npm run push:subscribers reads the live rows in `push-tokens` — the ones the sender will really fan out to — and groups them by topic. Devices and people are counted apart, because one reader with a laptop and a phone is two devices and one subscriber, and the difference matters when the whole list is small. Naming a topic also lists its devices. Legacy FCM-era rows carry no `kind` and are skipped here exactly as shared/push.mjs skips them, so the number is what a broadcast would reach rather than an inflated row count. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AGbWC1RD3FzZTgbW7S4zDx