effects — I am changing X; open these cards first

An index, not a spec. If a card and this table disagree, fix the card. First-order effects only; each card's If you change this carries the waterfall. What points into the box from outside (phone Shortcuts automation, the claude.ai artifact, the disabled cloud routine, Kuma/ntfy endpoints) is listed at the bottom because nothing on disk names it.

Changing…	Open	Do not forget
a disk (replace, reseat, cable)	objects/storage/disks.md → snapraid-array.md	write ~/disk-map-<date>.txt first; smartd.conf uses letters; snapraid status before any sync
/etc/fstab or a mount option	objects/storage/mergerfs-pool.md, disks.md	findmnt --verify; daemon-reload; the 16 pool consumers; x-systemd.requires stays
/etc/snapraid.conf (exclude, content, parity)	objects/storage/snapraid-array.md → processes/snapraid-sync.md	snapraid diff first; precheck limits 500/2000; a big exclude rewrites hundreds of GB of parity
the sync guard or precheck	processes/snapraid-sync.md	the rescue kit copies them nightly; README rebuild order depends on them
docker-compose.yml or a container	objects/containers/torrentvpn-stack.md	never bare up -d; gluetun+qbittorrent together; samba is a ghost
a download client, qBittorrent limits, an *arr	objects/containers/torrentvpn-stack.md, storage/hardening-plan.md item 18	removeCompletedDownloads=false is the interlock; enumerate who consumes the "torrent stopped" event
anything under /mnt/storage/immich	objects/containers/immich.md, storage/precious-copy.md	the motion-photo originals under encoded-video/; 02:00 dump → 02:20 copy order
the pool's top-level trees	objects/storage/datasets.md	parity diff counts; rsnapshot and precious watch two of them
a schedule (cron, timer)	objects/automation/schedules.md	the ordering block; watchdog assumes ≤ 30 h; training-sync deadlines
an alert destination	objects/automation/alerting.md	the ntfy topic lives in four files
sudoers / doing a root write	processes/root-maintenance.md, objects/host/sudo-and-tools.md	scoped grant, verify, revoke; firmware writes are user-run
rebooting or powering off	objects/host/machine.md	disk-map note; no jobs running; expect one rescue-watchdog alert; snapraid status before sync
~/CLAUDE.md, this map, memory	objects/host/claude-tooling.md, map/CONTEXT.md	router stays a router; run _meta/build-index.sh; memory points at cards, does not restate them
the Garmin MCP, routines, training policy	processes/training-sync.md, objects/containers/garmin-mcp.md	localhost only; foreground testing; the phone automation reads "Early wake"
a maintenance job in ~/maintenance/	processes/scheduled-maintenance.md	busy-guard; dated timers; reports close items on the hardening card
the hardening artifact	objects/storage/hardening-plan.md	read the live version first, edit content not design, republish with url

Points into the box from outside (invisible from inside)

• iPhone Shortcuts automation → reads calendar event "Early wake" (training-sync output).
• claude.ai artifact d6b76e92-… → the hardening report; only this map's card and memory name it.
• claude.ai cloud routine trig_01Ke8GAbB9q1y4J7UXqRhtm2 → disabled; would need to become three routines.
• Uptime Kuma push URLs and ntfy topic → in /etc/rescue-kit.conf, scrutiny.yaml, smartd-ntfy.sh, snapraid-alert@.service.
• OpenAI/ChatGPT via garmin-tunnel.service and codex-builder-tunnel.service (user units).
• Phone/web Claude usage → shares the 5-hour window the cron passes depend on.