fix: remove python shim for already removed codec files

Signed-off-by: anubhutiv <anubhutiv@nvidia.com>

Author: anubhutivyas

docker/rl/codec-file-removals.txt

@@ -6,3 +6,6 @@
 /opt/nemo_rl_venv/lib/python3.*/site-packages/_soundfile_data/libsndfile_*.so
 /opt/ray_venvs/*/lib/python3.*/site-packages/_soundfile_data/libsndfile_*.so
 /opt/uv_cache/archive-v0/*/_soundfile_data/libsndfile_*.so
+/opt/nemo_rl_venv/lib/python3.*/site-packages/soundfile.py
+/opt/ray_venvs/*/lib/python3.*/site-packages/soundfile.py
+/opt/uv_cache/archive-v0/*/soundfile.py

tests/smoke_gpu/test_rl_training.py

@@ -36,6 +36,11 @@
     "/opt/nemo_rl_venv/lib/python3.*/site-packages/_soundfile_data/libsndfile_*.so",
     "/opt/ray_venvs/*/lib/python3.*/site-packages/_soundfile_data/libsndfile_*.so",
     "/opt/uv_cache/archive-v0/*/_soundfile_data/libsndfile_*.so",
+    # The shim is removed with the codec; see docker/rl/codec-file-removals.txt for why
+    # keeping it is worse than not shipping soundfile at all.
+    "/opt/nemo_rl_venv/lib/python3.*/site-packages/soundfile.py",
+    "/opt/ray_venvs/*/lib/python3.*/site-packages/soundfile.py",
+    "/opt/uv_cache/archive-v0/*/soundfile.py",
 )
 
 # Actor FQN -> packages that must import inside that actor's venv.
@@ -230,6 +235,27 @@ def test_soundfile_libsndfile_removed():
     assert remaining == [], f"codec cleanup left scanner-visible libsndfile files: {remaining}"
 
 
+@pytest.mark.smoke_nmp_rl_training
+def test_transformers_audio_backend_probe_is_off():
+    """Removing the codec must also switch off the probe that guards its import.
+
+    ``transformers.audio_utils`` does ``if is_soundfile_available(): import soundfile``,
+    and that probe is ``find_spec("soundfile")`` -- file presence, not loadability. Delete
+    the codec but keep the module and the probe says yes to a backend that then fails to
+    dlopen, so ``from transformers import AutoProcessor`` raises. That import is at module
+    scope in ``nemo_rl.algorithms.grpo``, so the GRPO driver dies before it reads a config.
+    """
+    from transformers.utils.import_utils import is_soundfile_available
+
+    assert not is_soundfile_available()
+
+
+@pytest.mark.smoke_nmp_rl_training
+def test_grpo_driver_module_imports():
+    """The exact import the GRPO driver performs first, and the one the codec strip broke."""
+    from nemo_rl.algorithms.grpo import MasterConfig  # noqa: F401
+
+
 @pytest.mark.smoke_nmp_rl_training
 def test_no_git_directories_shipped():
     """The published image must not carry repository history.