v0.7.24

Fixed — found by running a real ResNet-18 on CIFAR-10 through it

Every demo is a small model. A standard benchmark run — ResNet-18, 30 epochs,
93.9% validation accuracy, now in the test corpus with the script that
produced it — showed the dashboard contradicting itself.

• A warm-up blip was reported as overfitting, for good. Validation loss
  rose twice while the one-cycle learning rate climbed, then set a new best
  the next epoch and fell to its lowest at the last one. The warning fired at
  epoch 5 and stood to the end, telling the reader to "stop at the best
  validation epoch" — epoch 30 — while the plain-English panel below said the
  model had learned rather than memorised. Both engines now withdraw the
  warning when validation loss beats its best from before the rise, and let
  it fire again if the run really does overfit later.
• A print(model) architecture listed stages and the blocks inside them:
  "Conv2d + BatchNorm2d + Sequential + BasicBlock ×2 + Sequential + …". Only
  the top level counts now, with Sequential/ModuleList opened one level,
  so a ResNet-18 reads as it is built: a stem, eight residual blocks, a
  classifier. PyTorch's folded repeats ((0-3): 4 x TransformerEncoderLayer)
  were skipped entirely; they are kept.
• The extension labelled layers from a hand-kept subset of the Python
  table (no ResNet, VGG or YOLO entries) and kept trivial layers the Python
  side drops. The table is generated from Python now, and both engines are
  pinned to Python's answers on every print(model) shape in the tests.
• A planned learning-rate decay filled the warning strip with three amber
  warnings. A learning-rate drop is not a problem; the strip shows problems,
  and the learning-rate chart shows the schedule.
• A metric printed after the story's metric on a line counted for the
  next epoch. The engine built a frame when the story's metric arrived and
  read the rest from history, so on ... val_accuracy=0.9357 lr=0.00442 the
  learning-rate drop at epoch 28 reported the epoch 26 -> 27 change, the last
  line's never reached any frame, and a log printing val_acc before
  val_loss had its overfit check a line late. The pipeline now hands the
  engine each line whole (StoryEngine.process_line); the extension's engine
  already worked a line at a time, and both agree in the tests.
• A warning message could name an object key in the dashboard (CodeQL
  js/remote-property-injection, found on this release's pull request before
  it shipped). Warning text comes from log files; the kinds are kept in a
  Map now.
• The grade was shown three times in one card; the stat tile repeating the
  large letter above it is gone. The learning-curve legend sat on the spot a
  good run's curve ends, and its loss label ran off the edge; it has its own
  row above the plot.

Fixed — docs

• A Windows contributor's first command failed. CONTRIBUTING.md opened
  with cd frontend && npm ci && npm test, and Windows PowerShell 5.1 rejects
  &&. Every shell example is one command per line now, and a test fails if
  one chains with && again.