fix(worker): carry the dp supervisor review follow-ups

.agents/skills/gpustack-operator-benchmark/SKILL.md

@@ -1,17 +1,17 @@
 ---
 name: gpustack-operator-benchmark
-description: "Benchmark ModelDeployment serving topologies on a GPU cluster: warm NodeModelStore cache, drive in-cluster load across EPD, P/D, Elastic EP, and External DP, and collect stage and transfer evidence."
+description: "Benchmark the operator on a GPU cluster — today the ModelDeployment serving topologies: warm NodeModelStore cache, drive in-cluster load across EPD, P/D, Elastic EP, and External DP, and collect stage and transfer evidence."
 disable-model-invocation: true
 allowed-tools: "Read, Agent, Bash(bash .claude/skills/_e2e-lib/scripts/preflight.sh*), Bash(bash .claude/skills/_e2e-lib/scripts/assert-core.sh*), Bash(bash .claude/skills/_e2e-lib/scripts/cluster-auth.sh*), Bash(bash .claude/skills/_e2e-lib/scripts/kube-context.sh*), Bash(bash .claude/skills/gpustack-operator-benchmark/cases/*), Bash(kubectl get*), Bash(kubectl describe*), Bash(kubectl logs*), Bash(kubectl cluster-info*), Bash(kubectl version*), Bash(kubectl config current-context), Bash(kubectl config get-contexts*), Bash(git diff*), Bash(git rev-parse*), Bash(command -v*), Bash(date*), Bash(mkdir -p .claude/reports/*), Bash(tee .claude/reports/*)"
 ---
 
-# GPUStack Operator — ModelDeployment Serving Benchmark
+# GPUStack Operator — Benchmark
 
 Use [the shared module map](../../../docs/README.md#module-map) to locate the relevant guides,
 owning specs, code entry points and related modules. [AGENTS.md](../../../AGENTS.md#finding-sources)
 owns the retrieval and evidence rules; this skill owns its execution procedure.
 
-Measure and compare the four `ModelDeployment` serving topologies on a real multi-GPU cluster (8 × GPU baseline) and collect both end-to-end latency/throughput distributions (`TTFT`, `ITL`, `TPOT`, `E2E`, `output_tok/s`, `req/s`) and per-stage execution / KV-transfer / EC-transfer evidence:
+Measure and compare the `ModelDeployment` serving scenarios on a real multi-GPU cluster (8 × GPU baseline) and collect both end-to-end latency/throughput distributions (`TTFT`, `ITL`, `TPOT`, `E2E`, `output_tok/s`, `req/s`) and per-stage execution / KV-transfer / EC-transfer evidence. The scenarios below are the suite's first five; a scenario outside serving joins this suite, not a new skill:
 
 - **Scenario 1 (`cases/case-1-epd.sh`)** — [Encode, Prefill and Decode](../../../docs/walkthroughs/model-deployment/encode-prefill-decode.md) ([benchmark](../../../docs/modules/model-deployment/encode-prefill-decode-benchmark.md)): 3-arm comparison (`Aggregated` vs `P/D` vs `E/P/D`) on a 3-GPU budget with `Qwen/Qwen2.5-VL-7B-Instruct` (`NIXL` encoder cache + `Mooncake` KV cache).
 - **Scenario 2 (`cases/case-2-prefill-decode.sh`)** — [Prefill and Decode](../../../docs/walkthroughs/model-deployment/prefill-decode.md) ([benchmark](../../../docs/modules/model-deployment/prefill-decode-benchmark.md)): 2-arm comparison (`Aggregated` 3 × `Server(TP=2)` vs `P/D` 1 × `Prefill(TP=2)` + 2 × `Decode(TP=2)`) on a 6-GPU budget with `Qwen/Qwen3-32B`.

.agents/skills/gpustack-operator-e2e/cases/case-100.sh

@@ -2,7 +2,7 @@
 #
 # CASE 100 — The weight identity in the KV store key: one artifact's deployments share blocks
 #            through the store, and a deployment of another commit on the same Binding reads none
-#            (MUTATING, cluster-scoped store objects; needs real accelerators)
+#            (MUTATING, self-cleaning: cluster-scoped store objects; needs real accelerators)
 #
 #   case-100.sh <NS>
 #

.agents/skills/gpustack-operator-e2e/cases/case-116.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 #
 # CASE 116 — Every encoder-role admission refusal fires, each naming what its rule refuses
-#   (MUTATING for the fixture only: one settings key and one injected node label, both restored by
+#   (MUTATING, self-recovering: one settings key and one injected node label, both restored by
 #    the trap; every verdict below is a server-side dry run that persists nothing)
 #   AUTO-SKIPS its matrix rows without an InstanceType that reports an admitted accelerator
 #   platform, which this case builds by injecting a CPU vendor label.

