A boundary clamp applied during integration must run again after collision resolution

Fact

The usual step order is integrate, clamp to the world bounds, then resolve collisions with other bodies. Positional collision resolution teleports the moving object out along the contact normal — and that displacement does not know about the world bounds the clamp just enforced.

So a body pressed against a wall pushes whatever it collides with through that wall. The object ends the step embedded in the boundary, past a clamp that ran correctly a few lines earlier.

It only happens within one object-radius of an edge, and only while something is holding position there, so it survives ordinary play and ordinary tests.

Why it matters

The clamp is present, is correct, and is the first thing read when the bug is reported — which makes the boundary code look exonerated. Attention then goes to the collision resolver, which is also correct. The defect is in neither: it is in the order.

Visually it is a sliver of an object outside the play area, easy to dismiss as a rendering or camera artifact rather than a state error, so it tends to be found by an invariant rather than by looking.

How to apply

• Re-apply the clamp after every step that can move the body: resolve collisions, then clamp again. It is idempotent, so running it twice costs nothing.
• Reflect velocity on the second pass only when the body is still moving into the boundary, or a rebound gets applied twice.
• Assert the invariant every step over a long automated run rather than trusting the clamp: assert(r <= x <= W - r) finds this in seconds and pins it against later reordering.