Findings from user field testing (Fedora host + Windows worker):

- workers previously loaded the dataset from their own disk by the
  path in the job file — nothing exists there on another machine. The
  coordinator now ships local dataset files (up to 50MB) inside
  TRAIN_START; workers write them to a temp dir and train from that.
  HF dataset ids still download per-node. Workers need zero files.
- worker setup looked like a hang (nothing printed between model load
  and the first round on slow CPUs). Workers now report each stage:
  dataset received, preparing model, round N: training K steps
- setup failures print on the worker itself, not only at the host
- single-machine 'onepool train' now serves the same live dashboard
  as pool runs (loss curve at localhost:7070)
- TESTING.md updated for calibration rounds and the new output

Verified: worker joined from a directory with no dataset present and
trained 3 rounds to loss 4.71; solo run with dashboard clean.