This is so cool, I didn't expect someone to figure it out so quick!
Quick note: there is currently an upper limit on vertices and indices inside the clip-space of the window, but it may be increased at some point.
Each sprite is 4 vertices and 6 indices.
The _r.layer(0) call hints at one of the reasons why the performance is so good.
The arcade renderer will batch similar objects into a single draw call.
So because everything is on the same layer, all squares are sent to the GPU in a single batch to be rendered.
This only works if all objects are on the same layer and have the same texture / blending mode.