Based on feedback from another professional game company, there are several improvements, including some API change...

Some dynamic memory allocations have been replace by pool allocation or stack allocations.
quantized aabb versus quantized aabb overlap check is made branch-free (helps a lot on consoles PS3/XBox 360)
Collision algorithms are now created through a new btDefaultCollisionConfiguration, to decouple dependency (this is the API change):
Example:
	btDefaultCollisionConfiguration* collisionConfiguration = new btDefaultCollisionConfiguration();
	m_dispatcher = new	btCollisionDispatcher(collisionConfiguration);

src/BulletCollision/CollisionDispatch/btCollisionWorld.h

@@ -98,7 +98,7 @@ protected:
 public:
 
 	//this constructor doesn't own the dispatcher and paircache/broadphase
-	btCollisionWorld(btDispatcher* dispatcher,btBroadphaseInterface* broadphasePairCache, int stackSize = 2*1024*1024);
+	btCollisionWorld(btDispatcher* dispatcher,btBroadphaseInterface* broadphasePairCache, int stackSize = 10*1024*1024);
 
 	virtual ~btCollisionWorld();