Revert an off-by-one regression in fn_task_update (introduced in 8e969bd72)
that translated 0-indexed step numbers to step-1, mis-targeting updates and
desyncing codeReviewVerdicts/stepCheckpoints keys with fn_review_step. Also
auto-advance step state from inside fn_review_step (in-progress on entry,
done on code-review APPROVE) so the dashboard reflects progress when an
agent runtime skips the follow-up fn_task_update call — observed with
permanent-agent CEO sessions on the openai-codex transport.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
