feat: degrade gracefully when no Kubernetes cluster is present

[mattDev0]

Implements Option 3 of adr_01, which is now marked Accepted.

The problem

Production moved to Docker Compose on 2026-07-14, but the Kubernetes half of the dashboard was never adjusted to match. To stop the agent blocking at startup, scripts/deploy.sh fabricated a kubeconfig pointing at https://localhost:8443.

That hid the problem rather than solving it, because Client::try_default() only parses configuration — it never contacts a cluster. So the agent reported:

{"status": "healthy", "k8s": true}

…against a cluster that does not exist. Downstream, the deployments table, pod-health panel and log streaming all queried nothing, and the SLO tiles displayed 100% availability and a full error budget for it. scripts/health-check.sh even carried a comment acknowledging /health returns 503 "when Kubernetes is not configured".

Approach

Keep Compose as the runtime, keep the manifests as a supported alternative target, and remove the dishonest parts rather than the capability.

Agent

• Verify before advertising. A real API call now confirms the server answers before Kubernetes is reported as available.
• Bounded startup. try_connect() replaces an unbounded backoff loop that could never let the process start without a cluster. Three attempts, then continue without one.
• AppState.kube_client becomes Option<Client>. Cluster-backed routes return 503, not 500 — an absent dependency is not a defect.
• /health returns 200 with k8s: false instead of 503. Running without a cluster is a supported configuration. This also fixes a real bug: the old 503 made the orchestrator raise AgentUnreachableException for a perfectly healthy agent.
• The deployment monitor only starts when a cluster is available.

Frontend

• Kubernetes panels are replaced by an explicit "no cluster connected" state.
• The pod / availability-SLI / error-budget tiles render — instead of 0 Pods and 100.0%.
• Cluster fetches are skipped entirely when the agent reports no cluster, so no pointless 503s.

Infrastructure

• Placeholder kubeconfig generation removed from deploy.sh (23 lines) along with its compose mount.
• Under infrastructure/k8s/ the service account supplies credentials, so the clustered path is unchanged — deploying the manifests restores every cluster feature automatically.

Validation

apps/agent          cargo test    5 passed (2 new)
apps/agent          cargo clippy -- -D warnings    clean
apps/orchestrator   mvn test      20 passed
apps/frontend       eslint        0 errors, 1 pre-existing warning
apps/frontend       vite build    OK
docker-compose*.yml YAML parse    OK

New tests:

• test_cluster_routes_report_unavailable_without_kubernetes — /deployments and /pods/health return 503 with no cluster.
• test_health_is_ok_and_honest_without_kubernetes — /health returns 200 with k8s: false.

Docs

• adr_01 moves from Proposed to Accepted, with the decision and the reasoning against Options 1 and 2 recorded. The three options are retained.
• .agents/rules/engineering-rules.md no longer claims the repos share a cluster, ingress and cloud.
• README.md: corrected the claims that the agent "runs as a Kubernetes pod" and queries K3s, and marked cluster-dependent features accordingly.

Deployment note

No secrets or deployment credentials change. On merge the agent will log Starting without Kubernetes... and report k8s: false — that is the expected, correct output for this environment, not a regression.