v0.7.32

Fixed — CatBoost, checked on a real run for the first time

CatBoost was listed as supported, but no real CatBoost output had ever been
read. Two runs are now recorded (tests/fixtures/logs/boosting_real_source.py)
and pinned for both engines; reading them found two faults.

• An AUC was told as a loss. CatBoost prints test: 0.9790356 with the
  metric's name nowhere on the line, and every such value was called a loss. A
  classifier whose AUC climbed from 0.979 to 0.992 was told "past its best".
  The direction comes from CatBoost's own running best, printed on every row:
  a best that follows the maximum makes the value a score, one that
  follows the minimum a loss. Rows are held until a row shows which;
  a metric that never moves is told as metric, claiming neither, and a
  live run is never held for more than ten rounds.
• Its timing was charted as a metric. total: 137ms remaining: 8.11s
  became a series called "ms remaining" on every row. A metric name can no
  longer start inside a word or a number.
• CatBoost's closing bestTest = / bestIteration = lines are claimed, so no
  other parser reads them as a new measurement.
• score, val_score and train_score are pinned higher-is-better
  (scikit-learn's convention). Unpinned, they took a CUSTOM run's
  lower-is-better default.
• VS Code: a parser that returns rows it held back together with the
  current one now gets a frame for each round; the extension had folded them
  into one reading.

Changed — a first look with nothing loaded

• The empty panel says what Epochix reads: what a training run prints,
  the loss and accuracy at each epoch, not your source code. Both reports that
  reached the tracker from outside in the first month were filed from an empty
  panel, and one expected Epochix to read their code.
• The report button with nothing loaded asks first. It offers the demo and
  opening a log file (in VS Code) or how to load a run from a terminal, before
  a report. A report filed anyway asks what you were trying to do and what you
  expected to see, rather than "what looks wrong" with nothing on screen. In
  English, French and Persian.

Dependencies

• Development tooling updates for the frontend and the extension (#106, #107,
  #108, #111, #112).