feat: opportunistic ip tracking scans on wake and shutdown

[mwpastore]

Summary

Stacked on the seriousm4x#1763 branch (claude/upsnap-magic-packet-routing-dc13ba is the base here, so the diff shows only the six new commits). Where the base PR makes ip tracking periodic, this makes it opportunistic — scans fire at the two moments a stale ip actually hurts:

• Waking a device. The wake wait used to ping the recorded ip blindly for up to wake_timeout (default 120s), so a device that renewed its DHCP lease during boot woke fine physically but was reported as "not online" — and stayed misreported until the next periodic sweep. Now a tracking scan of the device's subnet fires 15s after the wake attempt starts, and the wait loop re-reads the ip from the db before every ping attempt, so a moved ip is picked up seconds after the device answers ARP and the wake succeeds within its timeout.
• Acting on a device's ip. Scheduled and manual shutdowns, reboots, sleeps, and group shutdowns refresh the device's ip with a scan before pinging or substituting {{ DEVICE_IP }}. Scheduled actions are the primary beneficiary: they have no UI gating, and a stale ip previously made the shutdown cron's online check silently no-op. Manual actions are only offered while the device shows online, and a stale ip usually flips it offline first — the wake path above then heals it — but "online" can lie: the old ip may have been re-leased to another host that answers pings, and a custom ping_cmd runs verbatim and ignores the ip field entirely, so status accuracy implies nothing about ip accuracy. The pre-action scan re-locates the mac before the action targets an address, so shutdown_cmd and sleep can't hit the wrong host.

Opportunistic scans are gated the same way on both paths: the global track_ip_interval must be set and the device's track_ip toggle on.

Commits

1. fix: only write status changes when saving device status — every device-record flow that only mutates status (ping/wake/shutdown crons, all six handlers, boot-time reset) now saves with IgnoreUnchangedFields, so their full-record saves can never clobber an ip written concurrently by a tracking scan. (The ports status save is deliberately untouched.) Save baselines are refreshed (PostScan) after each intermediate pending write, so a later revert to the load-time status is never dropped as an unchanged field — pinned by a cronjobs regression test.
2. feat: follow tracked ip changes during wake and ping waits — WakeDevice/PingDevice take a getIp func() string (nil = the record's ip); networking.DeviceIPFunc(app, device) composes the db-reading getter, and the gate is exported as networking.DeviceTrackingEnabled.
3. feat: scan a woken device's subnet shortly after waking it — iptracking.TrackDeviceAfterWake schedules the +15s scan at all four wake initiation points. TrackOneSubnet now wraps the nmap run in a singleflight group keyed on the subnet, so a group wake coalesces into one scan and opportunistic scans join an in-flight sweep instead of doubling it (golang.org/x/sync promoted to a direct dependency). Nmap invocations are additionally serialized process-wide, since raising and restoring CAP_NET_RAW mutates process state.
4. feat: refresh tracked ips before shutdown actions — iptracking.TrackDevice (synchronous scan + re-read, returns the given record on gate-off or failure) at the top of the shutdown/reboot/group-shutdown flows and in the shutdown cron before its online check.
5. feat: run shutdown, reboot, and sleep actions asynchronously — the frontend passes async=true for those three actions (a capability the API already had), so requests return as soon as the action starts instead of blocking for the scan plus the action — which could outlive reverse-proxy timeouts even before this PR. The card's pending countdown plus realtime updates deliver the outcome, and the countdown re-syncs the record once if it expires while still pending. The action handlers also no-op with the current record when the device is already pending, so a double click can't start a second action. Wake stays synchronous because its response drives link_open.
6. feat: refresh the tracked ip before sleeping a device — sleep sends an http request to the device's ip, so it gets the same pre-action scan as the shutdown flows.

Behavior changes to note

• The shutdown cron no longer bails when status != "online". By that point it has just successfully pinged the device via its configured check (default ICMP or ping_cmd); a live ping supersedes a status column that is stale in exactly the scenario this PR targets (ip moved → ping cron marked the device offline). pending is still respected: checked on entry and re-checked after the potentially slow pre-action scan, so a wake initiated mid-scan is never raced by the shutdown.
• Scans coalesce. Concurrent scans of the same subnet share one nmap run and its result.
• The scanner remains an ip-only writer. Considered and rejected marking scanned devices online: ARP presence doesn't satisfy the per-device status contract (ping or custom ping_cmd), and it would flap for ICMP-blocking devices between each scan and the next ping tick.

Testing

• go build for linux, windows, and darwin; go vet; gofmt — clean.
• The backend suite passes under -race: the iptracking tests (gating on both toggles, the delayed post-wake scan, singleflight join proving one nmap run for two concurrent callers, TrackDevice's refresh/gate-off/unscannable paths, guard/skip/orchestration coverage), new networking tests for DeviceTrackingEnabled, DeviceIPFunc, and WakeDevice's getIp consultation, and a cronjobs regression test pinning that a failed scheduled wake persists its status revert. All hermetic — stubbed scanner, loopback subnets, real migrated schema via tests.NewTestApp.
• The pre-existing TestCheckPort/Invalid_IP failure is environment-dependent and unrelated (fails on the base branch too).

Not included

• No schema changes — the feature rides entirely on feat: optional per-device IP address tracking via periodic ARP scan seriousm4x/UpSnap#1763's track_ip / track_ip_interval. Frontend changes are confined to the three async=true call sites and the countdown re-sync.

🤖 Generated with Claude Code