Dual write timeout divergence

You're mid-migration from a monolith Postgres inventory table to a new sharded store, running a dual-write phase (app writes to both, reads still from old). At 09:30, reconciliation alerts fire: 0.3% of SKUs have divergent stock counts between old and new, and the divergence is growing ~0.1%/hour. Dashboards: both stores are up; the new store occasionally returns write timeouts (~0.5% of writes) under load, but the app logs those as warnings and moves on; no read-path impact yet. The cutover to read-from-new is scheduled for tomorrow. How do you triage the divergence, decide whether to proceed, and reconcile the two stores?