📋 Summary

Add Prometheus alert rules and an Alertmanager service so RabbitMQ dead-letter queues — which have no consumer by design — no longer go unwatched.

🎯 Motivation

The audit.log.notification.dlq (and email.notifications.dlq) hold poison messages that fail processing, but nothing observed them: there was no alerting anywhere in the stack. A silently-growing DLQ means lost audit/email events with no signal to anyone.

🔧 Changes

• Add alertmanager service to docker-compose.yml (UI on :9093)
• Wire alerting + rule_files into prometheus.yml
• Add rabbitmq-detailed scrape job hitting /metrics/detailed?family=queue_coarse_metrics — the rabbitmq_prometheus plugin's default /metrics omits per-queue depth, so this is required to see rabbitmq_detailed_queue_messages{queue=...}
• Add alert.rules.yml with 3 rules: DLQ-not-empty (any *.dlq), audit queue backlog, RabbitMQ down
• Add alertmanager.yml with a no-op default receiver + commented Slack/email templates
• Document the setup in README.md (zh/en/ja) and CLAUDE.md

🧪 How to Test

1. docker compose up -d rabbitmq alertmanager prometheus
2. Publish a message to the DLX: POST /api/exchanges/%2f/wallet.audit.logs.dlx/publish (routing key audit.logs.dead) via the RabbitMQ Management API
3. Expected result: within ~1 min RabbitMQDeadLetterQueueNotEmpty shows firing at http://localhost:9090/alerts and active in Alertmanager (http://localhost:9093). Purging the DLQ clears the alert.
   • Verified end-to-end locally: pending → firing → Alertmanager active → auto-resolved after purge.

⚠️ Notes / Risks

• Alertmanager ships with a no-op default receiver — alerts are evaluated and visible in the UIs but not pushed anywhere until a Slack/email receiver is filled in. Intentional (no credentials committed).
• Config-only change; no application code touched.

✅ Checklist

• Tested locally (live smoke test)
• Documentation updated (README + CLAUDE.md)
• No hardcoded secrets or credentials
• Tests written or updated (N/A — infra/config only)

🤖 Generated with Claude Code