fix(dev): stop preflight and the symbol check failing on out-of-scope input

[lusoris]

Summary

Three local checks were wrong. Two failed on input they should not judge and made make preflight report failures on a correct branch (#1474); the third never checked anything. None of them runs in CI.

• scripts/dev/preflight.sh, 32-bit sweep. It ran gcc -m32 -fsyntax-only over every changed C file, including the arm64/, x86/ SIMD and GPU trees. The Ubuntu i686 gcc lane it mirrors never compiles those: it configures -Denable_asm=false and no GPU backend. So a change to adm_neon.c or the AVX ADM files failed the stage on arm_neon.h and _mm_extract_epi64. The missing-header filter also knew only clang's file not found and the German gcc text, but the script runs under LC_ALL=C, so gcc's No such file or directory counted as a failure. The stage now sweeps only the changed files the gcc stage's CPU build compiles (its compile_commands.json), which also leaves out the per-backend GPU tests that only build with a backend define, skips the ISA trees, and accepts that message. The gcc stage also reused any existing build/ directory, even one holding only other build trees (build/cuda, ...), and then failed at ninja; it now requires build/build.ninja.

• core/test/check_exported_symbols.py. In a sanitizer build linked with GNU ld, libvmaf.so exports the linker-defined __start_asan_globals / __stop_asan_globals, the bounds of ASan's metadata section. lld, which the CI sanitizer lane uses, hides them. They are runtime artefacts, not API, so the checker now treats the __start_ / __stop_ bounds of the ASan, HWASan and SanitizerCoverage sections as runtime-owned. Any other section bound still fails. The file's three pre-existing ruff findings are cleared as well.

• scripts/dev/check-cuda-extern-c.sh (ADR-0747). It never checked a kernel. Its regex expected cuModuleGetFunction(&fn, "name"), but every call passes the module first, so it collected nothing and then aborted on the empty list under set -u. Rewritten: it reads every call, including split ones, blanks comments and strings before counting braces, and names the 23 macro-generated kernels it cannot locate instead of passing them. It reports 48 of 71 located and 0 unwrapped, and fails with the kernel names when an extern "C" block is removed.

The real i686+asm limitation behind the first point (_mm_extract_epi64 in adm_avx2.c / adm_avx512.c) is unchanged and stays as ADR-0151 records it.

Type

• feat — new feature
• fix — bug fix
• perf — performance improvement
• refactor — no behavior change
• docs — documentation only
• test — test-only
• build / ci — tooling / infra
• port — cherry-pick from upstream Netflix/vmaf
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally.
• Unit tests pass: meson test -C build.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2: no kernel code touched.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below: none touched.
• If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header (see CONTRIBUTING.md): no new files.
• If this is a breaking change, the commit message uses ! or BREAKING CHANGE: and the migration path is documented below: not breaking.
• If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md and the slug is appended to docs/adr/_index_fragments/_order.txt: no ADR.

Bug-status hygiene (ADR-0165)

• docs/state.md: T-PREFLIGHT-M32-FALSE-FAILURES-2026-09-18, T-EXPORTED-SYMBOLS-ASAN-GNU-LD-2026-09-18 and T-CUDA-EXTERN-C-CHECK-VACUOUS-2026-09-18, all closed.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• If I believe a golden value must change, I have explained why below AND pinged @lusoris for a CODEOWNERS exception: none.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial.
• Decision matrix — no alternatives: only-one-way fix.
• AGENTS.md invariant note — no rebase-sensitive invariants.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/preflight-m32-asan-symbols.md, changelog.d/fixed/cuda-extern-c-check.md.
• Rebase note — no rebase impact: fork-only tooling.

Reproducer

# on a branch that changes core/src/feature/arm64/adm_neon.c (e.g. #1474):
scripts/dev/preflight.sh --stage m32          # was FAIL on arm_neon.h, now PASS

# GNU-ld ASan build (CI links with lld, where the symbols stay hidden):
CC=clang meson setup build-asan core -Db_sanitize=address -Denable_cuda=false -Denable_sycl=false -Db_lto=false
ninja -C build-asan && meson test -C build-asan check_exported_symbols

Verified locally

Check	Result
preflight.sh --stage m32 on #1474's and the GPU branch's trees	FAIL before, PASS after; a planted 64-bit _Static_assert in a CPU file still fails it
check_exported_symbols.py on a GNU-ld ASan, a release and a SYCL libvmaf.so	all pass
The new rule on __start_mysection, __stop_vmaf_table, a leaked function name	all still reported
check-cuda-extern-c.sh on master / with one extern "C" block removed	exit 0, 48 of 71 located / exit 1 naming both kernels
shellcheck, ruff, black, pre-commit on the changed files	pass

Known follow-ups

• The 23 macro-generated CUDA kernels are still verified only from the built PTX, not by the script.