v0.7.33

Fixed — CatBoost with two eval sets

• A run with two eval sets was told on the wrong curve. CatBoost prints
  test: … test1: … best: …, and its running best and overfitting detector
  follow the last eval set. Epochix compared the best with the first, so
  a falling RMSE was called a score, and the run was told on its training
  loss and graded F, "past its best". Both sets also took one name, so the
  one CatBoost tracks was dropped. test1, test2 are now numbered eval
  sets: the last is the validation curve, the earlier ones are kept as
  eval0_…. Two more real CatBoost runs are recorded and pinned — two eval
  sets, and no eval set at all. (custom_metric was checked as well: CatBoost
  prints nothing extra for it.)

Internal — the extension's parsers are checked line by line

• The parsers exist twice, in Python and in the extension's TypeScript, and
  the ports had drifted and brought back fixed bugs before. Every line a
  Python test feeds a parser — every real log in the repository among them —
  is now recorded with what the parser returned, and the extension's test
  suite replays all of it through its own parsers: 3,876 calls, matching
  today. It caught the CatBoost fix above when only the Python half was in.