Uses up to k complete labeled training examples as demonstrations for each
predictor. Makes no model calls.
An example supplies a demonstration for a predictor when its inputs and
labels contain every field of that predictor's signature and the values are
valid. With sample: true (the default) examples are sampled without
replacement using local seeded random state; otherwise the first k are used.
Options: :k (default 16), :sample (default true).
Differences from DSPy: the build is unscored and returns no validation improvement claim; predictors without complete labeled examples keep no demos.