v0.7.16

Added — a comparison you can send

• The comparison view's explanation could not leave the browser. The race
  GIF carried the curves; the paragraph saying who finished ahead and why —
  or why the runs cannot honestly be ranked — and the numbers behind it
  existed only on screen. It is now a Markdown document: the same narrative,
  and each run's grade, metric, final and best value (with its epoch), epoch
  count and grade note. From epochix compare <a> <b> [-o file],
  GET /api/export/compare/md?runs=a,b, and a Download comparison button
  beside the race GIF. In English, Farsi and French.
• epochix compare was documented in the quickstart and did not exist. It
  does now.
• The comparison view's narrative was English whatever language the runs were
  told in; it now uses theirs. The view's own labels — metric, smoothing,
  run count, both download buttons, loading and error text — were English
  literals in every locale and are translated.

Added — every setting a parameter search tried, in the reports

• GridSearchCV's comparison never left the terminal. Fold rows were
  grouped by setting and the winner charted, then everything else the folds
  said was discarded; epochix check listed the candidates and no export
  could. The fold readings are now kept on the run, and the Markdown and PDF
  reports carry a Parameter search table — every setting, its mean, spread,
  range and fold count, best first, with the one that was charted marked —
  in English, Farsi and French. A plain cross-validation gets a
  Cross-validation table of its folds. The JSON export carries the raw
  readings. check ranks through the same function, so every surface names
  the same winner.
• The Markdown report's milestone suffix "(epoch N)" was English in every
  locale.

Added — a grade says when it is provisional

• An 11-epoch run and a 200-epoch run received equally confident letters.
  A two-reading run could show an A+ with nothing beside it. A letter now
  carries a note when the log shows it deserves less weight than it looks:
  it rests on fewer than five readings of the metric, or the metric was
  still setting new bests at the reading it was taken from, so it shows
  where the run got to rather than where it was heading. No confidence
  number is invented for this; both conditions are read straight off the log.
• Shown under the grade on the dashboard (browser and VS Code) and in the
  Markdown and PDF reports, in English, Farsi and French. Stored per frame
  (story_frames.grade_note, migration 0003; existing databases gain the
  column on open), in the JSON export, and computed identically by the VS Code
  extension's engine — the 38-log corpus truth now pins each log's final note
  for both engines.

Fixed

• The PDF cover's "since best … worse/better" row was English in Farsi and
  French reports; its translations existed and were not used.
• The VS Code extension's engine put the parser's confidence in a line into
  each frame's confidence, where the Python engine puts the run's
  advancement. Nothing displayed it, but a JSON export from the extension
  carried the wrong quantity under that name.

Added — the story says what to do next

• A run that went wrong now ends with a next step, not only a
  description. Past its peak: keep the checkpoint from the best epoch and
  use early stopping on the metric. Stalled: check the learning rate, then
  the data pipeline and the label mapping. Diverged: lower the learning rate
  or clip the gradients, and restart from a checkpoint at or before the last
  good epoch. The overfitting and plateau warnings, which described the
  problem and stopped, now say what to try.
• Before this, advice lived inside some narrative variants and not others,
  and the variant is chosen by run id, so whether a reader got a next step
  was luck. The step is now one fixed sentence per state, appended after
  every variant, in English, Farsi and French, identically in the Python
  engine and the VS Code extension's.

Fixed — an empty dashboard that waited forever (#35)

• A VS Code dashboard opened with nothing attached read "Waiting for
  training data…" forever. The extension told the webview every standalone
  panel was live, so a panel opened from the command palette or the shortcut,
  with no log and no terminal feeding it, waited for data nothing would ever
  send. A user reported exactly that screen: "epochs 0", no metric, no task.
  The panel now tells the webview whether anything feeds it, and when nothing
  does it says so and offers Try the demo, Open a log file and
  Watch the active terminal as buttons. A webview message may run only
  those three commands, never an arbitrary one.
• A log with no metrics in it also claimed data was on its way. A finished
  run with no frames now says no training metrics were found, and asks for a
  couple of example lines if the log does contain training output.
• The grade arc's own "Waiting for training data…" was drawn in English in
  every locale; it is translated now.