feat(batch-a): port four OPEN Netflix upstream PRs (ADR-0131/0132/0134/0135)

[lusoris]

Summary

Batch-A of the upstream small-fix sweep: ports substance from four open Netflix/vmaf PRs — each corrected during port where upstream had defects. All four upstream PRs are still unmerged as of 2026-04-20 (oldest is 21 months old, none has a merge signal), so the fork carries the fixes explicitly with ADRs documenting deviation.

• T0-1 (CUDA) — [#1381] Change cuMemFreeAsync to cuMemFree in vmaf_cuda_picture_free() Netflix/vmaf#1382 cuMemFreeAsync → cuMemFree in vmaf_cuda_picture_free. Fixes the Assertion 0 failed crash reported in VMAF CUDA vmaf_cuda_picture_free() Assertion 0 failed Netflix/vmaf#1381 when two or more VMAF CUDA sessions free pictures concurrently. This unblocks running multiple VMAF instances in the same ffmpeg pipeline on nvidia, which has been failing with the multi-session data hazard until now. ADR-0131.
• T4-4 (feature_collector) — Bugfix for adding/removing models. Netflix/vmaf#1406 mount/unmount list-corruption with ≥3 mounted models. Fork adds a correct local-cursor traversal and splits -ENOENT from -EINVAL. Test body refactored into shared 3-model helpers so the upstream inline 60-LoC duplication doesn't trip clang-tidy readability-function-size (JPL-P10 rule 4). ADR-0132.
• T4-5 (meson) — build: declare dependency to use it as subproject Netflix/vmaf#1451 declare_dependency + override_dependency('libvmaf', ...) so the fork is consumable as a meson subproject via dependency('libvmaf'). ADR-0134.
• T4-6 (public API) — Expose builtin model versions Netflix/vmaf#1424 vmaf_model_version_next iterator. Ported with three correctness corrections upstream omits: (1) NULL-pointer arithmetic UB via missing else, (2) off-by-one returning the {0} sentinel via idx < CNT, (3) const-qualifier mismatches in the test blocking -Werror. ADR-0135 documents each defect with C11 §§ references.

Six commits on the branch — four ports + one docs bundle + one clang-format fixup. All 30/30 meson tests pass locally; assertion-density + copyright headers green.

Type

• fix — bug fix (T0-1, T4-4)
• feat — new feature (T4-6 iterator)
• build / ci — tooling / infra (T4-5)
• port — cherry-pick from upstream Netflix/vmaf (all four)

Checklist

• Commits follow Conventional Commits.
• make format && make lint equivalent (pre-commit + scripts/ci/*) is green locally.
• Unit tests pass: meson test -C build → 30/30 OK.
• No SIMD/GPU numerical-path changes — T0-1 is a driver-API swap, not a kernel.
• No new .c/.cpp/.cu files; license headers on touched upstream files preserved per ADR-0025.
• Not a breaking change (iterator is purely additive).

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• No golden-value change claimed or needed.

Cross-backend numerical results

Not applicable — no kernel or scoring-path changes.

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/0009-batch-a-upstream-port-strategy.md (why port OPEN PRs now, per-PR defect analysis, port-from-PR-tip rationale, batch-shape justification).
• Decision matrix — each of ADR-0131 / 0132 / 0134 / 0135 carries its own ## Alternatives considered table with the wait-for-upstream / port-verbatim / port-with-corrections options weighed.
• AGENTS.md invariant note — added new Rebase-sensitive invariants entry to libvmaf/src/cuda/AGENTS.md (picture_cuda.c sync-free contract) and a bullet under the existing section in libvmaf/src/feature/AGENTS.md (feature_collector mount/unmount traversal + test helpers).
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG.md "lusoris fork" entry — two Added bullets (T4-5, T4-6) and two Fixed bullets (T0-1, T4-4) under the fork's Unreleased section.
• Rebase note — docs/rebase-notes.md §0031 — full Touches / Invariant / Re-test with per-file conflict predictions and "keep fork version" resolution policy for when upstream merges any of the four PRs.

Reproducer

# Build without GPU to exercise T4-4, T4-5, T4-6 on any machine.
meson setup build libvmaf -Denable_cuda=false -Denable_sycl=false
ninja -C build
meson test -C build  # expect 30/30 OK

# T4-6 sanity: iterate the built-in model versions via the new API.
cat > /tmp/iter.c <<'C'
#include <stdio.h>
#include <libvmaf/model.h>
int main(void) {
    const void *h = NULL; const char *v = NULL; unsigned n = 0;
    while ((h = vmaf_model_version_next(h, &v)) != NULL) { printf("%s\n", v); n++; }
    printf("total: %u\n", n);
    return 0;
}
C
cc /tmp/iter.c -Ibuild/include -Lbuild/src -lvmaf -o /tmp/iter
LD_LIBRARY_PATH=build/src /tmp/iter

# T0-1 sanity (requires CUDA): concurrent-session tear-down no longer asserts.
# Reproduce Netflix#1381 scenario — two VMAF sessions on one GPU.
meson setup build-cuda libvmaf -Denable_cuda=true -Denable_sycl=false
ninja -C build-cuda
build-cuda/test/test_pic_preallocation  # exercises picture lifecycle

Known follow-ups

• Ensure the parsing works properly with different locales Netflix/vmaf#1430 (thread-local RNG port, T4-3) is ~600 LoC and was held out of this batch; tracked as T4-3-standalone in the backlog audit.
• When Netflix merges any of [#1381] Change cuMemFreeAsync to cuMemFree in vmaf_cuda_picture_free() Netflix/vmaf#1382 / Bugfix for adding/removing models. Netflix/vmaf#1406 / Expose builtin model versions Netflix/vmaf#1424 / build: declare dependency to use it as subproject Netflix/vmaf#1451, /sync-upstream will see conflicts in the four files carrying fork corrections — resolution is always "keep fork version" (rationale in rebase-notes §0031).

🤖 Generated with Claude Code