Add a FedAvg fraud-detection application in two topologies: provider
mode (coordinator/trainer star) and P2P mode (full mesh with CRDT-gated
rounds), backed by an MLP trained on PaySim data with a synthetic
fallback, plus demo scripts for both modes.
Also harden the actor runtime:
- supervisor: ignore stale death messages (OneForAll restart storm),
drop normally-exited children instead of respawning, add a
restart-rate limit
- register actors before their thread starts so UUID dispatch and
getSelf cannot race the spawn
- killActor now kills the thread so links see Killed
- fulfill connection promises with the pool winner on simultaneous
connects
- clamp the PaySim test split on small datasets and split shared
transaction types across trainers in non-IID partitioning
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
dvdrw