Related to #16482

Why
Customers need proactive alerts when users experience unexpected workspace disconnects (network blips, agent crashes). Current Prometheus metrics (coderd_agentstats_connection_count, coderd_agents_connections{status="disconnected"}) only expose agent ↔ coderd state and cannot distinguish intentional session closures from unexpected client↔agent drops.

Proposal

1. Emit Prometheus metrics that increment on:
   • graceful (user-initiated) disconnects
   • ungraceful (timeout/crash/network) disconnects
2. Document new metrics in docs/admin/integrations/prometheus with example alert rules.
3. Ensure metrics are exposed per workspace and, if feasible, per user for targeted alerting.