[brofield]

Summary

The Linux allocation-failure hooks in tests/ts-regressions.cpp replaced nothrow operator new / new[] with __wrap_malloc. That is required: SimpleIni allocates with new (std::nothrow), which libstdc++ implements in a DSO, so --wrap=malloc never sees those calls.

Returning malloc memory from operator new made AddressSanitizer abort before any test body. testing::TestSuite::Run() calls std::stable_sort, libstdc++ get_temporary_buffer uses ::operator new(n, std::nothrow), and the matching free is sized ::operator delete. ASan reports malloc vs operator delete.

Keep the same fail-after-N / fail-small policy for both __wrap_malloc and the nothrow new replacements. On success, nothrow new now delegates to throwing ::operator new so delete[] matches. Do not call __wrap_malloc from operator new, or the budget would decrement twice.

Also cover every SI_NOMEM allocation site that the default Linux build can reach:

• First-load LoadData parse buffer (new SI_CHAR[uLen+1])
• LoadFile read buffer (new char[uSize+1])
• LoadFile read buffer succeeds, then LoadData parse buffer fails
• CopyString via SetValue (new returns nullptr)
• Incremental LoadData file-comment CopyString failure
• Incremental LoadData AddEntry / CopyString failure (existing leak test no longer starts with a file comment, so it actually hits AddEntry)

Not covered, on purpose: ICU UCharPathToUtf8 (not in the default build) and std::map bad_alloc (not turned into SI_NOMEM).

Validation

• Official Debug suite: 290 tests pass (GCC 13.3, C++17)
• ASan+UBSan with default alloc_dealloc_mismatch=1 and detect_leaks=1: 290 tests pass
• LoadDataRegression and LoadFileRegression OOM cases pass in both builds