main-app's shared Bun RedisClient, firehose's cache-invalidation publisher and hosting's revalidate queue client sit idle behind haproxy redis_write (timeout client/server 180s), so each was cut and reconnected every 180 s. Add startRedisKeepalive to @wispplace/constants (unref'd 60 s PING, skipped while not ready or a ping is outstanding, stopped on close) and use it for the three clients.