Skip to content

Research-2049: Preserve observation-only SVM tests while cleaning native lint

At 2143174c7bf933f97aec610f423c559a885b8e9a, exhaustive Cppcheck reports nine read-only model pointers in test_svm_parser.c and five read-only query arrays in test_svm_api.c. Fresh clang-tidy 22.1.8 also reports eleven parser warnings and seventeen API warnings: C null spelling plus the parser runner's eighteen branches. The runner exceeds the fifteen-branch limit because each mu_run_test expands to two branches.

The parser's model views and API query arrays are now const. Header-size and header-order cases move into two small driver helpers, preserving the original case order and first-error return. Both C files gain the already-required ADR-1138/ADR-1166 modernize-use-nullptr bracket; NULL remains compatible with MSVC's C frontend. No new diagnostic exemption or language policy is introduced.

Alternatives considered

Choice Result
Const local views and named parser groups Selected: remove the findings without changing public calls or test behavior.
Rewrite the vendored parser or its public types Rejected: these tests are observation-only under ADR-0952/ADR-0889.
Disable the branch limit or replace C NULL with nullptr Rejected: weakens existing lint or violates the established MSVC C portability rule.

Validation and limits

An independent private CPU build uses GCC 15.2, release optimization, assembly and default LTO. The original and changed binaries both pass the nine parser and nineteen runtime API tests. Full stdout/stderr agree byte-for-byte. A C-token comparison preserves all fifty assertion invocations, twenty-eight ordered test registrations, sixty-five public SVM calls and every string literal. svm.cpp, svm.h and Meson registration remain byte-identical.

The API executable's .text and .rodata sections are byte-identical. Parser sections differ after grouping, so no binary-identity claim is made for it. A separate retained runner control replaces only the public parser entry point with a stub: the all-pass case and each of nine injected failure positions produce identical old/new output, first-error return and test count. This checks driver propagation; the ordinary native runs above use the real library.

Actual scoped clang-tidy measures both touched TUs at zero warnings, compiler failures and uncited annotations. The guarded writer removes only their existing allowances, retains unselected entries and full-lane metadata, recomputes debt totals and appends scoped provenance. Exhaustive Cppcheck with the official POSIX model and both actual test-driver compile variants exits zero. No tests, assertion values, model inputs, ownership rules or registrations are added or removed.

Reproduce with meson test -C build --print-errorlogs test_svm_parser test_svm_api. Exact commands, source/tool hashes, original binaries, section comparisons, runner controls and hook results are retained locally under .workingdir2/evidence/svm-tests-native-lint-2026-09-08/. This is test-only validation: no complete RC1, Netflix Python golden, Windows/GPU or sanitizer acceptance is claimed. There is no user-discoverable surface or FFmpeg integration change, so no user-guide update is required.

References