perf: Reduce pgwatch log verbosity to error level

Closes #102 (closed)

Summary

• pgwatch was logging every metric fetch at INFO level, generating ~80% of all logs in postgres-ai-mon-production namespace
• With 4 database sources × ~25 metrics × 30-second intervals = ~12,000 log lines per hour of noise

Changes

• Add --log-level=warn to both pgwatch services in docker-compose.yml
• Add configurable logLevel (default: warn) to Helm values.yaml
• Update pgwatch-postgres and pgwatch-prometheus deployments to use the config

Evidence

$ gcloud logging read ... | sort | uniq -c | sort -nr
8138 pgwatch       <-- 81% of logs
1509 grafana
 306 postgres-exporter
  47 victoriametrics

Sample logs (pure noise):

[INFO] [source:prod_replica_postgres_ai] [metric:replication] [rows:0] measurements fetched
[INFO] [source:prod_replica_postgres_ai] [metric:pg_stat_replication] [rows:0] measurements fetched
...repeated every 30 seconds for every metric...

Test Plan

• Deploy to staging and verify pgwatch only logs warnings/errors
• Verify metrics collection still works (check Grafana dashboards)
• Monitor log volume reduction in GCP Logging

🤖 Generated with Claude Code