proto: experiment execution & analysis

## Goal
Run the experiments and compare the measurements with the model predictions — the data that settles the topology decision.

## Outcome
Per-stage findings (configs + seeds published) versus the laws of #76, feeding the recommendation (#85) and the architecture document. The runs also record what the nodes consume (CPU, memory, bandwidth), giving the node-requirements input for the architecture — absorbed from the former resource-model task (#89).

## Needed
- [ ] Stage runs: honest baseline → silent adversaries → bounded/bucketed layers
- [ ] Analysis against the model laws
- [ ] Resource observations per node alongside the metrics
- [ ] Findings write-up

## References
Experiment program: PR #74