M4 — ensemble search. Run an inner optimizer several times on different bootstrap resamples of the training set (sampling with replacement, the classic bagging trick that makes the resulting candidates complementary), collect the candidate programs, and combine them into one program with EP-5's ensemble combinator under a majority-vote reducer. The ensemble runs all members on the input and returns their modal answer, which is correct more often than any single member when the members err on different inputs.

Resampling is deterministic (a fixed linear-congruential stream seeded by the member index), so the search is reproducible run to run.

The returned artifact is structure-changing: it is an Ensemble combinator over the optimized members, not just a parameter rewrite of the student. The member parameters are persisted by encodeCompiled, but the reducer closure and exact member structure live in the program template held in code. Load saved state onto the matching ensemble template, not onto the plain student.