v0.7.18

Changed — the extension grades from the Python engine's own tables

• The VS Code extension's grading, direction and phase logic was a hand
  copy, and had drifted. It never got the Python engine's fix for a single
  reading — a bounded metric's first value is read on its own scale rather
  than as zero improvement — so a one-epoch run, or a fit-once result, ended
  its story in the wrong phase (single_epoch.log: learning where Python
  says mastering; a GridSearchCV result: awakening where Python says
  polishing). It also left the improvement ratio unclamped, and answered a
  made-up 0.05 of progress for a run whose length was unknown.
• Every grading, direction, phase, stall, past-peak and warning threshold is
  now generated from the Python engine (grading.generated.ts); the functions
  that read them are ported once, line for line, and replay 2,800 of Python's
  own answers (grading.golden.json) — every band edge of every task and
  metric, every metric name's direction, phases, trajectories, grade notes. A
  drift fails a test now. The corpus both engines are held to also pins every
  log's final grade and phase: the extension matches Python on all 39.
• LightGBM's raw l1 / l2 names get their direction pinned in the Python
  engine too; only the extension had them.