G13 — cross-architecture comparison guardrail¶

Registry: UC-PLAT-arch-guard (planned) Effort: S · Risk: low

Problem¶

Comparing an x86-64 build against an aarch64 build of the same library version returns a falsely reassuring COMPATIBLE_WITH_RISK verdict at 100% binary-compatibility. The ELF snapshot never captures e_machine (nor EI_CLASS 32/64, nor endianness), so a cross-architecture diff is not even recognised as such — the only incidental trace is a toolchain_flag_drift note scraped from DW_AT_producer. This is a real "false green" footgun for multi-arch CI matrices. PE and Mach-O already carry a machine field — ELF is the outlier.

Goal & acceptance criteria¶

[ ] ELF snapshot captures e_machine, EI_CLASS, and endianness.
[ ] compare / compare-release treat a machine mismatch as a hard guard: either refuse with a clear error, or emit a top-level ARCHITECTURE_MISMATCH finding that dominates the verdict (never a compatible/low-risk result).
[ ] Same-arch comparisons are unaffected.

Design¶

Add the machine fields to abicheck/elf_metadata.py and the ELF snapshot model (PE/Mach-O already expose an equivalent).
In the diff entry point, compare machine/class/endianness first; on mismatch short-circuit to the guard outcome (new ARCHITECTURE_MISMATCH kind in checker_policy.py, classified to dominate — or a refusing ValidationError, decided in the plan's first step).

Files & surfaces¶

abicheck/elf_metadata.py, abicheck/model.py (ELF block), abicheck/checker.py / abicheck/diff_platform.py (guard), abicheck/checker_policy.py (kind).

Tests¶

tests/test_arch_mismatch_guard.py: x86-64 vs aarch64 snapshot → guard outcome; x86-64 vs x86-64 → normal path.

Out of scope¶

Deliberate multi-arch fat binaries (Mach-O universal) — those expose multiple slices and are a separate concern.