Cohort windows that match the product, not the template
Vendor defaults love D1, D7, D30. A grocery app may earn a D7. A car insurance app that asks for a quote once a year will look dead on D30 even when the quote was the entire job. The template does not know this. The slide still gets made.
Match the window to the promise. If the product promises a weekly shop, measure a week. If it promises a completed claim, measure to claim-closed, however long that is, and stop pretending D7 is “engagement”.
Re-entry is not a bug
People come back. A cohort that forbids re-entry will punish a loyal customer who logged out. A cohort that allows infinite re-entry will count a confused session as a renaissance. Write the rule down. The cohere step exists for this argument.
Retention Forensics spends a week on windows because most “churn crises” we are invited to inspect are mismatched clocks. The users did not leave. The calendar did.