The migration runner now splits multi-statement SQL files and executes
each statement individually. "duplicate column name" errors are skipped
since the column already exists from a prior partial run. This fixes a
crash on startup when migration 004_voice_optimization.sql re-ran
against a database that already had the columns.