Bullet physics engine

[sstarosz]

Summary

Carves the Maya-free Bullet physics engine out of the oversized #3 into its own reviewable PR (part of the agreed 6-PR split). Everything here is pure C++17 + Bullet — no Maya SDK dependency — proven by a dedicated test target that links only Bullet + Catch2.

What's included

Core engine (mmd/core/)

• common.hpp — Double3 / Double4 / Matrix4 value types (constexpr, index access, data() for the Maya API boundary) + the documented precision model.
• physics_math.hpp — euler↔quat conversions (gimbal-lock aware), rest/anchor transform helpers, Bullet↔core matrix conversions.
• simulation.hpp / simulation.cpp — Simulation class wrapping a Bullet btDiscreteDynamicsWorld: body/joint definitions, kinematic anchors, dynamic reset, fixed 60 Hz stepping. Rule of five, e-prefixed enums, Doxygen, std::optional members (no heap hand-rolling).
• CMakeLists.txt — mmd_core static library.

Unit tests (tests/core/)

• test_physics_math.cpp, test_simulation.cpp — 18 Catch2 cases (rest pose, gravity, kinematic weld, reset, disabled bodies, euler round-trips).
• mmd_core_tests target links only Bullet + Catch2 — passing proves the engine is Maya-free.
• tests/core/.clang-tidy — relaxed profile for tests (Catch2 macro expansion noise).

Wiring

• CMakeLists.txt — new BUILD_TESTS option (off by default) + add_subdirectory(mmd/core).
• tests/CMakeLists.txt — tests/core subdirectory under BUILD_TESTS.
• .github/workflows/pr-checks.yml — configure with -DBUILD_TESTS=ON; new Run C++ unit tests step (ctest --preset default -E maya-integration).
• .clang-tidy — project-wide profile (bugprone / clang-analyzer / cppcoreguidelines / misc / performance / portability / readability, with narrow justified exclusions).
• docs/CPPDevelopment.md, CHANGELOG.md — updated.

Precision model (deliberate)

Public value types are double (Maya is double throughout); the Bullet world is float — MikuMikuDance itself runs Bullet in float, so float arithmetic is the fidelity reference. Conversions happen only at the core/Bullet boundary via explicit static_cast<btScalar> (no double-precision vcpkg feature).

Verification

• Build: clean (MSVC, Release).
• Tests: 18/18 pass (ctest --preset default -E maya-integration).
• clang-tidy (LLVM 22.1.8): 0 findings across simulation.cpp + both test files.
• git diff --check: clean.

CI will additionally compile the core on Linux/Clang and run the same unit tests in the C++ job.

[sstarosz]

Collision mask semantics ? worth pinning down before the adapter lands

b.mask is documented as "PMX non_collision_group, verbatim" and passed straight through as Bullet's filterMask. In PMX, bit i of non_collision_group set means "does NOT collide with group i" ? the opposite of Bullet's filterMask bit meaning ("collides with group i"). MMD passes ~non_collision_group as the Bullet mask, and blender_mmd_tools inverts in its collision_group_mask conversion too.

Either invert here (mask = ~non_collision_group) or document BodyDefinition::mask as a pre-inverted Bullet filter mask and drop the "verbatim" wording. If the adapter fills it verbatim, every collision-filtered pair behaves backwards. This path is also untested ? a collision-filter test would encode whichever contract you pick.

[sstarosz]

Group-15 edge case (latent)

1L << 15 = 32768 overflows the short that addRigidBody / addCollisionObject take (PMX group ids are 0?15, so 15 is valid) ? implementation-defined negative filter group (typically -32768), which makes that body's collision filtering wrong. MMD shares this limitation, so it may be accepted ? but a comment or guard would prevent a future "fix" from breaking it, and an explicit static_cast<short> on both group and mask would make the narrowing intent clear.