CS2 grenade bounce: exactly 0.45× on every surface

How it was measured

Bounce is the mechanic where a plausible-looking measurement is easiest to get wrong. Old demos update entities only every few ticks, so the tick a grenade actually touched a wall is rarely sampled, and the speed ratio you compute from the nearest samples is biased by however much flight got folded into the pair. An earlier pass on coarse-cadence data produced a per-surface reading of about 0.53 for concrete blocks, and it was recorded as UNRESOLVED rather than shipped.

On current-build demos, which update every tick, the same estimator run over 983 bounces converges: 0.4500 on the normal component and 0.4517 on the tangential. Two components agreeing to within the noise is the actual finding — it says CS2 scales the whole reflected vector by one number rather than treating friction and restitution separately.

Then the exact response law was transcribed from the game binary and replayed against 90 controlled captures: the sweep stops 1/32 u short of contact, the backoff is 2·dot plus that same 1/32, and the ground branches are separate. Replay is tick-exact — median residual zero, bounce counts and surface normals matching case by case. That is what retires the 0.53: the resolver applies a flat 0.45 with no surface term, and the flat law replays every surface the capture matrix touches, so the old figure was coarse-cadence noise rather than a material property.

The rest transition was the last piece. A grenade does not roll to a stop by friction; a collision on a qualifying normal whose post-clip speed is below 20 u/s zeroes the velocity and switches the entity to MOVETYPE_NONE on that same tick. All 85 recorded transitions replay on the exact tick, and walls never rest — 55 low-speed wall bounces kept moving, which is what pins the normal to a ground-facing one.