test transpilers on topologic

[Tokarzewski]

https://github.com/wassimj/Topologic
Do CPP to mojo transpilation

[Tokarzewski]

Ran the strict C++→Mojo engine against TopologicCore/src/*.cpp (51 files) from wassimj/Topologic, a real-world CAD/geometry-kernel library built on OpenCASCADE (OCCT). Fixes + tests pushed to claude/repo-review-n4ylcz.

Headline result

0/51 files transpile as shipped, 100% failing at the parse stage. Two independent causes stack:

1. Multi-file architecture. Every .cpp here is a bare method-definition file (namespace X { Ret Class::Method(...) {...} }) with the class declared in a separate .h. The engine's frontend strips every #include by design (toolchain-free, single-file model) and has no header-resolution mechanism, so libclang never sees the class the out-of-line methods belong to (use of undeclared identifier 'Bitwise', etc.). This is a deliberate design point (matches the examples/ corpus of self-contained slices / EnergyPlus "god-object" hand-flattening), not a bug — but it means arbitrary multi-file C++ repos need manual per-artifact flattening before this tool can touch them.
2. OpenCASCADE dependency. Once I hand-flattened header+source per file (recursively resolving local #include "X.h" chains), every non-trivial class still bottoms out in Utilities.h, which does #include <TopoDS_Shape.hxx> (OCCT) and declares an OCCT-typed helper. Since that same header is also where the project's TOPOLOGIC_API export macro lives, virtually every class transitively drags in an OCCT type the engine's std::-only preamble has no shim for. This is a hard, reasonable wall — OCCT is a huge third-party CAD kernel, well outside a general transpiler's std:: preamble.

Bugs found + fixed along the way (general, not Topologic-specific)

Flattening surfaced 4 real frontend gaps that would bite any real-world C++ corpus, not just this one. Fixed with regression tests in claude/repo-review-n4ylcz (538→543 tests, all green):

• Unresolved export/calling-convention macros (FOO_API, DLLEXPORT, WINAPI, ...) made every declaration in a file fail to parse — near-universal C++ library idiom, defined in a header we never see. parse_cpp now retries with such SCREAMING_CASE names neutralized via -D.
• std::list<T> shim was a bare opaque class (unlike the fuller std::vector<T> shim) — no begin()/end(), so range-for over a std::list failed to parse.
• std::list<T> had no backend type mapping — once parsing succeeded, every backend crashed with "unresolved type hole." Mapped alongside stack<T>/queue<T> → list[T].
• std::string had no (const char*) constructor in its shim, so the extremely common std::string("literal") failed to parse.
• (Most impactful) The header-guard-stripping fallback — used whenever the real clang -E step fails, which is virtually guaranteed on any real multi-file project with a missing system header — blindly stripped every #endif and bare #define globally instead of only matched #ifndef/#define/#endif triples. This corrupted ordinary #ifdef/#else/#endif blocks (ubiquitous for platform export macros) into malformed directives that confused libclang ("#else after #else"). Replaced with a depth-tracked scanner that only strips genuinely-paired guards.

With these fixes, TopologicCore/src/Bitwise.cpp (a self-contained utility class with no OCCT touch) transpiles cleanly end-to-end to valid Mojo.

Bug found, not fixed (filed separately: #80)

Bitwise::NOT(int) / Bitwise::NOT(unsigned int) — two overloads differing only by signedness — both collapse to an identical (Int) -> Int Mojo signature, emitting two methods with the same signature in one struct (guaranteed duplicate-definition compile error).

Bottom line

Topologic isn't a good near-term fit for this tool as-is: the interesting logic is inseparable from OCCT, and the multi-file layout needs manual flattening per artifact (same cost as the EnergyPlus corpus work). The 4 parser fixes are real wins for other corpora though, and are already regression-tested. If there's appetite for a deeper push here, the next lever would be an OCCT type-shim layer in the preamble (mirroring the existing std:: shim) — a much bigger undertaking than this pass.

---

Generated by Claude Code

[Tokarzewski]

Progress update (2026-07-12)

The multi-file parse wall from the Topologic stress-test now has an automated path via the codebase-memory-mcp knowledge graph.

New: transpile <src> --source cpp --target <t> --cbm <repo_root> (cli/main.py) queries the cbm graph for the file's cross-file class/method/macro surface and injects it as a parser preamble (``). This is the inverse of the manual header-flattening the strict frontend otherwise needs: out-of-line Class::Method definitions resolve against the recovered class declaration instead of failing with 'use of undeclared identifier'. `tests/test_cbm_preamble.py::test_parse_cpp_consumes_cbmpreamble` proves it deterministically (out-of-line `Bitwise::NOT(int)` parses with the preamble, fails without it).

Also added: auto OCCT (OpenCASCADE) opaque shim — when the graph shows a file references an OCCT type (TopoDS_*, TopAbs_*, BRep*, gp_*, ...) the preamble appends docs/occt_preamble.hpp-style opaque decls (issue #79's 'OCCT wall'), mirroring the existing manual shim but now automatic. Tested via test_occt_shim_appended_when_occt_type_detected.

Verified here: the transpilers repo is already indexed by cbm (175K nodes); live query_graph works against it. The Topologic 51-file pass itself still needs a fresh codebase-memory-mcp index of the Topologic repo on this machine (the one-shot index_repository CLI crashes on a worker subprocess here — needs the watcher-server mode / a released binary, open as follow-up).

Not solved: a Topologic file whose own logic calls deep into the OCCT API (BRepBuilderAPI_, Geom_, gp_*) still won't reach a compiling Mojo output — that needs a real target-side binding, which doesn't exist for Mojo. The shim only gets libclang to parse past the type.