diff --git a/CHANGELOG.md b/CHANGELOG.md
index 46360c36..82ba3f98 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,17 @@ trajectory so breaking changes can land in any minor release until the 1.0 stabi
 
 ### Changed
 
+- **Internal restructuring** (no behaviour change): the play-mode
+  shortcut-retargeting key list moved off the private
+  `cli/__init__.py:_PLAY_RETARGETED_KEYS` constant onto
+  `SimulationCfg.play_retargeted_keys()` (a static method on the domain
+  config in `genelab.configs`). The CLI's `env.` → `play_env.` retarget
+  loop now calls the method; the set of play-retargetable simulation
+  override paths (`env.simulation.{vis,gpu,steps,dt}`) lives next to the
+  `SimulationCfg` fields the `--vis` / `--gpu` / `--steps` / `--dt`
+  shortcuts target. `genelab play --help` is unchanged (R0.1 snapshot
+  gate green); `configs.py` stays torch-free at import (invariant #5).
+  Lands as ROADMAP §9 PR R3.2 — completes ADR-0005 (R3).
 - **Internal restructuring** (no behaviour change): `--eval-*` runner-arg
   parsing for in-training eval moved from `cli/__init__.py:_build_eval_callback`
   onto the domain config as `EvalCallbackCfg.from_args(runner_args) ->
diff --git a/src/genelab/cli/__init__.py b/src/genelab/cli/__init__.py
index 8bb93365..9990aa17 100644
--- a/src/genelab/cli/__init__.py
+++ b/src/genelab/cli/__init__.py
@@ -38,7 +38,7 @@
     render_registry,
 )
 from genelab.cli._scaffold import create_project_skeleton
-from genelab.configs import apply_overrides
+from genelab.configs import SimulationCfg, apply_overrides
 from genelab.registry import (
     TASKS,
     load_bundled_asset_zoo,
@@ -80,14 +80,6 @@ class _RegistryKindArg(str, Enum):
 
 _AGENT_KINDS: Final[frozenset[str]] = frozenset({"zero", "random", "trained"})
 
-_PLAY_RETARGETED_KEYS: Final[tuple[str, ...]] = (
-    "env.simulation.vis",
-    "env.simulation.gpu",
-    "env.simulation.steps",
-    "env.simulation.dt",
-)
-
-
 _RUN_FLAGS_HELP: Final[str] = """\
 Shorthand flags rewritten into env overrides:
 
@@ -556,7 +548,7 @@ def _configured_task(
     # task's play_env when one is configured. Keeps `genelab play TASK --vis` working
     # without forcing users to spell `play_env.simulation.vis`.
     if command == "play" and getattr(task.cfg, "play_env", None) is not None:
-        for short_key in _PLAY_RETARGETED_KEYS:
+        for short_key in SimulationCfg.play_retargeted_keys():
             if short_key in overrides:
                 overrides[short_key.replace("env.", "play_env.", 1)] = overrides.pop(short_key)
 
diff --git a/src/genelab/configs.py b/src/genelab/configs.py
index 09522684..5f876243 100644
--- a/src/genelab/configs.py
+++ b/src/genelab/configs.py
@@ -37,6 +37,25 @@ class SimulationCfg:
     # raise ``decimation`` if that's not what you want.
     render_fps: int | None = 60
 
+    @staticmethod
+    def play_retargeted_keys() -> tuple[str, ...]:
+        """Override paths the CLI rewrites ``env.`` → ``play_env.`` in play mode.
+
+        The ``--vis`` / ``--gpu`` / ``--steps`` / ``--dt`` shorthand flags expand
+        to ``env.simulation.<field>`` overrides. When a task defines a separate
+        ``play_env``, ``genelab play TASK --vis`` should target *that* env, so the
+        CLI retargets these keys onto ``play_env.simulation.<field>``. Owning the
+        list here (rather than as a private constant in ``cli/__init__.py``) keeps
+        the set of play-retargetable simulation overrides next to the fields
+        themselves (ADR-0005 / R3.2).
+        """
+        return (
+            "env.simulation.vis",
+            "env.simulation.gpu",
+            "env.simulation.steps",
+            "env.simulation.dt",
+        )
+
 
 @dataclass
 class InteractiveSceneCfg:
diff --git a/tests/test_configs.py b/tests/test_configs.py
new file mode 100644
index 00000000..800eb4d0
--- /dev/null
+++ b/tests/test_configs.py
@@ -0,0 +1,36 @@
+"""Tests for ``genelab.configs`` domain-config helpers.
+
+Currently covers ``SimulationCfg.play_retargeted_keys`` (ROADMAP §9 R3.2 /
+ADR-0005) — the set of override paths the CLI rewrites ``env.`` →
+``play_env.`` in play mode. The list moved off a private constant in
+``cli/__init__.py`` onto the domain config so it lives next to the
+``SimulationCfg`` fields the shortcut flags target.
+"""
+
+from __future__ import annotations
+
+from genelab.configs import SimulationCfg
+
+
+def test_play_retargeted_keys_exact_set() -> None:
+    """The four ``env.simulation.*`` shortcut-override paths, verbatim and ordered."""
+    assert SimulationCfg.play_retargeted_keys() == (
+        "env.simulation.vis",
+        "env.simulation.gpu",
+        "env.simulation.steps",
+        "env.simulation.dt",
+    )
+
+
+def test_play_retargeted_keys_target_real_simulation_fields() -> None:
+    """Every retargeted key names an actual ``SimulationCfg`` field (no stale paths)."""
+    fields = SimulationCfg.__dataclass_fields__
+    for key in SimulationCfg.play_retargeted_keys():
+        assert key.startswith("env.simulation."), key
+        field_name = key.rsplit(".", 1)[1]
+        assert field_name in fields, f"{field_name!r} is not a SimulationCfg field"
+
+
+def test_play_retargeted_keys_callable_on_class_and_instance() -> None:
+    """Static method — same result whether called on the class or an instance."""
+    assert SimulationCfg.play_retargeted_keys() == SimulationCfg().play_retargeted_keys()