[DO NOT MERGE] Confirm py3.15 segfault is a CPython alpha ABI skew

faulthandler showed the cp315 segfault is inside 'import torch'
(initModule -> CPython PyObject_SetAttr/_PyObject_MakeTpCall), i.e. an
ABI mismatch between the 3.15 alpha torch was built against and the
conda-forge python=3.15.0a8 runtime this job installs.

Replace the (unavailable in manylinux) gdb probe with a cross-interpreter
check: import the same libtorch_python.so under both the conda-forge
python and the manylinux python.org /opt/python/cp315-cp315 interpreter.
If it crashes under conda-forge but loads under manylinux, the conda-forge
alpha is confirmed as the culprit.

Revert before merging.

Author: atalman

.github/workflows/build_wheels_linux.yml

@@ -278,36 +278,43 @@ jobs:
 
           # --- DEBUG (Python 3.15 segfault in `setup.py clean`): remove before merge ---
           # On cp315 `python setup.py clean` dies with SIGSEGV (exit 139, core
-          # dumped). setup.py imports torch, so the crash is almost certainly in
-          # `import torch` from the cp315 nightly wheel. Capture where it happens
-          # with a Python faulthandler traceback and a native gdb backtrace.
-          # All debug commands are non-fatal (`|| true`) so they never mask the
-          # real failure below.
+          # dumped). faulthandler showed the crash is inside `import torch`
+          # (torch/__init__.py:445 -> initModule in libtorch_python.so), in
+          # CPython's own PyObject_SetAttr / _PyObject_MakeTpCall machinery --
+          # the signature of a C-API/ABI skew between the 3.15 *alpha* the torch
+          # wheel was compiled against and the runtime interpreter.
+          #
+          # The torch cp315 nightly wheel is built with the manylinux python.org
+          # interpreter (/opt/python/cp315-cp315), but this job runs under
+          # conda-forge python=3.15.0a8. Import the same torch .so under BOTH
+          # interpreters: if it crashes under conda-forge but loads under the
+          # manylinux python, the conda-forge alpha is the culprit (3.15's ABI is
+          # not frozen until beta). All debug commands are non-fatal (`|| true`)
+          # so they never mask the real failure below.
           echo "::group::DEBUG: interpreter / torch wheel info"
           ${CONDA_RUN} python -VV || true
           ${CONDA_RUN} python -c "import sysconfig; print('Py_GIL_DISABLED:', sysconfig.get_config_var('Py_GIL_DISABLED'))" || true
+          /opt/python/cp315-cp315/bin/python -VV || true
           ${CONDA_RUN} pip show torch || true
           echo "::endgroup::"
 
-          echo "::group::DEBUG: import torch under faulthandler"
+          echo "::group::DEBUG: import torch under conda-forge python (faulthandler)"
           # faulthandler dumps the Python stack at the moment of the fatal signal.
           ${CONDA_RUN} python -X faulthandler -c "import torch; print('torch', torch.__version__, 'imported OK')" \
-            || echo "DEBUG: 'import torch' crashed (see faulthandler traceback above)"
+            || echo "DEBUG: 'import torch' crashed under conda-forge python (see faulthandler traceback above)"
           echo "::endgroup::"
 
-          echo "::group::DEBUG: native backtrace via gdb"
-          # Allow core dumps and run the failing command under gdb for a C-level backtrace.
-          ulimit -c unlimited || true
-          command -v gdb >/dev/null 2>&1 || yum install -y gdb >/dev/null 2>&1 || true
-          if command -v gdb >/dev/null 2>&1; then
-            ${CONDA_RUN} gdb -q -batch \
-              -ex "set pagination off" \
-              -ex "run" \
-              -ex "bt full" \
-              -ex "thread apply all bt" \
-              --args python -X faulthandler setup.py clean || true
+          echo "::group::DEBUG: import same torch .so under manylinux python.org cp315"
+          # Point the manylinux interpreter at the conda env's site-packages so it
+          # loads the identical libtorch_python.so, isolating interpreter ABI from
+          # the torch build itself.
+          TORCH_SITE_PACKAGES="$(${CONDA_RUN} python -c 'import site; print(site.getsitepackages()[0])' 2>/dev/null || true)"
+          if [ -x /opt/python/cp315-cp315/bin/python ] && [ -n "${TORCH_SITE_PACKAGES}" ]; then
+            PYTHONPATH="${TORCH_SITE_PACKAGES}" /opt/python/cp315-cp315/bin/python -X faulthandler \
+              -c "import torch; print('torch', torch.__version__, 'imported OK under manylinux python.org cp315')" \
+              || echo "DEBUG: import also crashed under manylinux python.org cp315 -> NOT a conda-forge-only ABI skew"
           else
-            echo "DEBUG: gdb unavailable; skipping native backtrace"
+            echo "DEBUG: manylinux cp315 interpreter or torch site-packages not found; skipping cross-interpreter probe"
           fi
           echo "::endgroup::"
           # --- END DEBUG ---