Thanks!

The profile sats are (likewise) denormalized and were especially prone to this kind of divergence before I unified the increment logic.

I'll verify the accuracy of my aggregate queries and plan for a migration that recomputes the profile values.