Metric watch

• Run pnpm --filter analytics radar:watch --window daily; use --window hourly for provisioning and environment-key checks.
• Read the check status, numerator, denominator, seven-day comparison, and coverage. insufficient is not a pass. unavailable and insufficient remain explicit and make a completed watch exit 1; exit 2 means the run could not complete.
• Compare daily thresholds with the incident window and an ordinary week before changing them. Record both windows and the observed counts in the PR or report.
• Use pnpm --filter analytics radar:backtest to compare the Oct 9–23 incident window with the Oct 2–8 clean week.
• Action success uses a weighted action.response rate. Keep its app-specific threshold below the quiet-week distribution; do not copy the 95% starting value without a backtest.
• Read signup campaign tags beside signup drops. Paid search changes are context, not proof that product behavior changed.
• Event first-seen checks require 10 events that day; went-to-zero checks require at least 50 events across the preceding seven days.
• Read campaign-followup as identified recipient counts after send versus the same recipients' prior seven days, split by recipient and activity app. Anonymous landings cannot be attributed to recipients.
• The hourly Analytics alert rules cannot express ratios or trailing baselines; do not replace these checks with simple count rules.