Algorithm Change Check

Daily ML maintenance endpoint (Phase 4 rewire).

Called daily by GCP Cloud Scheduler. Per WIRE-2 = 1'a, drift detection is now performed in-tick via KLDivergenceDetector (which can also automatically reset segment gates on detection). This endpoint no longer performs the inline CUSUM+BOCPD detection that previously duplicated and shadowed the KL path.

What it does now:

1. Train the churn-prediction model on accumulated creator data
2. Re-discover dynamic segments via HDP
3. Report current drift status (read-only, from KL detector state)

Phase 4 wiring: train_churn_model and discover_segments needed periodic triggers; this is the natural daily slot. Both methods are no-ops until enough data accumulates (churn_min_creators / hdp_min_videos gates).