driver_version() was called unconditionally during Linker.__init__,
which fails in environments where nvJitLink is installed but the
CUDA driver is absent (e.g., build containers). Now catches the
exception and sets driver_major=None. When driver_major is unknown
and nvJitLink is available, optimistically selects the nvJitLink
backend.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Test helpers calling driver_version() at module scope would crash
in no-driver environments before test collection. Mirror the
production lazy-probe pattern: catch exceptions and pass None.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Linker_link was nulling self._drv_log_bufs right after cuLinkComplete,
releasing the bytearrays whose addresses were handed to the driver via
CU_JIT_INFO_LOG_BUFFER and CU_JIT_ERROR_LOG_BUFFER at cuLinkCreate time.
The CUlinkState retains those pointers until cuLinkDestroy, which runs
during Linker tp_dealloc. Freeing the bytearrays first left the driver
with dangling pointers and corrupted the heap; subsequent CUDA calls
(e.g. NVRTC compilation in the next test fixture) segfaulted.

This path became reachable in CI with the new per-instance backend
dispatch: CTK 12.9.1 + driver 13.0 runners now hit the driver linker
for cross-major linking, which was never exercised before.

Retain _drv_log_bufs until the cdef class is deallocated; pxd
declaration order ensures _culink_handle (and therefore cuLinkDestroy)
runs before the bytearrays are cleared.

…etime

The CUDA driver docs state: "optionValues must remain valid for the life
of the CUlinkState if output options are used." The driver writes log-
fill sizes (output) back into the optionValues slots for
CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES and CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES.
Linker_init previously declared c_jit_keys/c_jit_values as local
cdef vector[...] on the stack of Linker_init; they were destroyed when
the function returned, leaving the driver with dangling writes during
subsequent cuLinkAddData/cuLinkComplete/cuLinkDestroy calls.

This was always latent. It became reachable with the per-instance
backend dispatch (CTK 12.9.1 runners now select the driver linker when
they pair with a driver 13 install), and only manifested on driver 13
as heap corruption that killed the next NVRTC or link call.

Promote the two arrays to cdef class fields declared after
_culink_handle in the pxd. Cython's tp_dealloc destroys C++ fields in
pxd declaration order, so the vectors are destroyed after the shared_ptr
deleter runs cuLinkDestroy. The cuda.bindings high-level wrapper
(driver.cuLinkCreate) already handles this by attaching a keepalive to
CUlinkState; cuda.core's low-level cydriver.cuLinkCreate path did not.

Also drop the now-unused void_ptr ctypedef.

The as_bytes() method raises ValueError for unsupported backends (per
its docstring and matching the test directly above this one). The
driver-backend skip-guarded test was asserting RuntimeError, so it
always failed on CTK 12.9.1 runners where the skip condition does not
apply.

Adds parametrized cases for the build-container path where
cuDriverGetVersion is unqueryable: with nvJitLink present the
dispatcher picks nvjitlink optimistically; with nvJitLink absent
it falls back to driver for non-LTO and raises for LTO.

These paths are documented in _choose_backend's contract but were
previously uncovered.