A perfectly regular packing passes an impulse straight through instead of scattering

Fact

Resolve collisions one pair at a time — the normal way — and a geometrically exact packing behaves like a Newton's cradle. The impulse walks a single line of centres and leaves at the far body; everything in between ends up back at rest.

A racked triangle of billiard balls struck square is the clearest case: 14 of the 15 finish within a few millimetres of where they started, while one crosses the table. Every individual collision is correct, and the aggregate is nothing like the real thing.

Why it matters

It reads as an energy bug. The suspects that present themselves — restitution, the impulse formula, the timestep, tunnelling — are all fine, and instrumenting total momentum or the sum of speeds shows nothing wrong, because nothing is wrong locally. The cause is symmetry: contacts that are exactly aligned, resolved sequentially, never get a sideways component to split momentum with.

Real packings are never exact and real contacts are never purely normal, so the effect has no counterpart to compare against outside the simulation.

How to apply

• Perturb the initial state: a millimetre of jitter per body, comparable to the clearance between neighbours. Less than the clearance changes nothing.
• Model the off-normal part of a real contact. A tangential kick of a few percent of the normal impulse, in a random direction, is enough and is physical — friction between two surfaces does exactly this.
• Keep both deterministic. Draw from a stream seeded per event, not from the ambient generator, or the same inputs stop replaying to the same outcome.
• Diagnose by counting how many bodies moved, not how much energy remains. One fast body and a still crowd is the signature; a spread is what fixes look like.