v0.7.35

Fixed — a live run says what it is while it runs

Real runs were driven through epochix --live with the stream held open: a
CatBoost classifier, a CatBoost run whose AUC is 1.0 every round, and a Keras
run. CatBoost read correctly live — every round told as it arrived, and the
flat run's rows released after ten rounds rather than held for good — but the
dashboard was wrong in three ways.

• The learning curve drew a live run under the wrong metric. A run's record
  is created before its metric is known, holding val_loss and parser
  unknown, and only the end of the stream corrected it — which a live run
  never reaches. The curve read the record, so a live Keras run's validation
  accuracy (29% → 96%) was drawn as a falling "val loss" line for as long as
  it ran, and the run list called it CUSTOM. The record is now written as
  soon as the task, parser and story metric are known, and the curve takes
  the metric from the frames.
• "In plain English" called any rising number between 0 and 1 accuracy. A
  CatBoost AUC climbing from 0.979 to 0.992 read "97.9% to 99.2% accuracy"
  and "gets more than 9 in 10 right on data it hasn't seen"; so did AUC, F1,
  mAP50 and R² runs. It now goes by the metric's name: accuracy is told as
  accuracy, the "N in 10 right" meter appears only for validation accuracy,
  a loss that fell is "reduced", and any other metric is told by its own name
  without saying it improved.
• The epoch count was one short for logs that count from 0. Sixty
  CatBoost rounds read "over 59 epochs"; so did XGBoost and Lightning's
  Epoch 0: logs.

A CatBoost run whose AUC never moves is now a recorded fixture, pinned for both
engines.