Unity’s built-in 2D physics is an integration of the Box2D engine [12]. We build the same kind of2D game runtime from a different direction. Game scripts are written in a subset of C#, translated to a subset of C++ and then to C, and compiled together with a fork of Box2D v3 that we call Box2D-Packed. Our goal is a 2D runtime that is faster than Unity’s. This paper describes the two techniques we use to get there and measures each of them. The first is data packing: public handles shrink from 8 to 4 bytes, and the records the engine reads on every step are laid out by measured access patterns. Measured with Cachegrind’s exact cache simulation, last-level data cache misses fall by up to 65% on query-heavy scenes relative to Box2D v3, with unchanged instruction counts and bit-identical simulation results. The second is an intrusive engine: game logic is compiled into the physics engine at documented points, so it runswhile the engine’s data is still in cache, instead of through callbacks and event arrays. Injected pre-solve logic saves 1.3% of instructions, and injecting a platformer’s event handlers cuts L1 data misses by 2%, with identical game results. We also report where injection does not help: a per-body gravity hook that reads game data costs 7.65% more L1 misses, and event handlers followed by game work on the same objects add last-level misses. We compare against Box2D v3, not yet against Unity, and describe the measurements that remain.