blackfrost-ai/cyber-frost-harness/native-build-and-replay

Native build and replay

Prefer the shipped executable in /out. It was built with the matching project toolchain and sanitizer. Rebuild only when a concrete experiment requires instrumentation, a source change, or a missing target.

Before rebuilding

• Run file on the target and a sample of object files and static libraries.
• Read /src/build.sh, project build metadata, and /out/*.options.
• Record $CC, $CXX, $CFLAGS, $CXXFLAGS, $LIB_FUZZING_ENGINE, $SRC, $OUT, and $WORK.
• Remove or isolate stale objects when architecture or compiler provenance differs.
• Build in a new directory instead of layering over shipped objects.

Mixed EM_X86_64 and AArch64 objects, incompatible C++ standard libraries, and stale generated configuration files are hard failures. Detect them before linking:

find BUILD -type f \( -name '*.o' -o -name '*.a' \) -print0 |
  xargs -0 -r file | sort -u

Focused build

Use the project's own fuzz build script when possible. Preserve the OSS-Fuzz environment variables and request only the required target. For manual builds:

• compile libraries with -fsanitize=fuzzer-no-link,address or the image's supplied flags;
• link the final harness with $LIB_FUZZING_ENGINE or -fsanitize=fuzzer,address;
• keep debug symbols and frame pointers;
• disable optional components rather than repairing the entire application suite;
• verify the resulting binary on one known-good seed immediately.

Do not poll long builds with repeated sleeps. Launch once with a status file and inspect process state, status, and the first actionable errors in one command. Stop a build that cannot contribute to the current hypothesis.