Research-2050: SpEED test native lint cleanup¶

Diagnosis and implementation¶

At parent 2143174c7bf933f97aec610f423c559a885b8e9a, a fresh configured CPU build reproduces eight constVariablePointer findings in test_speed.c and five in test_speed_qa.c. The extractor descriptors are only inspected or passed to vmaf_feature_extractor_context_create, whose live declaration already accepts const VmafFeatureExtractor *. Const qualification therefore requires no production API or descriptor ownership change.

Clang-tidy 22.1.8 also diagnoses six and nineteen C NULL uses respectively, plus 26 branches in test_speed_qa_temporal_component_positive against the configured limit of 15. Both C tests receive the existing ADR-1138 bracket for Windows C support and upstream C compatibility. The temporal test's four picture allocations and four extractor/collector setup assertions move into two small private helpers. Immediate failure returns preserve the original runner behavior. No assertion is removed or weakened, and neither helper is registered as another test.

No alternatives: the read-only declarations implement the existing API contract, and grouping implements the existing touched-file branch rule. There is no new policy or architectural decision and no new ADR.

Preservation controls¶

The retained token control compares the exact ordered lists of all 67 assertion expressions/messages, 10 mu_run_test registrations, 47 vmaf_* call sites and 170 numeric/string input literals against the parent files. It permits only the three explicit equivalent pointer/address substitutions in the moved extractor setup calls. The four grey-picture allocation calls retain their arguments through the helper parameters. Apart from comments and the eight const insertions, test_speed.c is token-identical. Meson registration is byte-identical to the parent.

Both actual original/current Meson binaries exit zero with byte-identical stdout/stderr, retaining five passing cases each. A disposable linker wrapper then makes each of the eight moved setup assertions fail through its real API boundary: picture allocations 7–10, descriptor lookup 6, and context creation, context initialization and collector initialization 4. Original and current binaries return one with identical messages and stdout/stderr for all eight cases. Only the disposable test/wrapper objects disable LTO so wrapping cannot be hidden by cross-TU inlining; the real production objects and repository inputs are unchanged. This checks propagation, not an expansion of allocation-failure cleanup coverage.

Native measurement and limits¶

The private GCC 15.2 CPU release build enables float and assembly, disables optional GPU/DNN/MCP/docs components, and uses CPUs 28–29. Both complete touched files pass strict clang-tidy with zero warnings and zero uncited exceptions. Exhaustive Cppcheck 2.21.1 with the official POSIX model reports zero findings in either test. Its 133-command reduced context also contains test.c variants and retains one unusedFunction finding in the untouched feature_collector.h; that invocation exits one and is not a full lint pass. No category is suppressed for that contextual diagnostic.

The real two-TU scoped writer records test_speed.c 6→0 warnings and test_speed_qa.c 20→0, preserving every unmeasured entry and original full report metadata. Total recorded debt changes 1359→1333 warnings with 54 uncited exceptions unchanged. This CPU-only test cleanup does not claim a fresh whole-tree make lint, make test, Windows/Darwin, GPU, sanitizer or Netflix Python golden pass. Production kernels, public headers, fixture values and golden assertions are untouched.

Reproduce and retain¶

In the configured CPU build:

meson test -C build --print-errorlogs test_speed test_speed_qa
python3 scripts/ci/tidy-ratchet.py --lane cpu --build-dir build \
  --only core/test/test_speed.c --only core/test/test_speed_qa.c

Canonical source hashes, tool and container identities, full configure/analyzer commands, before/after outputs, token and failure controls, scoped writer receipt and normal hook results are retained under .workingdir2/evidence/speed-tests-native-lint-20260908/. Build objects remain in the matching .workingdir2/cache/ directory. This is a component for root-owned integration; standalone publication is not requested.