[erlis]

Two difftest runners import B-reps whose geometry owners don't match the names the runner hands the library. The topology solver depends on those owners:

• BrepObjects::curve_edge finds an edge from its curve's owner;
• BrepCache::build_trees adds an edge or face to the solver trees only when its curve's or surface's owner equals the caller's key.

In the C++ fixtures an edge owns its curve and a face its surface, so the inputs differed.

topology.brep_p6, op 3 MakeFaceTopology

• The import owns face surfaces under the runner's key_of names but leaves edge curves unowned.
• In cube_faces_6, the infinite-region ray from (1.50002)³ hits the edge y=1,z=1 at t=0.9427. The C++ solves it there, which sets best² = 0.8888 and prunes the x=0 and y=0 face boxes.
• Rust skips that edge (curve_edge is None), so it classifies face x=0 and creates 4 UV trims the C++ never makes.
• Fix: own every edge curve under key_of(Edge).

boolean.merge_p2, MergeEdge / MakeFaceuse

• p5::import owns the operands' geometry under arena identities (edge i → key i), but the runner names entities with the copy keys (edge i → 2 000 000 + i, face i → 4 000 000 + i).
• MakeFace's infinite-region rays use those names, so their tree held no edges and only the new face, and the shell rays failed with edge owner is not in source B-rep.
• The C++ solves the hit edge, treats the ray as ambiguous and fires the 0.789 / 0.189 retries.
• Fix: own both operands' curves and surfaces under the copy's destination and source names.

util::own_edge_curves / own_face_surfaces take the names; own_arena_geometry now calls them with arena identities (unchanged behaviour).

Results (Linux, pinned reference c9c5979; tolerances and fixtures unchanged)

This branch (merged with main 7935b4c):

PARTIAL topology.brep_p6 cases=595 passed=586 failed=0 deferred=9 max_rel_err=0e0 with_values=586 no_values=0 vacuous=0 blocked_by=topology.edgeuse_p3(9)
PARTIAL boolean.merge_p2 cases=45 passed=42 failed=0 deferred=3 max_rel_err=0e0 with_values=42 no_values=0 vacuous=0 blocked_by=curve.curve_p2(1),topology.face_p2(2)

• merge_p2 make_faceuse_square_to_make_face and make_faceuse_circle_region_given move from topology.edgeuse_p3 (GetOrCreateUVTrimCurve, read-only refusal) to topology.face_p2 SmFace::GetTolerance and curve.curve_p2 SmCurve::GetNaturalInterval. No other case of either unit changed.
• The other own_arena_geometry runners show no change on any case: query.brep, boolean.merge_p1, topology.brep_p2, boolean.merge_p5.

With #613 (shell ray fires create the missing trims), on main + #613:

FAIL topology.brep_p6 cases=595 passed=586 failed=2 deferred=7   (cube_faces_6, cube_faces_6_reversed)

and on main + #613 + this branch:

PARTIAL topology.brep_p6 cases=595 passed=588 failed=0 deferred=7 max_rel_err=0e0 with_values=588 no_values=0 vacuous=0 blocked_by=topology.edge(7)
PARTIAL boolean.merge_p2 cases=45 passed=42 failed=0 deferred=3 max_rel_err=0e0 with_values=42 no_values=0 vacuous=0 blocked_by=curve.curve_p2(1),topology.face_p2(2)

Checks (after merging main 7935b4c): cargo fmt --all --check; cargo clippy --workspace --all-targets -- -D warnings; cargo clippy -p difftest -p solidcore-ffi --features difftest/cpp --all-targets -- -D warnings; cargo test -p difftest --features cpp (258 passed); difftest --unit for the 7 units above. No library code changed.

How it was found: a gdb trace of the C++ reference (SmTopologySolver::MayContainAnswer, SmTLSPointEdge/Face::SolveIt, SmFace::PointClassify) next to a traced Rust run, while looking at the cases #541 moved to the read-only trim refusal.

🤖 Generated with Claude Code