.agents/skills/gpustack-operator-e2e/cases/case-123.sh

@@ -2,7 +2,8 @@
 #
 # CASE 123 — vllm-router admits vLLM encoder deployments on NVIDIA and AMD, refuses unadmitted
 #           engines and user-supplied --epd-config
-#           (NON-MUTATING on GPU clusters; fixture-mutating on CPU-only clusters, restored by trap)
+#           (MUTATING, self-recovering: the platform fixture on CPU-only clusters, restored by
+#            the trap — GPU clusters need no fixture, and every verdict is a server-side dry run)
 #           AUTO-SKIPS the platform rows without an InstanceType reporting nvidia or amd
 #
 #   case-123.sh <NS>

.agents/skills/gpustack-operator-e2e/cases/case-18.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 #
-# CASE 18 — CPU-manufacturer awareness reshapes the catalog   (MUTATING: flips an editable setting, self-restoring)
+# CASE 18 — CPU-manufacturer awareness reshapes the catalog   (MUTATING, self-recovering: flips an editable setting)
 #
 #   case-18.sh <NS>
 #

.agents/skills/gpustack-operator-e2e/cases/case-42.sh

@@ -51,6 +51,9 @@
 #              one-slice partitions requested together, then one two-accelerator request expected to
 #              be refused.
 # Verdicts:    One row per contract above, plus the setup gates. Any FAIL fails the case.
+# Expected:    Every contract in the Goal holds from both sides — the operator's records and the
+#              container or the vendor's own registry — and the table names both sides per row.
+# Cleanup:     An EXIT trap deletes the case's Pods and waits for their release.
 set -uo pipefail
 
 CASE_ID=42

.agents/skills/gpustack-operator-e2e/cases/case-45.sh

@@ -2,8 +2,8 @@
 #
 # CASE 45 — The ModelDeployment admission surface: every refusal fires, and each one fires from the
 #           layer that is supposed to own it
-#   (NON-MUTATING for the refusals, which are server-side dry-runs; one short-lived object for the
-#    controller-level check)
+#   (MUTATING, self-cleaning: the refusals are server-side dry-runs; one short-lived object
+#    serves the controller-level check)
 #
 #   case-45.sh <NS>
 #
@@ -74,6 +74,9 @@
 #              the behaviour is missing, but that asserting it would not DISCRIMINATE. See the row
 #              about what the unregistered domain costs, near the end.
 #
+# Expected:    Every refusal fires from the layer the Goal names, the field and message asserted
+#              per row; the created-object row observes creation before it judges.
+# Cleanup:     An EXIT trap deletes the one short-lived deployment (case45-nobind).
 set -o pipefail
 
 NS="${1:-}"

.agents/skills/gpustack-operator-e2e/cases/case-47.sh

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 #
 # CASE 47 — Two deployments on one Binding share a reuse domain; two on different Bindings do not
-#   (MUTATING: three ModelDeployments, one extra Binding and one probe Pod, all in the fixture's
-#    namespace, all removed by the shared teardown)
+#   (MUTATING, self-cleaning: three ModelDeployments, one extra Binding and one probe Pod, all
+#    removed by the shared teardown)
 #
 #   case-47.sh <NS>
 #
@@ -43,6 +43,12 @@
 #              image from. No GPU, no engine image.
 # Needs:       a live Mooncake store (the shared fixture brings up backend, pool and Bindings)
 #
+# Inputs:      All real, nothing mocked: the fixture's pool and Bindings, one additional Binding in
+#              the same namespace, and the engine image the deployments render.
+# Expected:    Deployments on the same Binding render the same reuse domain and the third
+#              deployment a different one; the probe writes land under the rendered values.
+# Cleanup:     The EXIT teardown deletes the three deployments, the extra Binding and the probe
+#              Pod; the shared fixture's teardown covers the rest.
 set -uo pipefail
 
 NS="${1:?usage: case-47.sh <NS>}"

