Coordinate instruction search over predictors.
The unchanged program is evaluated first. Then, for depth rounds, each
predictor in name order gets breadth proposed instructions from the
proposal role. The first round proposes from the current instruction and
training evidence; later rounds also see that predictor's earlier attempts
and their validation scores. Each proposal is evaluated with the other
predictors fixed at the incumbent, and a strictly better complete eligible
result becomes the new incumbent before the next predictor is tried.
Options: :breadth (default 4), :depth (default 2), :evidence_examples
(default 3), and :predictors (names to search; default all).
Differences from DSPy: candidates are scored on the explicit validation set,
not the training set; proposals are one typed call returning a list rather
than n sampled completions; no temperature is changed.