The test schema omitted `projection_turns.checkpoint_turn_count`, which
is the column `dropPendingTurnStarts` keys its "spare checkpoint rows"
guard on. Every existing test therefore ran the degraded path where the
clause is dropped for schema drift, so the guard's real behavior was
never exercised.

Added the column (migration 005 has it) and a case covering what the
clause is for: a pending row carrying a checkpoint survives the delete.
Confirmed the case fails when the clause is removed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>