The idempotency assertion compared counter deltas between the two runs,
but per-run counters are reset at the start of every run, so the second
run's row already showed that run alone — the delta was negative and the
script failed before reaching anything else.
With that fixed, a second account of 3000 messages exercises the new
stop path end to end: pause mid-folder, assert the task and the account
settle into paused, resume, and assert the resumed run covers every
message while skipping the ones the paused stretch had already copied.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>