Training size does not explain the pairwise pattern¶

datasets/results/projects-qt-creator_qt-creator-qt_qtbase-n788.json. qt-creator vs qtbase rerun at 788 training examples per side (TRAIN_SIZE=788, a nested subsample of the 1,517), the size the two qtdeclarative pairs trained at. Same held-out changes, one seed.

At the same 788 examples where qt-creator showed no advantage over qtdeclarative (+0.005), it still beats qtbase. The difference between the pairs is the pair, not the training size, which confirms the reading of the earlier entry: qt-creator's refinements carry something qtbase's adapter lacks and qtdeclarative's has. The effect roughly halves with half the data, as a learned property should, and qtbase still never wins at home.