py b/test/registered/unit/server_args/test_resolution_is_reproducible.py
index 6f140bd942de..dbdd8db6bbe7 100644
--- a/test/registered/unit/server_args/test_resolution_is_reproducible.py
+++ b/test/registered/unit/server_args/test_resolution_is_reproducible.py
@@ -151,6 +151,26 @@
 if torch.cuda.is_available():
     _SHAPES = _SHAPES + (("deepseek_dsa", _DEEPSEEK_MINI_CONFIG, {}),)
 
+# Since #34662 made cuda_ipc opt-in, no auto-resolution reaches the handler's
+# cuda_ipc arm any more -- an explicit request is the only way in, so without
+# this shape that arm (two raises and the pool-budget logging) is resolved by
+# nothing in this file. It is also the only shape whose env write differs from
+# what the *next* resolution would pick on its own, which is what keeps the
+# sticky-carry assertion in `test_a_resolution_does_not_leak_into_the_next`
+# from being vacuous. Gated on `is_cuda()` rather than
+# `torch.cuda.is_available()`: the handler raises for cuda_ipc off NVIDIA CUDA,
+# ROCm included.
+_CUDA_IPC_SHAPES = ()
+if is_cuda():
+    _CUDA_IPC_SHAPES = (
+        (
+            "multimodal_cuda_ipc",
+            _MULTIMODAL_MINI_CONFIG,
+            dict(mm_feature_transport="cuda_ipc"),
+        ),
+    )
+    _SHAPES = _SHAPES + _CUDA_IPC_SHAPES
+
 # The one field a previous resolution genuinely dictates for the next one in
 # this process: `_handle_multimodal_feature_transport` writes
 # SGLANG_USE_CUDA_IPC_TRANSPORT so tokenizer workers inherit the decision, and
@@ -298,6 +318,11 @@ def test_a_resolution_does_not_leak_into_the_next(self):
             ("multimodal", _MULTIMODAL_MINI_CONFIG, {}),
             ("torch_compile", _MINI_CONFIG, dict(enable_torch_compile=True)),
         )
+        # Auto-resolution picks cpu on every runner this case runs on now that
+        # cuda_ipc is opt-in, so the auto shape above writes what the next
+        # resolution would have picked anyway; the explicit shape is what makes
+        # the carry observable at all.
+        intermediates += _CUDA_IPC_SHAPES
         if torch.cuda.is_available():
             # Same device gate as _SHAPES: the DSA arm probes the device
             # capability during resolution.
@@ -321,7 +346,7 @@ def test_a_resolution_does_not_leak_into_the_next(self):
                 # against `default_before`, so clearing it here does not skew
                 # that comparison.
                 envs.SGLANG_USE_CUDA_IPC_TRANSPORT.clear()
-                self._resolved(self._config_dir(config), **kwargs)
+                intermediate = self._resolved(self._config_dir(config), **kwargs)
                 after = self._resolved(model_path)
                 without_sticky = lambda snapshot: {
                     k: v
@@ -332,15 +357,21 @@ def test_a_resolution_does_not_leak_into_the_next(self):
                     without_sticky(self._comparable(after)),
                     without_sticky(self._comparable(default_before)),
                 )
-                if label == "multimodal":
-                    # And the documented exception, asserted rather than
-                    # assumed: the multimodal handler's env write does reach
-                    # the next resolution. What it carries is the
-                    # intermediate's own device-dependent selection — cuda_ipc
-                    # on single-node CUDA, cpu on the CPU/ROCm runners (the
-                    # same `is_cuda()` gate the handler branches on).
-                    expected = "cuda_ipc" if is_cuda() else "cpu"
-                    self.assertEqual(after.mm_feature_transport, expected)
+                # And the documented exception, asserted rather than assumed,
+                # for every intermediate: each one runs the transport handler,
+                # so each one writes the variable the next resolution reads.
+                # What carries is the legacy *boolean*, not the tri-state field
+                # -- the handler writes 1 only for cuda_ipc -- so every other
+                # selection comes back as cpu, which is what keeps this honest
+                # if a cuda_vmm shape is ever added (its carry is cpu, not
+                # cuda_vmm). The `cuda_ipc` shape is the one whose carry
+                # differs from the cpu that `default_before` resolved to.
+                expected = (
+                    "cuda_ipc"
+                    if intermediate.mm_feature_transport == "cuda_ipc"
+                    else "cpu"
+                )
+                self.assertEqual(after.mm_feature_transport, expected)
 
     def test_resolving_a_sibling_leaves_the_first_alone(self):
         for label, config, kwargs in _SHAPES: