Skip to content

Research-2054: Preserve motion, SSIM and PSNR coverage during lint cleanup

At fb79390160fdfc3f8316dd5ab6cae4c05de2b58d, exhaustive Cppcheck reports seventeen read-only descriptor views across test_integer_motion_v2_coverage.c, test_ssim_coverage.c and test_integer_psnr_coverage.c. Strict clang-tidy also reports thirteen oversized test bodies and one missing loop brace. The drivers already meet the branch limit; their registrations remain intact.

Descriptor views become const, including PSNR's private descriptor-return helper. Thirteen test-specific setup helpers retain lookup, option insertion, context creation/init and collector creation in their original order. Each caller immediately returns the original setup failure. The remaining picture, extract, score and teardown statements are unchanged. SSIM's outer pixel loop gains braces. Existing C NULL portability brackets remain intact.

Alternatives considered

Choice Result
Const views and small test-specific setup stages Selected: preserves exact assertion text, options, order and failure behavior.
Generic parameterized test setup Rejected: would obscure distinct messages, options and ownership paths.
Suppress the size findings or change production APIs Rejected: unnecessary and outside this observation-only cleanup.

This applies existing test/lint policy; no new architectural decision or public surface is introduced. A user guide and FFmpeg patch changes are not required.

Validation and limits

All nineteen existing cases pass before and after in a private GCC 15.2 CPU release build with assembly and default LTO. Their stdout/stderr match exactly. Source checks preserve all 133 assertions, 183 API calls, nineteen ordered registrations and every string/numeric literal. Expanding each helper back to its original setup leaves identical whole-file tokens except const qualifiers and the single loop-brace pair.

A retained C probe includes the original or changed test source and intercepts only its five setup APIs, forwarding successful calls to the actual configured library objects. Across fifteen affected scenarios, each of sixty-six moved setup stages is failed independently. Original/current API traces, failure messages and immediate stopping points match. The fifteen normal probe paths also pass. The probes are local evidence, not new repository tests.

The actual release executables and diagnostic non-LTO objects differ after helper extraction. They are retained with section hashes for inspection; there is no binary-identity or performance claim.

Strict clang-tidy 22.1.8 measures all three TUs at zero warnings, uncited annotations and compile failures. The guarded scoped writer removes only their five/seven/two warning allowances and retains full-lane metadata. Exhaustive Cppcheck on all six configured test/driver commands reports no owned-TU findings. It exits one solely for the unchanged unused inline vmaf_feature_vector_get_score in feature_collector.h; this reduced-project header finding is retained, not suppressed or described as a whole-tree pass.

Reproduce the existing tests:

meson test -C build --print-errorlogs \
  test_integer_motion_v2_coverage test_ssim_coverage test_integer_psnr_coverage

Locally retained commands, source hashes, analysis databases, old/new binaries, original-source strict diagnostics, all failure traces and comparison scripts: .workingdir2/evidence/metric-coverage-const-2026-09-08/. Production sources/headers, Meson registration and the common test driver are unchanged. No golden assertions changed. Full combined RC1, Windows/GPU, Netflix Python golden and sanitizer acceptance remain separate.

References