Semi-Supervised Label Propagation

You have 1,000 examples but only 10 labels. Label propagation spreads the known labels along a similarity graph, so unlabelled points take the label of the region they sit in.

Implement label_propagation(X, labels, gamma=1.0, iters=1000, tol=1e-6), where labels uses -1 for unlabelled points and 0..K-1 otherwise:

1. Build RBF affinities $W_{ij} = \exp(-\gamma\lVert x_i - x_j\rVert^2)$ with a zero diagonal, and the row-normalised transition matrix $T = D^{-1}W$.
2. Start F with one-hot rows for labelled points and zeros for the rest.
3. Repeat F = T @ F, then clamp the labelled rows back to their one-hot values, until the largest change is below tol or iters is reached.
4. Return (predicted_labels, F), where predictions are the row argmax.