sqlmesh timeseries rolling windows metrics seem off

[ryscheng]

What is it?

https://www.opensource.observer/projects/opensource-observer

commit counts look low, separately from the fact that we're missing the backfill for reference https://github.com/opensource-observer/oso/graphs/contributors

Current hypothesis is that the sqlmesh factory is technically correct, but Clickhouse execution has so many errors, maybe there's writes that get lost and affect correctness.

Requires more investigation