Pin Error Contracts With Characterization Tests Before AI-Assisted Refactors

AI coding tools often suggest consolidating error-handling logic during refactors, preferring a single exception type or replacing None returns with raises — but this can silently break production callers that already depend on multiple distinct error shapes. A dispatcher function may legitimately return None, raise a ValueError, or return a status-keyed dict, and downstream code branches on all three outcomes. The real risk in such refactors is an unrecorded error taxonomy, not missing type annotations. The recommended safeguard is to write characterization tests that pin each fixture's exception type, return shape, status integer, and warning log count before touching any except clause. Only once those tests stay green should a single except clause be changed at a time.