One aggregated round: Python client, Rust round server Diagram

One round

The client registers once, then polls; only the sampled clients are told to train
Training stays in PyTorch, proof generation runs in the Rust core inside the client
The server closes the round itself when every assigned client has submitted

What the server checks

State hash, delta hash, learnable-only norm bound and the Groth16 delta-bound proof
A rejected client is dropped with a typed reason; the round still needs one survivor
The returned model carries the aggregation proof, so the client can verify it

When it stalls

`--round-timeout-secs` closes a stalled round with the updates that arrived
mTLS makes the certificate CN the client identity; a shared token only gates the deployment