Freshness checks that people actually read
Most freshness alerting fails for a social reason, not a technical one: the alerts arrive where nobody looks. We now wire freshness as an explicit contract per table — expected interval, tolerance, owner — and publish one daily digest to the channel the owning team already uses.
The mechanics are unglamorous. Each model declares `freshness_interval` and `freshness_tolerance`. A small job compares the newest partition timestamp against the declaration, and writes one row per table into a `_ops.freshness` table. Two alerts fall out of that table: a soft one in the daily digest, and a hard one that pages when a table feeding a contractual report crosses tolerance twice in a row.