[rambohe-ch]

Reason for Change:

The data-plane error-handling proposal has drifted from what charts/modelharness, charts/productionstack and the E2E suite implement. This PR updates the proposal to match the implementation; no code changes. The control-plane proposal already matches the reporter and is unchanged.

 §2.1 / §2.2  outage local_reply
-  cluster-wide local_reply in charts/productionstack
+  one per-namespace local_reply in charts/modelharness (mappers #1-#6)
-  bbr_unavailable twice (cluster + per-namespace defence-in-depth)
+  bbr_unavailable once (mapper #5)
-  500 mesh_config_invalid            # MeshConfig wiring no longer exists
 §2.2  invalid_request_body
-  not JSON | not chat schema | missing model  -> local_reply envelope
+  JSON without a model field -> model-not-found-direct direct_response
+  exception: non-JSON body -> BBR native 400, passed through (no envelope)
 §2.3  catch-all, model header absent
-  502 bbr_unavailable
+  400 invalid_request_body
 §2.3 / §3  cluster-filter HA
-  Envoy active health_checks + outlier detection on BBR and ext_authz
+  readiness probe -> EDS (active) on both; outlier detection on BBR only
+  ext_authz: replicas + readiness, short 502 window accepted
 Test Plan
-  *_test.go names, Gateway pod scale-down, ext_authz HA, upstream timeout
+  real *_spec.go names + Status column; gateway_unavailable via reporter

Why the proposal changes rather than the code:

• Cluster-wide local_reply: each inference Gateway has its own HTTP connection manager, and the MeshConfig-based cluster-wide wiring was removed because the AKS Istio add-on does not allow modifying the shared MeshConfig. A second filter on the same HCM would also bring back the non-deterministic EnvoyFilter merge order.
• Non-JSON body: the only local_reply match for this case (a local 400 without the model header) also catches the missing-model direct_response and Envoy's protocol-level 400s, misattributing the latter to bbr. OpenAI SDK clients never send a non-JSON body.
• Envoy health_checks: the BBR and ext_authz clusters are generated by Istio, and BBR serves gRPC health only on a dedicated port. Readiness probes reach the same result through EDS.

Remaining gaps are kept explicit: the llm-gateway-auth 401/403 envelope stays a cross-repo TBD, and the Test Plan marks the scenarios not covered yet. Status moves from provisional to implementable.

Requirements

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

Issue Fixed: