# Claude contract-pilot PR #51 · independent integration review

16 October 2026 · Head `3a330788a840e6757bd47cee4e0731c08499478f` · **Changes required before integration**

PR: https://github.com/215GrandAve/lost-souls/pull/51

The activated tooling-review agent hydrated only text into an isolated audit folder and ran the contract suite and focused probes. No Claude source, GitHub state, production contract, Blender scene or Unreal project was changed. Green CI does not cover all four reproduced cases below.

## Reproduced corrections

| Case | Actual result | Required change and regression evidence |
| --- | --- | --- |
| Nonpositive hollow-wall thickness | Setting fixture `envelope.exterior_wall_thickness` to 0 or -0.3 returns COMPLETE with no errors. A production-scoped/proposed-provenance variant remains issuable. Negative thickness expands the inner footprint outside the shell. | Reject nonpositive/nonfinite thickness before inset, for the envelope default and per-mass override. The issue gate must refuse these cases. |
| Sections ignore real openings | A section at x=-3 through fixture D01 is byte-identical when D01 changes from real opening to solid-wall marker, with a constant stamp. Section wall polygons do not subtract the hole. | Derive section wall intervals with opening cuts and glazing. A slice through the opening must differ from marker/solid wall; slices outside it stay solid. |
| Roof thickness differs between section and model | With RF slab thickness 0.8 and L01 thickness 0.3, analysis stays COMPLETE. Blockout uses 0.8 for the roof while the section draws 0.3. | Share a derived roof-slab record between drawings and model. A nonuniform-thickness fixture must give the same roof elevations in both. |
| Missing identity/revision can be issued | Removing asset_id or contract_revision from a production-scoped fixture still returns COMPLETE and issuable. | Require nonempty asset identity and contract revision. Test missing, null and empty values and refuse ambiguous UNKNOWN/NOREV issuance. |

Relevant implementation areas are `tools/contracts/contract.py`, `draw_svg.py` and `blockout.py` on the pinned PR head. Claude retains implementation ownership.

The probes used the actual imported modules: `C.analyse(candidate)`, `C.issue_gate(candidate, provenance, synthetic_production_contract_path)` and `D.section_sheet(report, constant_stamp, {"id":"DOOR","axis":"x","at_m":-3})`. The synthetic issue path in the audit was `production/assets/buildings/PI-BLD-9999-V01/BUILDING_DIMENSIONS.json`; it was a probe input only, not a new registered asset.

## R02 threshold and scope mapping

Read [SAMPLE_SCHEMA_MAPPING_REVIEW.md](SAMPLE_SCHEMA_MAPPING_REVIEW.md). At mass base z=0, hollow-mass floor slabs use the inset inner polygon. R02's y=4..10 footprint therefore becomes y=4.3..9.7 with 0.3 m walls, leaving unsupported threshold strips.

This is a convention to reconcile, not proof that the geometry is impossible. A below-sill perimeter-ring convention with base z=-0.25 may meet an inset slab at z=0, but it has not been selected or generated, and the connection checks currently equate mass base with walking elevation. Claude should explicitly implement and test either full outer-footprint slabs or a documented ring convention with separate walking elevation. Keep R02's intended full-floor/threshold behavior intact.

Do not add overlapping outer and inner slabs silently. Test downward floor support rays immediately before, within and after every threshold, and verify no duplicate coplanar floor surfaces. Reconcile the larger arch/portal and seawall content in the skeleton with R02's intentionally smaller generic proxies; do not combine their scopes automatically.

## Commands and evidence actually checked

- `python -m unittest discover -s tools/contracts/tests -p 'test_*.py' -v`: 46 tests run; seven Blender tests skipped because bpy was unavailable; the remaining tests passed.
- `python tools/contracts/validate.py --issue tools/contracts/fixtures/TEST_FIXTURE_ORIENT_01.json`: correctly refused the test-only fixture, exit 3.
- `python tools/contracts/validate.py --ci`: exit 0 in the isolated audit tree, which contains zero production contracts. This is not an independent full-repo validation.
- Both remote Verify checks on the pinned PR head: success.
- Fixture canonical hash matches its recorded evidence: `650b65a2adca6bb6ae600a09cb60f3f5060f07a1c2dacc3997ca19ccf7625d3c`.

Opening/marker/reservation separation and SVG provenance/determinism have useful tests. Starter changes preserve authored schema-v2 contracts; no bulk production migration is included. The new `.gitattributes` exceptions are limited to pilot evidence files.

Blender renders/GLB reproducibility remain Claude's recorded evidence; this independent review did not execute Blender or Unreal. Automatic `build_quik_handoffs.py` issue-gate integration is explicitly still unimplemented, so do not claim all packaging paths are enforced.

## Next handoff

Claude fixes the four reproduced cases, selects/tests the floor convention, and returns an updated head plus focused regression results. Codex then rechecks the changed cases and maps R02 into the reviewed schema. Quik's actual importer, character/camera, collision/navigation and traversal checks remain separate. PR #51 stays open at this head; main publication of the visual pilot does not accept the tooling.
