v0.7.12

Fixed — the VS Code extension's own engine

The extension ships a TypeScript copy of the story engine, used by everyone who
has not installed the Python package. AGENTS.md warns that the two copies drift;
this one had drifted a long way. Driven through the log shapes that exposed
Python bugs in 0.7.7 and 0.7.8, it:

• Told no story at all for six of nine common log shapes — a plain
  loss / val_loss log (the most common thing a training script prints), a
  YOLO box_loss / cls_loss log, an XGBoost log, a ROUGE log and a run that
  went to NaN. Metrics it did not recognise defaulted the task to
  "classification", which then waited for a val_accuracy that was never
  logged. Unrecognised runs are now a neutral task narrated by the metric they
  actually logged, graded on how far it improved.
• Graded every metric on accuracy bands, higher-is-better. A loss has no
  absolute scale, and FID, log loss, l2 and EER-style errors are lower-is-better.
  The per-metric direction table, grading on improvement, and "incomplete" for a
  single reading are ported from the Python engine.
• Named metrics the run never logged: a log loss would have read as
  "Accuracy", Dice as IoU, a rising R² as "Error has dropped". Wording written for
  one metric is now used only for that metric.
• Said "Patterns are starting to click" as validation accuracy fell from 0.79
  to 0.60. It had no past-peak, stalled or diverged wording at all; the Python
  versions are ported, including "stalled means flat, never fallen".
• Warned "plateau" on a steadily falling loss — its only check was "no value
  went up" — and announced every grade change, including a fall to F, as "Grade
  improved". Overfitting and divergence warnings are ported too.
• Graded XGBoost on its TRAINING loss. The universal parser read
  train-logloss and valid-logloss as one key and kept the first. Python's
  boosting parser (XGBoost, LightGBM, CatBoost) is ported, so validation is what
  gets graded and the boosting round is the x-axis.
• Pinned any log without Epoch x/N in its first phase forever ("the model
  is processing its first examples at epoch 55"). With no known total, progress
  is now the fraction of achievable improvement, as in Python.

Fixed — the Python engine

• Every FID-graded run got A+, improving or worsening. There are no generative
  grading bands, so FID fell through to the classification bands,
  higher-is-better: every FID above 0.95 is "A+". FID's scale depends on the
  dataset and feature extractor, so it is graded on improvement instead.
• A live run that went to NaN showed no divergence until it exited. 0.7.8
  reported it only from the end-of-stream step, and a live run has no end — a
  diverged one can print nan for hours. It is reported on the NaN line itself,
  using that line's own sequence number: any other would collide with the
  (run_id, seq) primary key, where the store's on_conflict_do_nothing would
  drop the frame silently.