[rambohe-ch]

Reason for Change:

Closes two data-plane E2E gaps against the data-plane error-handling proposal:

• 502 gateway_unavailable (outage local_reply mapper feat: add production stack arch contents #3) had no E2E coverage.
• The non-JSON body case accepted any status from 400 to 500, so it could not catch a regression to a misleading 404 or a 5xx.

 test/e2e/
+├── gateway_unavailable_spec.go   # new: mapper #3
 ├── cases.go                      # new case: gateway-unavailable
 ├── model_routing_spec.go         # non-JSON body: 400..500 -> exactly 400
 └── utils/ginkgo.go               # Outage label doc lists the new spec

gateway_unavailable needs a failure before model routing (no X-Gateway-Model-Name) with a router flag. Scaling the namespace Gateway itself to zero cannot produce that, because no Envoy is left to answer. Model discovery is the request path that reaches an upstream without model routing:

GET /v1/models                      # bodyless: BBR injects no model header
  Gateway route models-api-list
    upstream: productionstack-status-reporter   # scaled to 0
    -> local 503, flag UH
  outage local_reply mapper #3      # header absent + router flag
    -> 502 gateway_unavailable, x-kaito-error-source: gateway

The spec:

• is a regular case (gateway-unavailable) with one mocked model. The model is never requested, but on an App Routing cluster the Gateway hostname is published to DNS from the namespace's HTTPRoute hostnames, and only a modeldeployment renders an HTTPRoute;
• finds the reporter Deployment by its app.kubernetes.io/name label rather than a fixed namespace and name;
• restores the original replica count on cleanup;
• is Serial, because the reporter is a cluster-wide singleton.

Per #262 it does not carry StandardK8sOnly: scaling a shared Deployment works on a managed control plane.

For a non-JSON body, BBR answers with its own plain-text 400 (an ext_proc immediate response), which the gateway passes through without the error envelope. The test now asserts exactly 400 and records the raw response (status, content-type, x-kaito-error-source, body) with AddReportEntry, so the body is visible in the report without asserting its shape.

Requirements

• added unit tests and e2e tests (if applicable).

Issue Fixed:

Notes for Reviewers:

Evidence

The new specs have not been run on a live cluster yet; they run in the PR e2e workflow (!Nightly && !Perf && !Scaling). Local checks:

• go vet ./test/e2e/... (with -mod=readonly): passed.
• gofmt -l test/e2e: no output.
• make verify-boilerplate: passed.
• codespell on the changed files: no findings.
• ginkgo --dry-run --label-filter='Outage && ModelDiscovery' ./test/e2e: selects exactly the 2 new specs.

Merge Danger

Door: two-way. Test-only change, no chart or code change.

Blast Radius: e2e-runtime. The new spec briefly scales the status reporter to zero, which pauses model discovery and control-plane Events cluster-wide; Serial keeps every other spec from overlapping it.

The companion proposal update is in a separate PR.

[rambohe-ch]

Companion proposal update: #264 (its Test Plan references the specs added here).