docs(devops): fix-forward-in-git rule — ban escalating reconcilable changes as manual/board actions (GRO-2536)

Adds "fix forward in git" rule to devops skill: any breakage a controller (Flux/OpenTofu/SealedSecrets) can reconcile must be resolved via PR to groombook/infra — not board approval, not hand-run kubectl. Prevents recurrence of GRO-2536 stall.

Co-Authored-By: Paperclip <noreply@paperclip.ing>
Co-authored-by: Scrubs McBarkley <18+gb_scrubs@noreply.git.farh.net>
Co-committed-by: Scrubs McBarkley <18+gb_scrubs@noreply.git.farh.net>

skills/devops/SKILL.md

@@ -59,6 +59,24 @@ Images currently use `:latest` with `imagePullPolicy: Always`; pin to a CalVer t
 
 **Policy — Flux Image Tag Automation is DENIED.** Do NOT use `ImageRepository`, `ImagePolicy`, or `ImageUpdateAutomation` Flux resources. Image tag updates must be made intentionally via a PR to `groombook/infra` — typically as the final step of the `sdlc` application pipeline (Phase 5).
 
+## When a cluster is broken: fix forward in git — never escalate a manual action
+
+The cluster is reconciled by controllers (Flux, the OpenTofu Controller, the Sealed Secrets controller). **Any change one of these controllers can reconcile MUST be delivered as a PR to `groombook/infra`** — it is never a board approval and never a hand-run `kubectl` / `kubeseal` / `tofu` command.
+
+This is the corollary of the read-only-prod and "no `kubectl apply` to production" rules in `safety`: agents are read-only on `groombook` **by design**, precisely because the write path is git. "I lack cluster-admin" therefore resolves to **"open a PR,"** not **"ask a human to run the command."**
+
+Contract:
+
+- **Do NOT** file an issue, board approval, or escalation that asks a human to run an imperative cluster command (`kubectl delete/apply/patch`, `kubeseal`, `flux reconcile`, `tofu apply`) that a controller would otherwise reconcile from git. That request is unfillable and wrong on a GitOps cluster — fix the desired state in the repo and let the controller converge.
+  - SealedSecret won't unseal / wrong scope → re-seal the `SealedSecret` and commit it.
+  - Missing or not-ready Flux `Receiver`, `Kustomization`, `Terraform`, RBAC, etc. → commit/correct the manifest in the overlay.
+  - Stale or wrong `sourceRef`, annotations, ownership → fix them declaratively in the overlay.
+- **A reconcile blocked on a pre-existing in-cluster object** (e.g. a `SealedSecret` the controller won't adopt because an unmanaged or Reflector-mirrored `Secret` already exists) is still solved declaratively: correct ownership/annotations in git so the controller adopts it. Only if **no controller can adopt the object** is a one-time imperative step justified — and then it is a single, specifically-scoped, reviewed exception stating the exact reason, **not** a multi-day approval queue standing in for missing engineering.
+- **Board approval is reserved** for genuinely irreversible or out-of-band actions no controller reconciles — destroying stateful data, rotating the cluster bootstrap, bootstrapping a brand-new cluster. Routine reconcilable breakage never qualifies. (See `safety` for destructive-action approval.)
+- The Flux bootstrap/cluster repo is **not** `groombook/infra` (see GitOps above). A genuinely missing `GitRepository` or other bootstrap object is a PR to that externally-managed cluster-config repo — still a PR, still not a hand-run apply.
+
+If you are about to write "escalated to board — a human must run …" for a reconcilable change, stop: that is the failure mode, not the fix. Open the PR.
+
 ## Infrastructure as Code
 
 Terraform (OpenTofu) is deployed via the **Flux OpenTofu Controller** in a GitOps fashion. Submit Terraform configurations via a PR to `groombook/infra` — the tofu controller reconciles them on merge. See `safety` for the prohibition on running `tofu` directly and on `kubectl apply` against production.