[zorak1103]

Machine-generated (Claude Sonnet 5.5 agent) on behalf of @zorak1103.

What

boolean.merge_p1 reaches 0 deferred and is promoted to green. Its 4 deferred cases all stopped at the UNPORTED(boolean.topology_intersector_p3) stub for SmTopologyIntersector::MergeCoincidentSurfaces, called from SmMerge::FixFoundFaceCoincidence. The algorithm was already ported (TopologyIntersector::merge_coincident_surfaces, PR #211); the follow-up its notes left open was an IntersectorBrepP3 for merge p4's ArenaPair. This PR does that and wires the call.

• boolean/merge/p1.rs: fix_found_face_coincidence is split into check_face_coincidence (the all-edges-mated return) and finish_face_coincidence (rbModifiedTopology from the edge counts). The merge between them needs &mut access to the adapter's own mate store, which the old signature held borrowed.
• p4/arena_pair.rs: ArenaPair implements IntersectorBrepP3 (modelled on the tested difftest Two) with loop-use and face-use handle kinds, and, since Port MergeInterval imprinting (intersect.curve_class_p6) #721 made MergeIntervalBrep a supertrait of IntersectorBrepP3, MergeIntervalBrep and ClassGeometryMut. merge_context hands out the operand's arena, the entity, the handle mapping and the pair's notification records, attributes and context; the identity callbacks it borrows are held by the pair. A curve identity that caller-supplied keys refuse cannot pass through the context's infallible callback, so the first refusal is kept and the merge fails with it (no rollback of the edit in progress; this error path has no unit test). The private Objs classification wrapper now delegates classification_vertex_point, topology_zone_tol_3d, topology_entity and topology_handle_in; their absence surfaced as an SmVertex::GetTolerance deferral inside the merge. vertex_face_gap keeps the trait default, as the pair does.
• p4/kernel_adapter.rs, p4.rs: ManifoldKernelAdapter::fix_face_coincidence runs the merge on an ArenaPair with the retained mate store; the ManifoldIntersector::fix_face_coincidence trait method gained kernels and attributes (threaded through classify_edge_faces).
• Plan: plan/units/boolean.merge_p1.toml to green and an entry in plan/findings.md. (boolean.topology_intersector_p3.toml is not touched: that unit's numbers do not change, and its one-line notes field conflicts with nearly every PR in this area; its "stays UNPORTED ... follow-up" back-edge sentence is stale once this lands.)

Harness changes

Op 2 previously cloned and discarded the other B-rep and compared only the primary's graph plus the face lists, so the merge's edits of the second B-rep were invisible. Now:

• the C++ shim (harness/cpp/shim/boolean_merge_p1.cpp) appends for op 2 the returned faces numbered by walks after the call, both directions of m_vTI's edge mates, and the other B-rep's raw export (length-prefixed);
• the runner drives the real adapter with the context-only SmMerge tolerances (SM_EFF_ZERO_SQ, 30 degrees) and compares other_graph, other_geometry and the mates; the trailer parser rejects short or trailing data;
• the generator gives op 2 another rectangle of the face's plane as other (overlapping, contained, enclosing, touching, sharing edges, and disjoint).

Results

• difftest --unit boolean.merge_p1: PASS cases=248 passed=248 failed=0 deferred=0 with_values=236 (main before this PR: 244 passed, 4 deferred).
• --seeds 6000: PASS cases=6152 passed=6152 failed=0 deferred=0.
• Full gate on the tree before the last merge of main (7292bda1d): cargo fmt --check; clippy --workspace --all-targets -D warnings with and without difftest/cpp; cargo test -p solidcore (2096) and the difftest-crate tests (273); cargo check --target wasm32-unknown-unknown; --green: 260 units, 213 passed, 47 partial, 0 failed.
• After the later merges of main (Reach the owning faces' caches in FaceIntersect, CreateParallelProjectionCurves and sm_FindSectorOfFace #723, port: proto face updates on real owning topology #726, difftest: boolean.api_brep builds real box/sphere operands #730, Classify on the owning face in the coincidence checks, boundary start points, OrientTrimmedSurfaces and MergeCoincidentSurfaces' drop #734 and others; head 95c30590a): fmt, clippy both ways, cargo test -p solidcore --lib (2099), wasm, and difftest of boolean.merge_p1 (248/248; 6152/6152 at 6000 seeds) through _p5, boolean.api_brep and boolean.topology_intersector_p1 to _p3: 0 failed, deferral counts as on main. The full --green run was not repeated after that last merge (it was stopped twice for memory pressure); CI will run it.
• In an all-units run before the first merge of main, no unit's numbers other than boolean.merge_p1 changed, including the other users of the path (boolean.merge_p4 311/140/0/168, boolean.api_brep 287/151/0/136).
• New unit test unmated_faces_reach_the_coincident_surface_merge (the path used to return Unported).

Notes for review

• What the new variants check. With port: topology.face_p4 (stacked on #693, merged with main) #718 and Port MergeInterval imprinting (intersect.curve_class_p6) #721 on main all rectangle variants pass. The touching and disjoint rectangles compare the two B-reps. For the overlapping, contained, enclosing and edge-sharing ones the merge returns an error status that C++ and Rust agree on (both_errored; no_values rises from 12 to 160 at --seeds 6000 when one such variant is generated alone), so they check the status, not the topology. No generated case has a merge that succeeds and splits faces, so the post-call face numbering and the edge-mate comparison are written for that case but not exercised by a splitting merge yet.
• Not compared: notification records and attributes (op 2 uses fresh ones) and the geometry of curves the merge creates. The merge's temporary curves stay in the adapter's curve stores (handles are stable; memory only).
• Possible textual conflicts with open PRs that also touch arena_pair.rs (wire: topology-faces (3 target units) #334), kernel_adapter.rs (wire: topology-edge-brep (7 target units) #335) or append to plan/findings.md (wire: topology-faces (3 target units) #334).
• An independent review of the diff was done; its findings (face numbering after a split, uncompared mates, trailing-data check, operand checks, stale comments) are addressed in this PR.