[SmaCow666]

OpenBLAS supports negative NRM2 increments, and the public wrapper starts at the last logical element for such calls. Both RVV complex NRM2 kernels store the byte stride in unsigned int. A negative stride is consequently passed to a strided vector load as a large positive offset, so a two-element public DZNRM2/SCNRM2 call can fault.

Use BLASLONG for the byte stride in znrm2_vector.c and znrm2_rvv.c. Add C/Z Fortran and CBLAS regressions for negative unit, negative non-unit and positive non-unit increments. Keep the tests usable with NO_CBLAS and check result finiteness before approximate comparisons.

Local/QEMU validation used upstream 02f585444a0fba2505cfda77cc6684de008ea477, Clang 21.1.8, LP64D/32-bit BLAS integers, S/D/C/Z BLAS, pthreads with a two-thread limit, and QEMU 9.1.0:

• Complete builds from the pinned upstream tree for RISCV64_ZVL128B and RISCV64_ZVL256B, with each PR applied independently.
• Both targets: 153/153 unit tests and 1479/1479 extension tests pass, together with all enabled official Fortran BLAS and CBLAS Level-1/2/3 drivers in all four precisions.
• VLEN=512: ZVL128B passes 153/153 unit and 1479/1479 extension tests. ZVL256B has three unit failures and 34 extension failures, exactly matching the named failures in the unmodified library at the same VLEN and thread count; all twelve added regressions and the targeted API checks pass. This configuration is not an all-passing full-library result. Primary runs use the target's minimum VLEN and request two threads; VLEN512 runs request one.
• Four current full-library/target/VLEN cells each pass 2,484 guarded finite-value API cases, including 960 negative-increment cases. Input bytes remain unchanged.
• Each unmodified full-library control reproduces four small-input SIGSEGV cases in isolated children; patched calls return the expected norm.
• Ten warning-as-error precision/NO_CBLAS compilations passed. The NO_CBLAS test objects have no CBLAS references and their nine-test runners pass for both targets.
• Library-member comparison confines changes to the affected kernels; exported symbol sets are unchanged. The patch passes git apply --check --whitespace=error-all against the pinned upstream tree.

BLAS was built with NO_LAPACK=1. A modern GCC RVV build and legacy C910V intrinsics have not been tested. No performance improvement is claimed.

Native RISC-V Linux validation (VLEN=256):

• The same independently patched binaries pass all four target/thread cells: RISCV64_ZVL128B and RISCV64_ZVL256B, each with one and two requested threads. Each cell passes 153/153 unit tests, 1479/1479 extension tests, and all 24 enabled official S/D/C/Z Fortran BLAS and CBLAS Level-1/2/3 drivers.
• This patch accounts for 120 native program executions; the shared unmodified-library baseline accounts for another 108. Both independently tested patches together account for 348 executions, rather than 348 per patch.
• 9,936 guarded finite-value API calls pass, including 3,840 negative-increment calls, with unchanged input bytes. Unmodified-library controls reproduce four expected child SIGSEGV cases in each cell (16 in total). Each NO_CBLAS runner passes 9/9 tests.

Native and QEMU results are separate correctness evidence; no hardware performance measurement is claimed.