There are a few minor breaking changes, see this document.

New functionality

The Samples application now runs on iOS and Android. See Build/README.md for more information about the controls.
Added support for 64-bit ObjectLayer. This e.g. allows for more groups in case you're using ObjectLayerPairFilterMask.
Added ModifyShapes overload to MutableCompoundShape. Adds the ability to change position, rotation and shape at the same time.
EstimateCollisionResponse now takes surface velocities into account.

Bug fixes

MSVC Build Tools 14.52 in combination with /fp:fast introduced some extra optimizations that were incompatible with Jolt. This caused failures in Quat multiplications and FlipSign. This completely broke the simulation.
MinGW32 defines __STDCPP_DEFAULT_NEW_ALIGNMENT__ as 16 but its allocations are actually 8 byte aligned. This caused unaligned read access violations in TriangleSplitterBinning::mBins.

Avoiding overflow to INF if inverse effective mass is denormal. When restricting the rotation of an object using EAllowedDOFs, a tiny rotation in a restricted axis can cause the effective mass of the angular friction component to overflow to INF and cause a FP exception.
Clamping warm start impulse ratio to avoid the simulation from exploding when updating the physics system with wildly different delta times. Note that this is not a perfect solution, so it is still recommended to step physics at a fixed time step.
Fix out of bounds read for a body activated during the continous collision detection step.
VehicleEngineSettings::mInertia and mAngularDamping were not serialized in an ObjectStream or from SaveBinaryState/RestoreBinaryState.
Fixed an issue where a body with motion quality LinearCast could pass through another body if its center of mass was offset.
Fixed an issue where GearConstraint would apply the wrong torque to body 2.
CharacterVirtual::WalkStairs and CharacterVirtual::StickToFloor could trigger an assert when stepping onto another virtual character.

If GJKClosestPoint::CastShape failed to converge, it was accidentally using the rejected support point to calculate the contact point. In some cases this could lead to returning the wrong contact point.
When colliding with very long and thin triangles, GJK could fail to find the closest point to the triangle because it used a fixed epsilon to determine if the triangle was too thin. Changed this epsilon so that it is dependent on the length of the edges of the triangle so that we fall back to closest point to edge in those cases.
Handling of convex/sphere collision/cast checks against degenerate triangles. A MeshShape attached to a Body with a certain transform could result in an almost degenerate triangle becoming degenerate. The code now handles this.
HeightFieldShape::SetHeights corrupted the bounding volume tree when the number of blocks was not a power of 2. This resulted in collision queries possibly failing to report the correct collision.
Fixed HeightFieldShape::SetHeights not updating the last range block of a level with an odd number of blocks. This resulted in collision queries possibly failing to report the correct collision.
ConvexHullShape::GetSupportingFace could return a face that did not contain the deepest vertex, resulting in objects being able to penetrate more than they should.

Fixed jitter in hair simulation on macOS/iOS. The shader compiler shuffled two instructions that it shouldn't have.
ComputeSystemVK now falls back to non-cached host memory on GPUs that don't have memory that is both host coherent and host cached (e.g. Mali).
ComputeSystemVK no longer requires the VK_EXT_debug_utils instance extension to be present.

Changed the queue in JobSystemThreadPool so that a worker thread that doesn't wake up cannot prevent new jobs from being added to the queue.
Fixed macOS thread performance checker warnings. Note that this requires suppressing 1 warning on the application side. See Build/README.md.
Semaphore::Acquire ignored sem_wait's EINTR. JobSystemWithBarrier could finish a wait early on Linux.
Fixed a race condition in JobSystemWithBarrier where a job that was added to a barrier could finish before the barrier counted it. When this happened, WaitForJobs could return too early resulting in memory corruption and a hang.
Suppressing clang thread safety analysis warnings which appeared under XCode 27.

Removed PhysicsSettings::mDeterministicSimulation. Turning this off has negative impact on stacking. The simulation is now always deterministic.
Removed support for using std::vector instead of JPH::Array (define JPH_USE_STD_VECTOR). This is a pretty obscure build option and not worth the continous testing needed to support it.