The proposed primitive
Use bounded rational cells and a fixed number of packed messages per active node. Let connectivity evolve on a slower schedule than weight updates. The ambition is to keep the routine step regular enough for accelerator kernels while the stored graph gains capacity.
Growth has to earn its cost
A responsibility signal could propose splitting a cell when persistent residual error suggests a missing specialization. A function-preserving split would initialize the new capacity without abruptly changing the output. Routing, pruning, and validation would decide whether it stays.
The hard test
Measure wall-clock training, memory traffic, routing overhead, and generalization against dense and sparse baselines at equal active compute. Stored capacity and active compute are different budgets. This is a proposal, and we have not established dense-network training speed.