v0.7.13

Fixed — invented milestones

• "Crossed 25%, 50%, 75% and 90%" at epoch 1 for a loss of 1.2. The
  milestone tracker took its direction from the task, not the metric the run is
  graded on, so a custom run graded on val_loss — and every FID — was read
  higher-is-better. A loss climbing from 0.5 to 1.3 "crossed 90%" on the way
  up, and "best so far" tracked the worst value. Percentage milestones now fire
  only for metrics that are fractions (accuracy, IoU, mAP…) logged on a 0–1
  scale.
• "Biggest single-epoch improvement" reported the biggest move either way,
  so a run's worst deterioration was celebrated. It is now the largest step in
  the right direction, and a run that never improved claims none.

Fixed — a validation metric filed as a training metric

• Python: validation_accuracy, validation_f1, test_l1, val_mae_cm and
  val_error_rate lost their split and were graded as training metrics — the
  numbers that keep improving while a model overfits. A held-out prefix now
  keeps its split whenever the metric has one.
• VS Code extension: its hand-ported name tables had drifted from Python's —
  170 of 319 names canonicalised differently. valid_loss and eval_loss were
  read as TRAINING loss, and a Keras regression's mae and val_mae merged into
  one series. The tables are now generated from the Python source
  (make gen-ts-tables), and a golden file of Python's answers for 375 names is
  replayed by the extension's tests, so they cannot drift silently again.

Fixed — W&B offline import

• epochix import-wandb <dir> failed on every current wandb. wandb 0.30
  moved its writer to Go and deleted the Python reader the importer used, and
  the importer then told people who had wandb installed to install it. The
  .wandb file's framing is now read directly (the format is small and stable);
  only wandb's protobuf definitions are still used. A run that is still being
  written is read up to its last whole record, and corruption is an error
  rather than a guess. None of this was caught because no CI job installed
  wandb — see below.

Fixed — VS Code extension settings that did nothing

• epochix.theme was ignored: the panel had its own resolver that only
  followed VS Code. "light" and "dark" now apply, and changing the setting
  updates an open dashboard.
• epochix.taskHint never reached an engine, and task detection would have
  overwritten it anyway. A pinned task is now kept, and is sent to the sidecar
  when a run is saved there. segmentation and generative can be pinned too.
• epochix.llmFallback was read and never used; the extension has no LLM
  path. It is marked deprecated, with a message saying so.
• Saving a log to the sidecar recorded its last epoch twice. The "run
  finished" marker was sent as an extra copy of the final metric on a new seq,
  which the server stored as a real measurement.

Fixed — stories in the language you chose

• The VS Code extension told every story in English, whatever
  epochix.locale said: its engine carried its own English-only copy of the
  templates. They are now generated from the Python engine's templates in
  every locale (en, fa, fr), so the two engines tell the same stories.
• Runs saved through the sidecar were narrated in English too — the
  extension never sent the locale when it created the run there.
• Warnings and milestones were English in every locale, in both engines.
  They now come from one translated catalogue (story_engine/messages.py).
• "validation MAE" appeared mid-sentence in French and Farsi stories. The
  split word is now in the story's own language and word order ("MAE de
  validation").
• A rising R² was narrated as "Error falls steadily" in both engines: the
  regression templates describe an error, and R² is regression's first-choice
  metric. Those templates are now used only for error metrics.
• The extension now reports a single result with no epoch (a script that fits
  once and prints a score) as a result, as Python does, instead of narrating
  it as the first epoch of a training run.

Fixed — dashboard localisation

• The epoch label, phase names, connection status and task names were English
  in every locale, while fa.json and fr.json held translations no code read
  (40 of 60 keys were dead). They are now read at runtime; the unused keys are
  gone, and a test fails on any key that is missing, unused, or absent from a
  locale.
• Timeline cards from the extension's own engine were headed with raw
  identifiers ("grade_transition"). Every milestone kind either engine emits
  now has a title in every locale.
• The phase badge and the phase journey drew different icons for "mastering".

CI

• The integrations job now also runs the tests that need torch, numpy, wandb or
  pypdfium2 — activation capture, architecture capture, the accelerator report,
  the W&B offline import and the PDF render check. They skipped everywhere
  before, and the job fails if any of them skips.

Performance

• The universal parser is 1.5x faster (16.7k to 25k lines/sec on CI), with
  byte-identical output over 644,355 lines from every log in the repository.
  Each of its patterns was scanning a line twice. The parser benchmark had
  never run until now; it has per-parser targets, and the universal parser's
  gap to 50k is recorded in ROADMAP.md rather than hidden.

Removed

• scripts/_gen_templates.py and scripts/_gen_i18n_templates.py: one-shot
  generators for narrative templates that have since been corrected by hand.
  Running either would have overwritten 30 template files and reverted those
  fixes.