Verify test-judged code-patch options by substitution plus step models
When multiple-choice code repairs are graded by running tests, every option that passes is accepted, so substitute each option literally, run the suite, then sweep a broad input domain against a reference to prove the suite rejects every behaviourally wrong option. If a step budget enforces a complexity bound and the target interpreter's step semantics are unknown, measure three models with sys.settrace (loop-header line events, all line events, AST-node-weighted line events) and choose a budget that gives the same verdict under all three. Searching an integer domain (e.g. integer sqrt with n up to 1e8) separates O(log n) from O(sqrt n) by over 100x without long list literals in test args. Raising an exception from the local trace function cleanly aborts infinite loops.