Improve Assembly collision simulation responsiveness

[10-X-eng]

Closes #163.

User-visible outcome

• Runs dynamic collision checks across independent solver frames in parallel while preserving deterministic frame ordering and authenticated collision evidence.
• Gives each worker its own deep OCCT topology and triangulation copy, preventing unsafe concurrent mutation/read access.
• Uses at most 75% of logical CPUs and balances concurrency against the cost of cloning the full collision model.
• Reports collision frame progress, percentage, measured rate, and approximate remaining time through the existing worker progress path and VibeCAD status bar.
• Adds the optional collision_mode=off Assembly simulation mode for kinematic playback. Existing callers continue to default to full collision analysis.
• Represents skipped analysis explicitly as not_checked; it can never be reported as collision-free.
• Uses a positive OCCT proximity tolerance because FreeCAD does not configure OCCT proximity when passed zero.

Performance evidence

Production-sized Windows workload: 27 components and 141 solver-output frames.

• Previous serial collision phase: 232.469 s
• New 12-worker collision phase: 35.516 s
• Collision speedup: approximately 6.5x
• Serial and parallel collision frame records were byte-for-byte equivalent after deterministic ordering.
• Trace digest: 731f02277ea814d31f3a59777c8e6d4f5ae37d558dd97468df9ff8b510846f48

The complete clean portable Windows build was launched, exercised by the owner, and confirmed to fix the reported behavior.

Red/green TDD

The new tests were added first and failed on the absent frame-worker API, collision policy, progress metrics, skipped-analysis contract, and status rendering. After implementation:

$env:PYTHONPATH='C:\Users\robit\vibecad-issue-163\src\Mod\VibeCAD;C:\Users\robit\vibecad-issue-163\package\rattler-build\.pixi\envs\default\Lib\site-packages'
C:\Users\robit\vibecad\build\pr-test-venv\Scripts\python.exe -m pytest -q src/Mod/VibeCAD/vibecad_tests/test_mechanism_geometry_parallelism.py src/Mod/VibeCAD/vibecad_tests/test_mechanism_engine.py src/Mod/VibeCAD/vibecad_tests/test_engine_contracts.py src/Mod/VibeCAD/vibecad_tests/test_modeling_surface_architecture.py src/Mod/VibeCAD/vibecad_tests/test_vibescript_file_io.py

Result: 228 passed in 6.85 s.

Additional verification completed during development:

• Mechanism/geometry focused suite: 28 passed.
• Engine/file-I/O focused suite: 44 passed.
• Modeling architecture suite: 156 passed.
• Provider subprocess suite: 51 passed, 1 skipped.
• Full Python suite: 4,354 passed, 9 skipped; the initial isolated-worktree run had one missing-Fasteners-license failure caused solely by the submodule not being initialized. After initializing the pinned Fasteners submodule, its packaging contract passed 6/6.
• Native FreeCAD integration exercised actual overlapping solids, imported STEP geometry, strict containment, deep-copy non-partnership, and both full/off isolated Assembly workers.
• py_compile passed for every modified runtime and test module.
• git diff --check passed.

Clean Windows package verification

pixi install -e default --frozen
pixi install -e package --frozen
$env:BUILD_TAG='v26.3.1-RC6-build1'
$env:MAKE_INSTALLER='false'
$env:MAKE_PORTABLE_ARCHIVE='true'
pixi run -e package create_bundle

Result: complete Rattler native compile, portable-bundle startup/dependency smoke tests, provider subprocess smoke, Codex app-server smoke, geometry worker smoke, McMaster WebView2 smoke, and windowless-provider smoke all passed.

Portable archive SHA-256: 1eec8503bb9ae345aadb4d80c68d68b360960a549dc2d1d13102499ac569a032.

Compatibility

• Existing public functions/APIs remain present and behaviorally compatible.
• No preference keys, tool names, or schema fields were renamed or removed.
• Existing simulations still perform full collision analysis by default.
• New arguments and progress fields are additive and optional.
• Skipped collision analysis is explicitly non-authoritative.
• No breaking changes.

Risk and containment

Frame-parallel collision checking increases temporary memory use because safety requires one deep shape/mesh copy per worker. Worker selection is bounded by frame count, 75% of logical CPUs, and the square root of frame count to balance setup cost and throughput. Pair-level parallelism is disabled inside frame workers to avoid nested oversubscription.