Merge pull request 'docs(memory): record the WSP-32957 PIM migration project state' (#104) from docs/memory-wsp-32957-pim-migration into main

Reviewed-on: #104

home/claude/memory/MEMORY.md

@@ -15,3 +15,4 @@
 - [Dev clusters disposable](dev_clusters_disposable.md) — Lyra's dev clusters are recreatable; mutate/break freely, no confirmation needed
 - [Nix shell tooling](nix_shell_tooling.md) — any nixpkgs tool runs ad hoc via `nix run`/`nix shell nixpkgs#<pkg>`; a missing command is never a dead end
 - [WSP local build and test](wsp_local_build_and_test.md) — core-services-cloud on this box: dotnet via nix, artifactory creds from `~/.artifactoryenv` sourced per command, how to tell auth failure from a code failure
+- [WSP-32957 PIM migration](wsp_32957_pim_migration.md) — AKS RBAC to PIM + AutoPerm decommission; prod is in the Technical Preview subscription, three tenants; resume via `~/code/WSP-32957-CONTINUATION.md`

home/claude/memory/wsp_32957_pim_migration.md

@@ -0,0 +1,58 @@
+---
+name: wsp-32957-pim-migration
+description: State of the WSP AKS-RBAC-to-PIM migration and AutoPerm decommission, and how to resume it
+metadata:
+  node_type: memory
+  type: project
+---
+
+Long-running epic (August 2026) moving WSP's AKS cluster RBAC off **AutoPerm
+Manager**-maintained `wsp-*` groups onto the per-subscription **`CEO-*` Azure PIM**
+groups, then retiring AutoPerm. Epic **WSP-32957**; on the critical path for Zensar
+L1 on-call (WSP-32193). Work happens in **`~/code/multicluster`**
+(`terraform/cluster`, `terraform/cluster-k8s-resources`, `products/*/environments/*`).
+
+**Resume from `~/code/WSP-32957-CONTINUATION.md`** — full state, branch list,
+verified object IDs, findings and next steps. Jira is the durable record; that file
+is the index. Keep it updated as work lands ([[docs-keep-updated]]).
+
+**Landed:** WSP-33141 (multicluster PR #1808, merged `b047163c`) added object-ID
+inputs — `admin_group_oids` on `cluster`, and `cluster_user_group_oids` /
+`cluster_viewer_group_oids` / `cluster_superuser_group_oids` on
+`cluster-k8s-resources`. Supplying IDs _replaces_ the display-name lookup and leaves
+the `data "azuread_group"` unread, which is what will let the legacy groups be
+deleted. Nothing sets them yet, so behaviour is unchanged. Repoint branches for test
+(WSP-33067) and staging (WSP-33068) are pushed but have **no PR** — both gated on
+decisions, not code.
+
+**Facts that cost real effort to establish, do not re-derive:**
+
+- **Production runs in `fc7af6ae-…` (_Workspace Platform Technical Preview_), not
+  `d6d75d07-…` (_Workspace Platform Production_)**, which holds no clusters. The
+  epic was wrong about this for its whole life and every production `CEO-*` group
+  name and object ID had to change. Because `CEO-*` names embed the subscription
+  name, **always re-verify object IDs against live Entra rather than trusting the
+  epic table.**
+- **Three tenants**, not two: `6f4fe054` (prod, prod JP), `335836de` (staging,
+  staging JP, test), `3eae2746` (dev). Each `wsp-*` name is a distinct object in
+  each tenant.
+- `wsp-staging-cluster-admins` and `wsp-test-cluster-admins` are **empty**, so
+  `wsp-owner` is the _only_ path to `cluster-admin` in staging. Never drop it before
+  `SuperAdmin-*` is proven — hence the staging branch is split into an additive
+  commit and a cutover commit.
+- **No break-glass exists**: `disableLocalAccounts = True` on every cluster; only the
+  deployment SPNs authenticate non-interactively.
+- Graph **PIM-for-Groups is unreadable via `az`** (the CLI's first-party client lacks
+  `PrivilegedAccess.Read.AzureADGroup`, on both `v1.0` and `beta`). Use the portal or
+  `Connect-MgGraph -Scopes PrivilegedAccess.Read.AzureADGroup`.
+
+**Watch for:** the epic gets rewritten by James Weldrake between sessions — re-read
+the description before acting, and check which child tickets are still live
+(WSP-33062/33063/33064/33066 were cancelled 2026-08-24, and dev was put out of
+scope). Verified findings have repeatedly contradicted the epic text
+([[copilot-review-false-positives]] is the same instinct: check against reality
+first).
+
+Queued Slack messages and the leaver report live as `~/code/*.txt` alongside the
+continuation file; see the table in it for what has and has not been sent
+([[workflow-review-and-comments]] — show them before they go out).