Registered mesh-transfer foundation (#4381)

This slice adds native/WASM planning, not a new UI command or a measured real scan registration. It composes a registered opaque image observation over the existing target appearance through the same atlas and IFC material planner used by finite pages. Gaps remain explicit unknown observations.

The independent run invokes actual rebuilt WASM, applies its typed mutation plan through IfcOpenShell, serializes/reopens IFC, checks independent geometry, decodes PNG with Pillow and evaluates UV/color samples with NumPy. The controlled 1 m triangle has a smaller observed triangular patch. Four source image quadrants distinguish U/V orientation; an uncovered point retains its explicit original RGB. The source is a declared invariant fixture, not a real scan/IFC pair. Five interior/background samples pass a 0.025 normalized-channel ceiling; the actual worst delta is recorded in JSON.

Native invariants additionally reject thin-wall look-through, overlapping source faces and UV-seam ambiguity, while allowing continuous shared edges. They check partial coverage area accounting, real IFC/PNG reopen, missing/stale identity, unsupported alpha/tint, reflected placement, changed pixels and exhausted work. The actual WASM contract repeats coverage, payload binding, no-applicable-plan and refusal checks. Existing page composition tests protect source appearance, previously applied pages and material properties through the shared refactor.

Reproduce after bash scripts/build-wasm.sh:

cargo test -p ifc-lite-processing appearance::transfer --lib
pnpm test:wasm-contract
python3 tools/texture-authoring/mesh-transfer-oracle.py

The oracle needs IfcOpenShell, Pillow and NumPy installed locally. It does not fetch assets or provide an application IFC writer. Browser acceptance, source lease/frozen-frame wiring, ordinary Apply/Undo/export/share and real distributed held-out correspondences remain separate gates under #4381. CRAS plane-support evidence does not supply those missing correspondences.

Real boulder geometry: conservative coverage and capacity

The public CC0 albedo-only boulder GLB, already qualified for captured-mesh creation, supplies 66,122 source triangles. This probe uses the existing captured IFCZIP from the portability acceptance, with a 40,087-triangle proxy. It first verifies that target coordinates are the known Y-up→Z-up conversion of that exact GLB. The four fit/four check observations are therefore known-derived controls, not independent scan/BIM correspondences. The embedded fit residuals measure that control transformation only.

The complete target refuses the aggregate BVH work budget; it produces no partial mutation plan. This foundation does not yet qualify model-wide transfer at that size. A future measured optimization should reuse conservative candidates across target tiles/triangles while preserving unknown coverage and one final atomic Apply; raising the cap is not the acceptance gate.

The first 500 target triangles, retaining original coordinates and parallel UV maps, complete within the same bounds. The sampled area estimate is about 77% observed. Of 22,915 centroid/interior samples, 5,538 are ambiguous; no samples fail distance or normal criteria in this measured prefix. Exact coplanar UV-continuous neighbors are exempt from ties; curved facets and seams remain conservative unknowns. This result is a bounded prefix measurement, not full-boulder coverage or proof of smooth-surface continuity. The preserved old boulder image supplies unknown albedo. Source and effective derived IFC identities are in each report.

Reproduce with the existing qualified GLB and captured IFCZIP:

python3 tools/texture-authoring/boulder-transfer-probe.py source.glb captured.ifczip
python3 tools/texture-authoring/boulder-transfer-probe.py source.glb captured.ifczip 500

The script is deliberately a hash-qualified offline evidence decoder, not a new application ingest path. It requires SciPy as well as the oracle dependencies.