[devin-ai-integration]

Note: This is just an experiment to implement the entire C++ transpiler using LLM - it will never be actually merged.

Summary

Adds a second compiler backend (src/engine/internal/cpp/) that turns Scratch scripts into C++ source code instead of LLVM IR. It plugs in at the existing ICodeBuilderFactory boundary. The default backend is still LLVM, so nothing changes for current users.

Covers backlog items #662–#681, #683–#687 and #689–#691. Not included: #682 (inlining), #693 (type analyzer) and #692 (in progress elsewhere).

How it works

• CppCodeBuilder implements all of ICodeBuilder. Each script, procedure, reporter or predicate becomes a heap Frame subclass with a switch-based step() state machine. This is C++17 (no C++20 coroutines needed). All registers live as frame members, so they survive yields.

  int step() override {
      switch (state) {
          case 0:;
          ...
          state = 1; return cpp::CPP_SUSPEND; case 1:;   // yield / end of loop iteration
          ...
      }
      return cpp::CPP_DONE;   // CPP_END_THREAD for "stop this script"
  }

• Value representations: double, bool, cpp::Str (pooled StringPtr), cpp::Val (ValueData) and void *. Conversions use the public value_* / string_* functions. Variables and lists are resolved from varData/listData (or the stage) in resolve(), which runs again after invalidateTarget() and procedure calls.
• cppprelude.h is the runtime header for generated code. It only uses public headers and is embedded into the library as a string at configure time (CppCompilerContext::prelude()).
• CppCompilerContext::moduleSource() outputs a module whose static initializer calls cpp_register_function(moduleKey, name, fn). moduleKey is an FNV-1a hash of the generated code. CppExecutableCode looks the function up by (moduleKey, name). So code only runs if the matching generated source was compiled into the binary.
• CppExecutableCode follows LLVMExecutableCode: it waits for unresolved promises, reset() destroys the suspended frame, and the (void *)0x1 end-thread sentinel is passed up through procedure calls.
• Pointer constants that point to a target's list or variable (e.g. from listblocks) are emitted as resolved references instead of raw addresses. Other pointer constants are still only valid in the current process.

Tests (#687): test/cpp has a cpp_test_generator executable. At build time it runs the shared test-case builders (cases.cpp) through CppCodeBuilder and writes generated_cases.cpp, which is compiled into cpp_test. cpp_test then rebuilds the same cases and runs the resulting ExecutableCode, checking stdout, run counts and return values. It also covers promise/kill/reset and the codebase generator. 24 tests pass locally, with and without LIBSCRATCHCPP_ENABLE_SANITIZER.

Codebase generation (#683–#686): CppCodebaseGenerator::generate(sb3, outDir, name) loads the project with a recording C++ factory and writes:

• project.cpp: the prelude plus all modules
• project_data.cpp: the .sb3 as a byte array
• main.cpp: extracts the .sb3 to a temp file, calls cpp_use_cpp_backend(), then runs Project::load() and run()
• CMakeLists.txt: uses either LIBSCRATCHCPP_SOURCE_DIR or a prebuilt LIBSCRATCHCPP_INCLUDE_DIR/LIBSCRATCHCPP_LIBRARY

The new scratchcpp-transpiler <project.sb3> <outdir> [name] tool (option LIBSCRATCHCPP_BUILD_TRANSPILER, default ON) wraps the generator. I checked this locally on test/custom_blocks.sb3 and test/bubble_sort.sb3: the generated projects build and run against libscratchcpp. 3_threads.sb3 never finishes, but it doesn't finish with the LLVM backend either.

Known limitations / review focus

• No type analysis yet: variables and list items are always ValueData.
• Generated code calls block functions through extern "C" declarations built from the addFunctionCall arg types. These have to match the real signatures (same requirement as LLVM).
• Generated codebases only load the project when they start; they don't embed assets separately (Embed project assets into generated C++ codebase #685 is covered by embedding the whole .sb3).

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring