neural_ca: state the universality claim precisely. The CA realizes and verifies the billiard-ball mechanics (ballistic transport, deflecting reversible collisions); universality is Fredkin and Toffoli's result for that model and the gate layout is their construction, not built here.

README.md

@@ -689,14 +689,14 @@ tile that is itself a permutation with the same 0.5 analog margin as
 so `variants/neural_ca.safetensors` stores the rule of the medium, not a
 processor.
 
-Its dynamics are those of the billiard-ball model of computation (Fredkin and
-Toffoli): isolated particles travel ballistically along diagonals, and
-collisions deflect them reversibly, both verified here. Particle configurations
-are therefore signals and collision geometries are logic; a Margolus
-billiard-ball automaton of this class is Turing-universal, so a piece of this
-homogeneous reversible medium computes by evolving. It is the family's maximal
-dissolution of architecture: the same rule everywhere, no center, computation as
-discrete reversible physics.
+Its dynamics are those of the billiard-ball model of computation: isolated
+particles travel ballistically along diagonals, and collisions deflect them
+reversibly, both verified here. Those are exactly the mechanics Fredkin and
+Toffoli showed are computation-universal, with ballistic particles as signals,
+fixed clusters as reflectors, and collision geometries realizing the Fredkin
+gate. This medium realizes and verifies the mechanics; the gate layout is their
+construction. It is the family's maximal dissolution of architecture: the same
+rule everywhere, no center, computation as discrete reversible physics.
 
 ```bash
 python src/ca.py             # rule bijection, lattice reversibility, ballistic motion, collisions