.agents/skills/gpustack-operator-e2e/cases/case-48.sh

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 #
 # CASE 48 — The deliberate break: what a deployment looks like when the cache cannot work
-#   (MUTATING: one external KVCacheBackend and one KVCachePool, both CLUSTER-SCOPED, plus one
-#    KVCachePoolBinding and two ModelDeployments in the fixture's namespace — all removed here)
+#   (MUTATING, self-cleaning: one external KVCacheBackend and one KVCachePool, both
+#    cluster-scoped, plus one KVCachePoolBinding and two ModelDeployments — all removed here)
 #
 #   case-48.sh <NS>
 #
@@ -47,6 +47,13 @@
 # Needs:       a live Mooncake store (the shared fixture brings up backend, pool and Bindings) and an
 #              InstanceType, for the role to reference
 #
+# Inputs:      All real, nothing mocked: the fixture's pool and Bindings beside the case's own
+#              unreachable pool, and the engine image the deployments render.
+# Expected:    The unreachable pool renders no connector onto its deployment while the reachable
+#              pool's image still gets its render; every CacheAttached row pairs the condition
+#              with a fact not true of everything.
+# Cleanup:     The EXIT teardown deletes the namespace objects and both cluster-scoped objects
+#              the case created.
 set -uo pipefail
 
 NS="${1:?usage: case-48.sh <NS>}"

.agents/skills/gpustack-operator-e2e/cases/case-51.sh

@@ -2,8 +2,8 @@
 #
 # CASE 51 — Every P/D refusal fires, from the layer that owns it, and a replicas change replaces
 #           nobody
-#   (NON-MUTATING for the refusals, which are server-side dry-runs; one short-lived deployment for
-#    the convergence rows)
+#   (MUTATING, self-cleaning: the refusals are server-side dry-runs; one short-lived deployment
+#    serves the convergence rows)
 #
 #   case-51.sh <NS>
 #
@@ -43,6 +43,9 @@
 #              and whether a short pool leaves both roles queued is case-50's. This file stops at
 #              the operator's own writes, which is what it can assert without an engine.
 #
+# Expected:    Each refusal names its owning layer and message; the replicas change converges
+#              without replacing the replica already running.
+# Cleanup:     An EXIT trap deletes the short-lived deployment and its groups' Workloads.
 set -o pipefail
 
 NS="${1:-}"

.agents/skills/gpustack-operator-e2e/cases/case-68.sh

@@ -83,6 +83,8 @@
 # against one object would be asserting one of them against a state that cannot produce it. Phase E
 # needs a THIRD object for the same reason: Phase C deactivates the deployment it runs on, and a
 # deactivated workload is never admitted again whatever the cluster does.
+# Cleanup:     An EXIT trap deletes the three ModelDeployments and the held InstanceType the case
+#              created.
 set -uo pipefail
 
 # Route every kubectl through the retrying shim. Against a remote API endpoint a read can fail on

.agents/skills/gpustack-operator-e2e/cases/case-70.sh

@@ -6,6 +6,9 @@
 #
 #   case-70.sh <NS>
 #
+# Goal:        A routed deployment owns exactly one complete router object set through both
+#              router transitions, and its status reports every cache-event publication state
+#              observable from the cluster.
 # Environment: A single-node cluster with no accelerator, the operator deployed, the general
 #              InstanceType materialized by CASE 1, and pull access to the Mooncake fixture image.
 #              No serving engine image is required: the assertions read rendered objects, Pod
@@ -18,6 +21,11 @@
 #              garbage collection remove all six. The deployment remains Starting with no ready
 #              role replicas. KVEventsPublishing reaches NotApplicable, NoRouter,
 #              PublisherDisabled, RoleUnmanaged and Publishing, with status and reason asserted.
+# Inputs:      All real: the operator's charts and images, the Mooncake fixture image and a Ready
+#              Binding. No serving engine image — the assertions read rendered objects, Pod
+#              specifications and status.
+# Cleanup:     The teardown releases any finalizer-held Pods it created and deletes the lifecycle
+#              objects.
 set -uo pipefail
 
 E2E_SHIM_DIR="$(cd "$(dirname "$0")/../../_e2e-lib/scripts/kubectl-shim" 2>/dev/null && pwd)"

.agents/skills/gpustack-operator-e2e/cases/case-71.sh

@@ -8,6 +8,16 @@
 # This case is separate from CASE 70 because the upstream router images may be unavailable. An image
 # pull failure is a stated SKIP. Any other failure to become Ready is a FAIL. No accelerator or
 # serving engine is needed: the router runs outside Kueue and discovers the role Pods dynamically.
+# Goal:        The router endpoint publishes only once the router Deployment has a ready replica —
+#              the readiness contract status.endpoint carries.
+# Environment: A single-node cluster with no accelerator, the operator deployed, the general
+#              InstanceType from CASE 1, and pull access to a router image.
+# Inputs:      All real: the router image is pulled from its registry; no accelerator and no
+#              serving engine image is needed.
+# Expected:    With the image pullable, the deployment publishes its endpoint once the router
+#              replica is Ready and RouterReady reads True/Ready. An image pull failure is the
+#              stated SKIP; any other failure to become Ready is a FAIL.
+# Cleanup:     An EXIT trap deletes the deployment and any Pods left under it.
 set -uo pipefail
 
