Kick off today's payload tuning for the three discovery scrapers and return at once.

Reads the last few days out of remy_cron_history plus how each of those payloads performed (remy_cron_logs cap vs fetched_value) and the handles they already pulled, then asks the model for a fresh, rotated payload per job. Runs detached: the model call is slow enough that a timeout would have the scheduler retry and store a second set of payloads for the same day. Any job the model returns nothing usable for falls back to its seed payload.