A bind mount arrives owned by root however the image left its mount point, so the daemon could not write /data/state.json. The write threw out of the firehose handler, which dropped the event and left the cursor where it was, so every reconnect resumed from the live head and lost whatever landed while the socket was down. A durability problem became a correctness one. The store now warns once and carries on in memory, and the image takes the mount as root before dropping to the unprivileged user, so the mounted case persists again. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>