Fix four bugs in plot_template_data.py causing compute_oracle and pipeline to fail

[Copilot]

plot_template_data.py had multiple bugs that collectively caused the entire script to fail silently or with errors — the most subtle being that select_variables_and_clean returned an empty table, making all downstream logic operate on no data.

Bugs fixed

• select_variables_and_clean (root cause): groupby(keys).count() replaced real effectif values with row counts. Since "Session" is part of the key, every (Session, formation) combo is unique — so groups[cible] > 1 always produced an empty DataFrame. Replaced with duplicated(subset=keys, keep=False) + inversion to keep only rows with a unique key combination while preserving actual values:

  # Before (broken)
  groups = df[[*keys, cible]].groupby(keys).count()
  filtered = groups[groups[cible] > 1].reset_index(drop=False)
  mask = filtered.duplicated(subset=keys, keep=False)
  return filtered[~mask][[*keys, cible]], cible
  
  # After
  subset = df[[*keys, cible]]
  mask = subset.duplicated(subset=keys, keep=False)
  return subset[~mask].reset_index(drop=True), cible

• compute_oracle: .dropna(axis=0) was commented out as # fails, but it is required — without it the pivot table retains NaN for formations present in only one year, causing mean_absolute_error to raise ValueError. Uncommented it.

• split_train_test: Parameter typo cuble → cible, and test-set rows used train_test mask instead of ~train_test.

• make_pipeline: c != "cible" (string literal) → c != cible (variable), so the target column is correctly excluded from the feature list.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.