Imported from inspiredrobot/mwcc (AGENTS.md). Install upstream with npx skills add inspiredrobot/mwcc. Copyright stays with the author.

Project rules

• Treat orig/, compiler binaries, DLLs, and generated Ghidra databases as local-only proprietary inputs. Never commit them.
• Treat every downloaded binary as untrusted. Parse object files and libraries as data only. Never execute a third-party EXE or load a third-party DLL on the host. If execution becomes necessary, use a disposable sandbox with no network, read-only inputs, and a dedicated writable scratch directory.
• Do not run candidate tools through host Wine or Wibo. A sandboxed run must also use a read-only root filesystem, drop Linux capabilities, enable no-new-privileges, and impose process and resource limits. Preserve the input hash, exact command line, and sandbox configuration with the result.
• Verify the configured SHA-256 before drawing conclusions from an executable.
• Record whether every name or structure is proven by this binary, inferred from its strings/control flow, or borrowed as a hypothesis from an external reference.
• This project's own contents are CC0 1.0 (public domain); anyone may reuse them for anything, with no attribution or permission needed. Keep new files under that dedication and do not add per-file copyright headers.
• The compiler studied here is proprietary and is never redistributed by this project. This is noncommercial research intended as fair use, and it will be taken down at a rights holder's request; never commit anything that would make the repository a substitute for the product.
• Do not copy third-party code unless its license or explicit permission allows it. Record copied files and modifications in docs/PROVENANCE.md.
• mwcc-debugger and Ninji's MWCC decomp are reference-only until their licenses are clarified. Facts learned from them must be independently checked against stock GC/1.2.5.
• Treat GC/1.2.5 as the primary historical target. GC/1.2.5n is a derived patch configuration, not a replacement for the stock artifact.
• Do not assume the compiler was self-hosted. Keep host compiler and linker identification evidence in docs/HOST_TOOLCHAIN.md and require output comparison before calling a candidate confirmed.
• Prefer typed, structured C over assembly-shaped output. Do not encode a guessed structure merely to improve a superficial diff.
• Prefer typed static helper functions over statement-like macros when shared behavior has meaningful inputs or state. Keep macros for declarations and compile-time concerns, not reconstructed control flow.
• Preserve the order and reason for compiler passes as first-class findings.
• Before core backend work, read docs/CORE_SUBSYSTEMS.md, docs/DATA_MODEL.md, docs/DECOMP_WORKFLOW.md, and docs/DECOMP_ASSISTANCE.md. Update the subsystem manifest when an address-backed role or name is confirmed or rejected.
• Keep commits focused and run ninja check plus relevant object diffs before committing. Run python3 tools/check_format.py --fix before committing C or header changes; ninja check enforces the result.