Research tools

Start with current status. The supported research entrypoint is python -m experiments.baseline --help: the guarded retry control, with separate motion checks and exact replay. It is not a completed puzzle.

Reusable modules:

Module	Responsibility
`numerical_geometry`, `native_geometry`	Numerical polyhedra, reference/native kernels, benchmark helpers
`matching`, `search_geometry`	Rigid matches, subtraction, numerical audits
`search_state`	State, splitting, rollback/repair primitives, saved geometry
`numerical_proposals`, `boundary_proposals`	Numerical cut/pose proposal machinery
`search_components`	Shared guarded search and historical comparison modes
`assembly_guidance`, `paired_guidance`	Independent numerical removal and proposal screens
`exact_replay`, `certification/`	Exact reconstruction and independent acceptance checks
`state_io`, `artifact_paths`, `rollback`	Saved-state loading, archive aliases, rollback control

Other diagnostics and proposal libraries remain active dependencies. Shared historical modes are retained for comparisons; only the guarded retry is the documented baseline. The original ten campaign runners live in archive. The browser demos are independent of these Python entrypoints.

Run the reference suite with uv run --with scipy python -m unittest discover -s tests -p 'test_*.py', or the native suite with uv run --extra native python -m unittest discover -s tests -p 'test_*.py'. Archived runner tests still run: moving a campaign out of the active path does not discard its regression coverage.