Outcome ledger live verification UX gaps

Found while driving dev's console end to end for a sandbox coder run (Task 12443). Each item is a 1–2 sentence requirement. 1. Add node → "Add to a host": mark a host whose disk is full (or whose node pauses on disk) as not pickable, or warn before Create; today the node is created and immediately shows "Paused: disk 100% used". 2. Cloud connections: flag a saved subnet or security group that no longer exists in the account (the edit dialog shows the stale ids with no warning), so the first sign is not a failed RunInstances during Add node. 3. Add node failure: after "AWS refused RunInstances", offer a link to edit the connection from the error, not just Retry. 4. Task editor: show the Account (and Agent) picker on an existing task's Advanced section, as the new-task form does, so a task stuck on a bad account can be re-pinned without deleting and recreating it. 5. New project → Nodes: when every listed node is offline, say so and offer Add node in place; today the wizard picks an offline node and the onboarding task just waits. Out of scope: the daemon socket reply cap (fixed in Task 12443's PR); the scheduler's account choice (filed separately). Done means: - Each item above is either implemented with a test or explicitly declined with a reason in the PR body. - Required checks: ./dc check quick, changed-area tests, merge lane full gate.

Built all five items with tests: disk-full hosts not offered, stale subnet/SG flagged, Edit connection link on a failed launch, account re-pin on an existing task, all-offline wizard banner with Add node. PR #7688 merged (adee85f605) through the lane's full gate; dev deploy passed. Visual check done.