Object-matching audit helpers
These scripts preserve focused COFF comparison, donor-search, disassembly, and residual-ranking utilities used during exact-object reconstruction. Run them from the repository root unless a script accepts explicit paths.
The disassembly helpers require the Python capstone package. All COFF
parsing and exactness decisions come from the repository's tracked
tools.coff_compare implementation; these helpers do not alter admission or
matching status.
error_geometry_runtime_differential.py compares the real January and rebuilt
Error Geometry caller graphs in separate Unicorn instances. It requires the
Python unicorn package, the local cachebeta.exe, symbol manifest, and built
objects. Run without Python optimization (-O is rejected):
python -B tools/audit/error_geometry_runtime_differential.py --output scratch/error-geometry-runtime.json
Use --target, --candidate, --expected-target-sha256, and
--expected-candidate-sha256 to pin frozen inputs. The 24 bounded cases cover
file lifecycle, names, primitive/bounds calls and variadic comments. Genuine
external APIs are modeled: this does not execute host CRT file operations or
prove locale-rendered output, non-identity transforms, arbitrary inputs,
byte exactness, or whole-object/link ownership. The report records these limits
and input/tool hashes; only the requested JSON is written. Unsupported format
conversions and attempts to overwrite verification inputs fail closed.
Files with dates in their names are retained research snapshots. Their hard-coded inputs intentionally document the evidence wave they were written for, while all imports and repository paths are portable.
For a source-level credibility pass, run python tools/fake_match_scan.py.
It reports deterministic review leads for inert, contradictory, or explicit
codegen-steering constructs without treating a heuristic hit as proof. See
docs/fake_match_review.md for scope, vendored exclusions, and exit codes.