add README from _to_move_to_dataset_repo/

README.md

@@ -0,0 +1,32 @@
+# Dataset-curation scripts (holding folder)
+
+These scripts validate that fixture ground-truth STEP files are well-formed.
+They are **dataset-authoring tools**, not part of the public benchmark code.
+
+They live here temporarily until the `cadgenbench-data` HF dataset repo is
+created, at which point they will be moved there with `git mv`. Until then,
+they remain in this holding folder so they are preserved in git history but
+explicitly excluded from the `cadgenbench` package build (see
+`pyproject.toml`'s `tool.setuptools.packages.find` exclusion).
+
+## Contents
+
+- `sanity_check_gt.py` — validates that each fixture's `ground_truth.step`
+  is canonical-posed, watertight, and BREP-valid.
+- `sanity_check_jig_metric.py` — runs the interface-match metric on
+  each fixture's jig sub-volumes against its GT and asserts saturation.
+- `sanity_check_submission.py` — validates a candidate submission directory
+  conforms to the submission spec (`docs/benchmark/submission.md`).
+- `sanity_check_topo.py` — validates Betti-number agreement between the
+  GT BREP and its tessellation.
+
+## Why they were moved out
+
+These tools require the GT STEPs to run, which we may eventually keep
+private. Coupling the public benchmark code to internal authoring
+workflows would make a future split painful. Keeping them in this
+holding folder makes the lift a one-command `git mv`.
+
+Any metric-exercising logic that is genuinely useful as a regression
+test for the metric code itself should be lifted into `tests/eval/`
+rather than living here.