[iancarscadden]

Refs #1054.

Problem

With enhanced-determinism, transcendental functions are meant to go through libm (simba/libm_force). That only happens for calls made through simba's traits. A method call such as x.sin() on a Real resolves to the inherent f32::sin whenever the std feature is on (the default, and what serde-serialize needs), and that calls the platform's math library. Platform libms round these functions differently, so results differ across platforms even with enhanced-determinism on.

Fix

Call each of these sites through simba::scalar::ComplexField / RealField, like the rest of the solver already does (simd_asin, ComplexField::cos, ...). Without enhanced-determinism they still resolve to the standard library, so nothing changes for other users.

Site	Function	Runs when
joint_constraint_helper.rs, AngularLimitParams::new	sin_cos	every joint with an angular limit (2D and 3D)
generic_joint_constraint_builder.rs, motor_angular_generic	sin ×2	position-target motor of an impulse joint attached to a multibody link
revolute_joint.rs, RevoluteJoint::angle	asin ×2	3D
soft_element_constraint_assembly.rs	exp, powf	soft bodies: deformation damping, cell size
soft_constraint_plasticity.rs	powf ×2	soft bodies: plasticity
soft_scalar_constraint.rs	acos	soft bodies: dihedral constraints (3D)
soft_body_shape_matching.rs	atan2	soft bodies: shape matching (2D)
soft_body_material.rs	exp	soft bodies: tear smoothing
soft_body_builder_shapes.rs	sin/cos ×2	soft-body tube (3D) and disk (2D) builders
soft_body_builder_mesh_helpers.rs	acos	soft-body dihedral rest angle (3D)

The list comes from a grep for every transcendental method call on a concrete float, cross-checked with clippy's disallowed-methods over f32::/f64::. What is left after this PR is powi(2) (one multiplication), Rot2::sin()/cos() (accessors), and test code.

Repro

The new platform_math_determinism tests (2D and 3D) step the joint paths and compare the result bit for bit with a golden. On master (cherry-pick only the two test files):

cargo test -p rapier3d --release --features enhanced-determinism --test platform_math_determinism
cargo test -p rapier3d --release --features enhanced-determinism --test platform_math_determinism --target wasm32-wasip1

	macOS arm64, master	wasm32-wasip1, master	every target, this PR
3D revolute_joint_angle	0xdd919012d0f49c14	0x384ef971c0fe5a21	0x384ef971c0fe5a21
3D motor_position_on_multibody_link	0x1c9bbfec0e740d88	0xec521711c406c841	0xec521711c406c841
2D motor_position_on_multibody_link	0xc067a7ae165a7432	0x4260ecae0c6ed6bd	0x4260ecae0c6ed6bd

angular_limits happens to agree between these two targets for the ranges I tried (Apple's sinf/cosf rarely differ from libm for small arguments); it is there as a golden for the other platforms. #1054's repro repository shows that path diverging on macOS in 2D.

Testing

• New tests added to the tests-determinism, tests-determinism-arm64 and wasm-determinism jobs and to run-ci-checks.sh.
• Checked locally with this PR: the goldens hold on macOS arm64, Linux arm64, Linux x86_64 (emulated) and wasm32-wasip1 (Node 22).
• cargo fmt -- --check, cargo clippy with -D warnings, cargo test -p rapier2d -p rapier2d-f64 -p rapier3d -p rapier3d-f64 (799 passed), cargo check --features enhanced-determinism for both crates, the no_std checks on the host target. The existing snapshot_portability (native and wasm32) and simd_backend_determinism goldens are unchanged.

Not in this PR

• The same pattern exists in parry (Bvh optimization log2, VectorExt::angle, ConvexPolyhedron sin_cos, mesh intersection): fix: make enhanced-determinism cover the math calls that go to the platform's libm under std parry#468.
• and in glamx: SymmetricEigen3::eigenvalues (SymmetricEigen3::eigenvalues calls platform acos/cos in std builds, breaking enhanced-determinism glamx#9, fix in Route eigenvalue trigonometry through ComplexField under std glamx#10) and Svd2 (fix(svd2): route trigonometry through simba so the libm feature applies under std glamx#12). The eigenvalue one is what makes 3D bodies with a non-diagonal inertia tensor (compound or rotated colliders) diverge between native and wasm32.
• The multibody IK case from Rapier (& Parry) use of platform floating point fns breaks determinism in Rapier 0.36.1 #1054 is a different cause (not a libm call) and is untouched.

For context: with these three fixes applied, 600-tick 2D and 3D scenes in our engine give identical state hashes and byte-identical snapshots on macOS arm64, Linux arm64 and x86_64, Android arm64, wasm32-wasip1 and the iOS Simulator, and the test binaries import no platform math symbols.