fix(examples): stop trusting candidate-reported values in example evaluators

[zichen0116]

Summary

Fixes #514: the matmul and function_minimization example evaluators scored values the candidate program reported about itself. Those values are produced inside the EVOLVE-BLOCK — i.e. under the LLM's control during evolution — so a candidate could score above the state of the art without solving the task.

The exploits (verified at HEAD a1e9d91 before the fix)

matmul (examples/alphaevolve_math_problems/matmul/evaluator.py):

• verify_tensor_decomposition printed a warning and returned early when any factor array was empty, so verification was silently skipped.
• The score was BENCHMARK / rank using the candidate-reported rank.
• A candidate returning (np.zeros(0), np.zeros(0), np.zeros(0)), 2, 4, 5, 0.0, 2 scored 16.0 (32.0 with a claimed rank of 1); a correct rank-32 decomposition scores 1.0.

function_minimization (examples/function_minimization/evaluator.py):

• When run_search() returned three values, the evaluator used the returned value directly.
• Returning (0.5, 0.5, -1.519) scored 0.793 (value_score 1.0) at a point whose true objective value is 0.6931 (honest (0.5, 0.5) return scores 0.449).

Changes

• matmul: empty factor arrays now fail verification (score 0) instead of skipping it; the reported (n, m, p) are checked against the fixed <2, 4, 5> target; the rank is derived from factor_matrix_1.shape[1] rather than the reported value.
• function_minimization: the objective is always recomputed from the returned (x, y) via a new _objective_function helper, in both evaluate() and evaluate_stage1().
• Added tests/test_example_evaluator_integrity.py with 5 regression tests covering both exploits plus the dimension check.

Verification

• Reproduced both exploits on the pre-fix code (scores matched the issue: 16.0 and 0.793); after the fix the cheats score 0.0 / equal-to-honest.
• New tests: 5/5 pass on fixed code; 4/5 fail on pre-fix code (the 5th locks in the dimension-check behavior, which the old shape validation also rejected).
• Honest path unchanged: the bundled function_minimization/initial_program.py still evaluates end-to-end with sensible scores (value_score 0.974); a valid rank-40 decomposition with a lied-about rank now correctly scores 32/40.
• New test file is black-clean (line-length 100, per repo config); function_minimization/evaluator.py had pre-existing black drift outside the touched hunks, left alone.

Closes #514

[CLAassistant]

All committers have signed the CLA.