gcp-systems-cost-reduction overview

End-to-end GCP cost-reduction. All tasks are `kind: 'agent'`; agents run `gcloud`/`bq`/`jq` via Bash and own their scratch dirs under `tasks/<effectId>/artifacts/`. NO inter-task filesystem contract — data flows as JSON results + explicit artifact paths. The single user-controlled path is `inputs.outputDir`, where the final report lands. Tiers: 1 cosmetic (dangling DNS, missing labels) / 2 rightsize (GCE machine-type, pd-ssd->pd-balanced, GCS Standard->Nearline->Coldline-> Archive, HA Cloud SQL->zonal in dev) / 3 CUDs 1y/3y IRREVERSIBLE / 4 decommission with backup (orphan static IPs $7.30/mo, unattached PDs, snapshots of deleted VMs, TERMINATED GCE with attached disks billing, idle Cloud SQL) / 5 cluster (shrink GKE node pool, delete dev cluster, Autopilot switch, Cloud Run min-instances=0). GCP waste patterns: orphan reserved static IPs (huge sleeper), unattached PDs, snapshots of deleted VMs, TERMINATED GCE with attached disks, GCS Standard on cold objects, Cloud SQL HA in dev (2x cost), GKE node pool min>0 in dev, Pub/Sub subs with no subscribers (msgs accumulate, billed).