Author new graded assignments (notebooks now ship full worked solutions)

[jonfroehlich]

Background

Issue #6 is done (PR #25): the published notebooks are now complete teaching artifacts — every exercise prompt / "your turn" keeps its framing, but a clearly-labeled ### Example solution (markdown + worked, executed code) now sits right below it. That means the previous student assignments are "spoiled" — the answers are visible inline.

Jon's plan was to author new graded assignments separately. This issue tracks that.

Where the now-published solutions live

• GestureRecognizer-ShapeBased: generate_kfolds (from scratch), a DTW matcher (~91%), preprocessing demo.
• GestureRecognizer-FeatureBased: frequency-domain + 2D feature "your turn" answers, preprocessing demo. (Cross-user / leave-one-out classification was already worked in v2.)
• Feature Selection and Hyperparameter Tuning: preprocessing demo.
• StepTracker-Exercises (now the single consolidated notebook): offline step detector + a real-time buffered detector.
• Tutorials (Quantization, Frequency Analysis, Comparing Signals): model answers for the open-ended exercises.

Task

• Design new A3/A4-style assignments that don't simply reproduce the published solutions (e.g. new gesture sets, new features, a different matcher, stricter cross-user protocol, real-time tuning targets).
• Decide how to distribute starter code now that the public notebooks are answer keys (separate assignment repo? stripped templates generated from these?).

Heads-up: fix these correctness issues before reusing as graded material

A few pre-existing audit issues overlap with this content and should be resolved first:

• FeatureBased: train/test leakage in per-user leave-one-out cross-validation #17 — FeatureBased cross-user leave-one-out has train/test leakage (matters if cross-user becomes an assignment).
• StepTracker (both notebooks): top-2/top-3 frequencies duplicated + real-time peak slope wraps at buffer start #20 — StepTracker real-time peak detector slope wraps at the buffer start (the original baseline cell; the new Example solution starts the loop at j=1 and avoids it).
• Feature Selection: manual grid-search uses stale 'filtered_df' for X instead of the freshly built df #22 — Feature Selection grid-search uses a stale filtered_df for X.

Context: follow-through from #6 / PR #25.