Scaling tangential velocity by a fixed factor per contact bleeds a rolling body to a stop

Fact

The obvious way to add friction to an impulse-based collision resolver is to shrink the tangential part of the relative velocity by a constant:

rv.t *= (1 - mu);

It is wrong for resting and rolling contact, because a body lying on a surface is re-resolved every substep. At a 240 Hz step, mu = 0.06 is not a 6% loss; it is 6% two hundred and forty times a second. A ball rolls down a 25° ramp at a crawl, and lowering mu only postpones the stall.

Coulomb friction has no such term. The tangential impulse is bounded by the normal one, and a resting contact carries only the substep's worth of gravity:

const jn = -(1 + e) * vn;            // normal impulse
const jt = Math.min(tangentSpeed, mu * jn);

So the same coefficient that grips hard on a fast impact costs a rolling body almost nothing.

Why it matters

The symptom is behavioural, not numerical: nothing is NaN, energy is not diverging, and every collision is individually plausible. Bodies simply refuse to travel along shallow surfaces, which reads as bad level geometry — a ramp that is "not steep enough" — and gets fixed by tilting the world instead of the model.

It also scales with the timestep, so raising the substep rate to fix a tunnelling problem makes the friction worse, and the two changes look unrelated.

How to apply

• Cap the tangential impulse by the normal impulse; never scale the tangential velocity by a bare coefficient.
• Put deliberate, small energy loss somewhere frame-rate independent instead — a drag term applied per unit time, v *= 1 - k * dt.
• Test it on the shallowest slope in the world, from rest. A body that creeps where it should roll is this bug, not the slope.