"""Training fleet — multi-host distributed training coordinator.

Mirrors the collection-side fleet pattern:

  - Single canonical training_manifest.toml (operator-edited)
  - Workers self-detect capability + report to the receiver
  - Receiver maintains a SQLite job queue, atomic claim + heartbeat
  - Workers loop: claim → train → ship artifact → repeat
  - Operator controls deployment via the manifest only

The collection fleet is embarrassingly parallel (every host runs the
same plan). Training jobs must be assigned at most once across the
fleet, so the receiver coordinates claims; everything else is symmetric.
"""