 E2E_SHIM_DIR="$(cd "$(dirname "$0")/../../_e2e-lib/scripts/kubectl-shim" 2>/dev/null && pwd)"

.agents/skills/gpustack-operator-e2e/cases/case-89.sh

@@ -1,6 +1,7 @@
 #!/usr/bin/env bash
 #
 # CASE 89 — A ModelDeployment role's rendered Pod requests the selected fabric device.
+#           (NON-MUTATING; E2E_REJECT_MANIFEST, when set, is a server-side dry run)
 #
 # case-89.sh <NS> <MODEL_DEPLOYMENT> <ROLE> <COUNT> <KEY|none>
 #

.agents/skills/gpustack-operator-e2e/cases/case-90.sh

@@ -1,6 +1,7 @@
 #!/usr/bin/env bash
 #
 # CASE 90 — A ModelDeployment metrics snapshot reflects served traffic and Pod endpoints.
+#           (MUTATING, self-cleaning: one temporary probe Pod, deleted by the trap)
 #
 # case-90.sh <NS> <MODEL_DEPLOYMENT>
 #

.agents/skills/gpustack-operator-e2e/cases/case-91.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 #
 # CASE 91 — A routed prefill/decode pair moves real KV bytes over its direct leg, and writes the
-#           store when it has one.
+#           store when it has one.   (NON-MUTATING)
 #
 # case-91.sh <NS> <MODEL_DEPLOYMENT>
 #

docs/modules/model-deployment/routing.md

@@ -34,7 +34,7 @@ pairing:
 | Mode | `llm-d-router` | `vllm-router` | `sglang-gateway` |
 |---|---|---|---|
 | Multi-replica *Server* | One fixed scoring profile — prefix match, queue depth, KV utilization — with no policy switch; measured 187/170 on vLLM and 375/0, every request on one replica, on SGLang | `cache_aware` by default; `--policy round_robin` spreads evenly, measured 60 per replica | The same policies as `vllm-router`; `round_robin` measured 60 per replica |
-| *PD* pair | The disaggregated pairing's own router: the token-counted prefill/decode split, per-role connector documents, the decode sidecar | Its own *PD* mode over the pair | Native `--pd-disaggregation` |
+| *PD* pair | The disaggregated pairing's own router: the token-counted prefill/decode split, per-role connector documents, the decode sidecar | Its own *PD* mode over the pair | Native `--pd-disaggregation`; combining it with `--dp-aware` produced an invalid *decode* destination in the recorded container run, so the operator never renders the two together |
 | *EPD* | An `encode` scheduling profile beside the pair; the decode sidecar coordinates the encoder transfer | Drives encode, prefill and decode itself through operator-owned `--epd-config`, no sidecar | Routes the pair natively; the prefiller reaches the encoder through `--encoder-urls` |
 | *Elastic EP* | The engine's own internal load balancing spreads work inside the collective, and the deployment publishes one endpoint, so a router is not how this mode balances | Same | Same |
 | *External DP* | Admitted, but no `--policy` to spread the ranks evenly and no recorded run — use the engine's own router | Every rank endpoint behind `--policy round_robin` (recorded) | Per-member endpoints; the registration boundary is the paragraph below |
@@ -63,10 +63,10 @@ carries a rank suffix (`http://host:port@0`, `http://host:port@1`), and every su
 member's one physical serving listener — the suffix is a rank, not a port.
 
 That registration is the gateway patch this project ships: an image carrying it is
-`gpustack/llm-router:v0.4.0` or later, while the default
-[`model-deployment-router-image`](../../reference/settings.md) still names `v0.3.1`, which
-predates the patch. The registration activates only where *`spec.router.image`* or the setting
-names an image that carries the patch.
+`gpustack/llm-router:v0.4.0` or later, which is also the default
+[`model-deployment-router-image`](../../reference/settings.md), so a deployment that names no
+image of its own registers per rank out of the box. An older default or a named image that
+predates the patch runs the gateway without the fan-out.
 
 No runtime run has exercised the flag: the render gate and the patched fan-out are verified in
 source and unit tests on the recorded shape.