# 21_percona_eks_hardened#3 — FAILED | field | value | |---|---| | campaign | COVERAGE (mongodb × aws) | | archetype | generative | | date | 2026-10-15T18:57Z | | gate version | v1 | | deployment | dep-aea61518 | | deploy status | failed (4217s) | | gate | healthz=— connection=— logs=— metrics=— | | teardown clean | True | ## Attribution - **bucket:** AGENT - **class:** kubernetes-manifest-beside-its-own-cluster - **evidence:** Five plan failures in a row on a kubernetes_manifest resource placed beside the cluster the same configuration creates (terraform cannot construct the REST client before the cluster exists) — a construct the agent's own reference material prescribed at the time. Attempt 2 switched to kubectl_manifest and the cluster came up; the operator's Helm release then fought the Terraform-created namespace over ownership metadata, flipped create_namespace, and died 'namespaces mongodb already exists' three times (attempts 3-5) until the budget capped. - **resolution:** Fixed (75c8333d, 2911201d): the Percona manifest is a kubectl_manifest in the exemplars, a pre-init gate and a plan-stage invariant refuse the kubernetes_manifest-beside-cluster shape, and the release must adopt the Terraform namespace, never create it. The repeat guards (reference-suspect block, gate self-demotion) were built from this slot. Scenario 21 passed on its 30 Oct re-run. ## Recorded reason (scrubbed) ``` ladder exhausted (auto-repair gave up) ```