fix(test): run the GPU pool UAF test without MALLOC_PERTURB_ so it cannot fill the host's memory

[lusoris]

Summary

test_gpu_picture_pool_uaf (suite fast) can take all of a host's RAM and swap. It now runs with MALLOC_PERTURB_=0, and its two cases are unchanged.

The test asks vmaf_gpu_picture_pool_init() for 0x7FFFFFFF pictures, about 192 GB, to reach the allocation-failure cleanup. Meson sets a random MALLOC_PERTURB_ for every test, so glibc fills each allocation it returns. On a host with vm.overcommit_memory = 1 the request succeeds and the fill writes every byte of it. On the RC workstation (60 GB RAM + 60 GB swap) the merge train's stack gate timed out on it after 91 s (local gate failed on the stack of 5 (fast tests fail ...: test_gpu_picture_pool_uaf TIMEOUT 90.9s), 00:47 on 2026-10-08). A local gate run timed out after 32 s; both runs fill swap while they last.

Meson keeps an explicit MALLOC_PERTURB_=0 and does not set its own value (mesonbuild/mtest.py: "Setting MALLOC_PERTURB_="0" will completely disable this feature"). The allocation is then not touched. The failure path and its two cases still run.

Type

• fix — bug fix
• test — test-only

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally (the commit hooks pass).
• Unit tests: the test binary with MALLOC_PERTURB_=88 in a 3 GiB memory scope (systemd-run --scope -p MemoryMax=3G) is OOM-killed (137) after 3 s; with MALLOC_PERTURB_=0 both cases pass at once. meson introspect --tests on this branch shows the test's environment with MALLOC_PERTURB_: 0.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff. — not applicable.
• If I touched a feature extractor with SIMD/GPU twins, I updated every twin. — not applicable.
• New .c / .h file with a license header. — not applicable.
• Breaking change. — not breaking.
• ADR added. — not applicable: a test environment fix.

Bug-status hygiene (ADR-0165)

• docs/state.md: T-GPU-POOL-UAF-TEST-FILLS-HOST-MEMORY-2026-10-08 opened and closed here (PR fix(test): run the GPU pool UAF test without MALLOC_PERTURB_ so it cannot fill the host's memory #2547).

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 — not applicable.

Cross-backend numerical results

Not applicable: no score path changed.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial test environment fix.
• Decision matrix — no alternatives: only-one-way fix (the test must not write the oversized allocation; the cleanup path it tests is unchanged).
• AGENTS.md invariant note — no rebase-sensitive invariants.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — no changelog: test-only change, no user-visible delta.
• Rebase note — no rebase impact: fork-only test registration.

Reproducer

cat /proc/sys/vm/overcommit_memory      # 1
MALLOC_PERTURB_=88 systemd-run --user --scope -p MemoryMax=3G -p MemorySwapMax=1G build/test/test_gpu_picture_pool_uaf   # 137 after 3 s
MALLOC_PERTURB_=0  build/test/test_gpu_picture_pool_uaf   # 2 tests run, 2 passed