Grafana alert pack: codify the operational watch-list as importable alerts

[hardbyte]

Motivation

The 0.5→0.6 upgrade guide and the managed-Postgres guide define a precise watch-list (ring rotation, prune blockers, queue lag, xact age), but operators must hand-build alerts from prose. docs/grafana/ ships dashboards; it should ship alert rules too.

Scope

Importable Prometheus/Grafana alert rules beside docs/grafana/awa-dashboard.json, covering at minimum:

• pg_stat_activity.xact_start age beyond threshold (the Validate MVCC / dead-tuple behaviour under sustained load #169 MVCC-discipline alert promised in that issue's playbook)
• awa_maintenance_rotate_attempts_total{awa_ring_outcome="skipped_busy"} sustained elevation (producers outpacing consumers)
• awa_maintenance_prune_attempts_total{awa_ring_outcome="blocked"} non-zero (held-tx blocking reclaim)
• awa_queue_lag_seconds p95 vs SLO
• DLQ depth delta over window
• awa_maintenance_branch_overrun_total non-zero (also the telemetry gate for maintenance: split rescue + promote into a separate tokio task (v0.7) #303)
• backpressure signals once Backpressure policy: queue depth limits and producer-side flow control #341 lands

Each rule documented with meaning, likely cause, and the matching awa doctor check / troubleshooting entry.

Acceptance

• Rules import cleanly into a stock Prometheus/Grafana stack; validated in CI (syntax/lint at minimum).
• Cross-referenced from the upgrade guide and operations handbook so prose watch-lists point at the shipped rules.

References

Roadmap: docs/0.7-roadmap.md (WS-4, NEW-12). Related: #169 playbook, #302/#303, #341, docs/upgrade-0.5-to-0.6.md watch-list.