chore: bump OCCTSwift floor to 2.0.0, fix #168 and AAG face-index breaks

[gsdali]

What & why

OCCTSwift v2.0.0 (2026-08-09) is a correctness major: 17 breaking API changes (12 compile errors, 5 silent value changes), full table in docs/SEMVER.md#v200. This bumps the direct pin and fixes what the audit found.

#168 (already open, P1): Shape.centerOfMass returns nil instead of the bounding-box centre for anything enclosing no volume (OCCTSwift#605) — every vertex-anchor site now reads Shape.vertices() via a new SelectionTools.vertexPoint(_:) helper instead (SelectionTools, RemapTools, two CorrespondenceTools sites). TopologyIdentityTests' face/edge/vertex round-trip moved to surfaceInertia/edgeMidpoint/vertexPoint since centerOfMass no longer works as a generic identity probe.

AAG face-index scoping (#642/#699): AAG.detectPockets()/detectHoles()'s floorFaceIndex/wallFaceIndices/faceIndex are now occurrence indices into orientedFaces(), not faces() — only diverges on a body with a face shared between two solids (a boolean/pattern result). Three consumers fixed:

• GapFillerTools.selectByFeature: mintFaceSelection now indexes orientedFaces() directly — this was silently grabbing the wrong face on a shared-face compound, a confirmed real bug, not just a theoretical risk.
• AutoDimensionTool.autoDimension: converts to distinctFaceIndex before calling edgesInFace(at:), which expects a faces()-space index.
• AnalysisTools.recognizeFeatures/featureRecognize: extracted into a shared buildFeatureReport(shape:bodyId:kinds:) helper (was two independently-copy-pasted blocks) that converts every reported index to the stable distinctFaceIndex before returning it to the LLM caller.

New AAGFaceIndexTests pins all three against a multi-solid fixture (an off-center blind hole, split down the middle so the cut face is shared between the two halves).

#541 changed Shape.faces() itself to the deduplicated convention BRepGraph already used, so TopologyIdentityTests.multiShellSharedFaceDivergence (which relied on faces() staying occurrence-based) moved to the still-occurrence-based Shape.orientedFaces().

The remaining compile-error and silent-value-change symbols (PathParser, BisectorPoint, VinertGKResult, selfIntersectionCount, continuityOrder, ContinuityAnalysis, chamfer2D, buildCurves3d, drawMesh/evaluateGrid, nbEdges/nbFaces/nbVertices, and the mass-property surface bar centerOfMass) have zero call sites in this repo.

Update: all four sibling repos have since shipped their own 2.0.0-compatible releases (OCCTSwiftIO v1.7.7, OCCTSwiftMesh v1.7.4, OCCTSwiftTools v1.6.3, OCCTSwiftScripts v1.6.0), so this PR now also re-pins OCCTMCP's own transitive floors to match. swift build && swift test clean (213 tests, 37 suites) against both the local-sibling-checkout path and OCCTMCP_FORCE_REMOTE_DEPS=1 (the real published dependency graph) — this is ready to merge and release.

One known, non-blocking issue surfaced along the way and filed upstream: OCCTSwift#830, a likely kernel regression in Shape.pipeShell(solid: true) on a helix spine (~14% volume drift), found via OCCTSwiftScripts' own recipe smoke tests. Not consumed by this repo's own code.

Closes #168
Closes #171

Checklist

• New or changed behavior is covered by a unit test in the same PR — AAGFaceIndexTests (new) plus the updated TopologyIdentityTests.

Notes for the reviewer

• The AutoDimensionTool/GapFillerTools bugs this fixes are real but narrow: they only trigger on a body with a face shared between two solids (a boolean or pattern result), never on an ordinary single-solid part.
• OCCTSwiftViewport and OCCTSwiftAIS needed no re-pin: Viewport has no OCCTSwift dependency at all, and AIS floors OCCTSwiftTools (not OCCTSwift directly) at a version Tools' new release already satisfies.