Scale FMoW to 200 Satellites

[Nova-David]

Scaling to 200 satellites

• Fixed chunk transport to tolerate out-of-order delivery under concurrency, added a 30s stale-transfer timeout, doubled chunk size to 4MB
• Fixed RandomSelector never resetting selected_ends/all_selected between rounds which made agg_goal unreachable and triggered false SIM_STARVATION after the first round
• Reduced per-round payload size: frozen-layer parameters excluded from delta_weights, empty-dataset trainers now send a status-only message instead of a full payload, added send jitter to avoid synchronized traffic bursts

Evaluation

• Final round now always evaluates (previously skippable), runs synchronously so the run doesn't exit before it finishes

Guards ahead of a dg-fork-main merge

• eval_every_n_commits set 1 to neutralize a new change that would otherwise silently drop ~half of scheduled evaluations
• heartbeats config access now uses getattr with a default, since the field is being removed from the shared schema upstream