ci: check out test sources at the commit the wheels were built from

Nightly runs reuse wheels from the latest successful CI run on main, but
the test job checked out the workflow's own commit.  When main moved on
after that run, new tests ran against old wheels: the aarch64
cuda.bindings job saw a cascade of failures and then a segfault in
test_ffi_coverage_jit_wall_time_pointer_is_valid, where the old wheel
returns a NULL cptr that the test dereferences through ctypes.

Check out `inputs.sha || github.sha` in the Linux and Windows test
workflows so sources match the wheels under test, and assert cptr is
non-NULL in that test so a regression fails instead of crashing.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

Author: mdboom

.github/workflows/test-wheel-linux.yml

@@ -128,6 +128,10 @@ jobs:
     steps:
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          # Test sources must match the wheels under test (nightly runs
+          # reuse wheels built from an older commit of main).
+          ref: ${{ inputs.sha || github.sha }}
 
       - name: Setup proxy cache
         uses: nv-gha-runners/setup-proxy-cache@main

.github/workflows/test-wheel-windows.yml

@@ -107,6 +107,10 @@ jobs:
     steps:
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          # Test sources must match the wheels under test (nightly runs
+          # reuse wheels built from an older commit of main).
+          ref: ${{ inputs.sha || github.sha }}
 
       - name: Setup proxy cache
         uses: nv-gha-runners/setup-proxy-cache@main

cuda_bindings/tests/test_cuda.py

@@ -655,6 +655,8 @@ def test_ffi_coverage_jit_option_binary_loader_thread_count():
 def test_ffi_coverage_jit_wall_time_pointer_is_valid():
     """CU_JIT_WALL_TIME must give CUDA a valid writable float address, not a garbage pointer."""
     helper = cuda._HelperCUjit_option(cuda.CUjit_option.CU_JIT_WALL_TIME, 0.0)
+    # Guard against NULL so a regression fails instead of segfaulting.
+    assert helper.cptr != 0
     # Write a sentinel float via ctypes to confirm the pointer is addressable.
     ctypes.c_float.from_address(helper.cptr).value = 3.14
     assert ctypes.c_float.from_address(helper.cptr).value == pytest.approx(3.14, rel=1e-5)