GCP operator verbs live proof - 2026-09-15 (mldp8)

Five-verb packed session for order 12 (operator-finops-catalog): one GKE Autopilot preview, 13/13 acceptance cells PASS, then health, logs, exec (plus the --service deploy rejection), cost, and cleanup reconcile exercised live against the running stack. The session also live-proved the destroy-after-expiry fix: an expired YAML refused deploy and completed destroy. AWS cost --live (Fargate vCPU $17.74, memory $3.87, Valkey $10.51/month) and cost --budget (preview not-configured, staging lists five account budgets) were proven creds-only the same day; those reads fixed five Price List filter defects and a Budgets page-size defect. Do not reuse mldp8 (WIF pool name tombstoned 30d).

Scope

Field	Value
Project	GCP acceptance project (accountRef redacted in sealed JSONL)
Region	europe-west1
Name	mldp8
Profile	preview
Runtime	gke-autopilot
Catalog	scripts/acceptance/cells-gcp-preview.txt (13 cells)
Artifact A	…/magento-249-rc1-static-owned-dbhost-20260816@sha256:8588b13f…fdb2be4
Seed	magento-249-sanitized-definer-free.sql.gz (102 KiB)
KEEP	true during verbs, manual destroy after

False start

The first up attempt inherited a stale MAGELIFT_ACCEPTANCE_CELL_CATALOG pointing at an OVH cell file from order 11 and failed closed on unsupported gcp acceptance cell with zero cloud changes. The operator purged the leaked variables, removed the bogus checkpoint entry, and resumed; all 13 cells below come from the clean resume. The sealed bundle keeps final records only.

Verbs

All five against the live mldp8-preview stack:

• health --mode runtime: healthy, deployment 1/1 ready, Magento HTTP 200.
• logs --service web: real nginx lines, early rollout 500s converging, matching the mldp3 pattern.
• exec --service web -- php -r ...: PHP 8.5.9. exec --service deploy refused with the guided pointer to deploy logs.
• cost: account-free inputs (Autopilot web, Cloud SQL, Memorystore, edge, db queue) plus an explicit unpriced list; no live GCP prices claimed.
• cleanup claim|record|plan|reconcile --yes on a probe ledger: live Cloud SQL inventory ran over ADC, the gone identity resolved, report complete, ledger persisted. (Wiring the GCP cleanup provider into the CLI was part of this order; the hook existed but no binary populated it.)

Expiry

deploy with an expired YAML refused (preview environment has expired); destroy --yes --destroy-backups with the same YAML deleted 22 resources. Pre-fix the destroy died in the Pulumi program; the allow-expired flag now rides the spec in all five providers.

Teardown

Pulumi destroy left the service-networking peering plus the network it blocked (same GCP-side refusal as mldp7). Manual sweep deleted NAT, router, service account, three secrets, media bucket, four subnets, PSA range, and the VPC; deleting the VPC removed the stuck peering with it. Pulumi stack removed, state bucket removed, WIF pool deleted. Final sweep: zero mldp8 resources in every listing.

Spend

Paid resources (Autopilot cluster, Cloud SQL, Memorystore, LB) lived for the run window (~75 min create-to-destroy); no leftovers.

Non-claims

• GCP live unit prices (cost stays account-free there)
• RDS / OpenSearch / Amazon MQ live-price shapes (fixed, await a session that runs them)
• Cloud SQL Delete over a live instance (unit-covered; the probe resolved a gone identity)