feat: v0.9.2 Ask Claude on an alert starts an investigation session

- Ops Alerts → Ask Claude now calls `devdesk attach --investigate SESSION UID`.
  The server looks the alert up, writes the prompt and starts Claude in a new
  alert-<name> session with it as the first prompt, already sent. Prompts are
  stored in prompts/<session>.txt and consumed once by devdesk-claude-run.
  A session that is already running is just opened.
- The prompt has the rule, query, summary, details and labels, and where it
  runs: the CT/VM id, the node it is on now, status and HA state, and
  `ssh <node> pct exec <id> --`. It reminds Claude that HA guests move
  (`devdesk guests <id>`), and gives aggregate rules a hint.
- `devdesk guests`: Proxmox inventory over SSH (DEVDESK_PVE_NODES). pvesh
  gives the cluster view; each node gives lxc-info and guest-agent IPs.
  Cached in guests.json, keeping last-seen IPs for stopped guests; --alerts
  refreshes it in the background every 5 minutes.
- Alerts carry uid, query, description and targets, matched by instance
  IP/host or by name; each card shows the target.
- cmd_alerts moves to libexec/devdesk-ops.sh.

Closes #39

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

etc/env

@@ -30,6 +30,13 @@ DEVDESK_GRAFANA_URL=http://10.10.10.49:3000
 DEVDESK_GRAFANA_PUBLIC_URL=https://dashboard.lotusguild.org
 DEVDESK_GRAFANA_TOKEN_FILE=/etc/devdesk/grafana-token
 
+# Proxmox nodes to ask where guests run (`devdesk guests`, and the CT/VM and node
+# shown on each alert and given to Claude by "Ask Claude"). Space-separated SSH
+# hosts; the first that answers gives the cluster view, then every online node is
+# asked for its containers' IPs, by node name. Needs root SSH keys to the nodes.
+# Empty = off.
+DEVDESK_PVE_NODES=
+
 # Claude Code binary and extra arguments for every session.
 CLAUDE_BIN=/root/.local/bin/claude
 CLAUDE_ARGS=