Solidrift 0.1.2: Native geometry microscope with checked report integrals

Recompute retained mesh output integrals during report verification and bind mirrored failure signatures. This addresses a real GitHub Codex review: altered baseline/witness/undo geometry and scalar records passed older verification; they now reject. The authentic native report still verifies. Includes recorded-report compatibility, small-scale tolerance, corruption controls and cross-Python accumulation handling. Both Python 3.10/3.14 hosted checks passed 27 tests.

Fix the 0.1.0 Witness/reference-overlay rendering bug when no vertex is selected. The regression reproduces the original exception and now exercises Witness, Repair and Undo. Real browser rendering and vertex ancestry passed with no new console errors. Native analysis code and recorded geometry are unchanged.

Solidrift turns a supported broken mesh Boolean into a valid small reproducer, source ancestry and a reversible geometry intervention tested on the full original inputs.

Try the recorded native 3D investigation or install the attached wheel with python -m pip install solidrift-0.1.2-py3-none-any.whl. Python 3.10+ and a separately installed Blender with FLOAT/MANIFOLD are required for analysis; Python runtime dependencies are empty. python -m solidrift demo --blender /path/to/blender --out case performs a fresh native search. The evidence ZIP contains an offline HTML viewer, original bytes, OBJ exports, patch and stored report.

Verified on Blender 5.3 Alpha bad8871dc58d: 132 → 12 input triangles, 60 accepted valid contractions, 395 native queries. A three-witness-vertex boundary intervention lifts to 13 original vertices; full repaired solver surfaces have matching oriented triangles with separation bounded by 4.27162e-6 model units. Actual input undo reproduces the failed surface. Separate native review, conventional simplification controls, healthy/budget/timeout checks, 27 Python contracts, 7 viewer controls, Linux Python 3.10/3.14 installed-package CI and a fresh installed-wheel native run passed.

Scope: one FLOAT-union omission family, bounded triangular single outward closed solids, greedy search and a conservative correspondence comparator. No universal repair, global minimum or formal geometry-proof claim. Repairs change inputs. The demo fixtures are independently authored Manifold regression data with exact source identity and Apache-2.0 license retained; original implementation is MIT. Methods and evidence.

Implementation and automated reviews used Codex assistance. No human approval, upstream acceptance or outside adoption is implied.