Two delay lines are cross-coupled: the left one's output feeds both the left output channel and the right line's input, and the right one's does the mirror image. A single impulse therefore emerges on the left after one delay time, on the right after two, on the left after three, and so on. Nothing counts repeats or schedules sides, the alternation is a property of the loop, which is why it stays correct at any delay time and needs no re-scheduling when the time knob moves.
The input is summed to mono before it enters the loop, and that is deliberate rather than a limitation. If a stereo source kept its channels, both lines would be seeded at once and the first repeat would arrive on both sides simultaneously, which is precisely the bounce not happening. The dry path keeps the source's own image; only the repeats are mono-seeded.
Feedback is applied at each crossing, so it multiplies twice per full cycle back to the starting side. That is the practical difference from a plain delay's feedback knob: the same number gives a shorter tail here, and the cap the two effects share is measured against the per-crossing gain, which makes it conservative for the round trip rather than marginal.
This is the browser preview: two DelayNodes, two gains and a channel merger, structured after the reference ping-pong delay rather than compiled from the same kernel, and making no parity claim against it. There is no tempo sync and no filtering in the feedback path, so the repeats stay as bright as the source instead of darkening the way an analogue-voiced delay would.
