fix(secrets): populate the empty sops placeholders and drop the unused one

The reasoning engine starves every other request while one long prompt
prefills. Measured: 262 tok/s contended prefill, ~250s for a 65k prompt,
with Avg generation throughput 0.0-0.1 tok/s at Running: 2-4 reqs.

Root cause is NOT what the issue first diagnosed. The issue proposed
--max-num-partial-prefills / --max-long-partial-prefills, which are V0-only
knobs. Verified in the live container: vLLM 0.11.0 ships V1 only (vllm/core/
is absent, so there is no V0 to fall back to), and EngineArgs.
_is_v1_supported_oracle returns False whenever either flag differs from its
default, calling _raise_or_fallback("Concurrent Partial Prefill") ->
"Falling back to V0" -> AssertionError. With strategy Recreate that is an
outage, not a slow rollout. Both flags are therefore NOT set, and a comment
records why so they are not reintroduced.

V1 needs no such flag: it has no prefill/decode phase at all. Every step it
walks self.running and hands out tokens from one shared budget
(vllm/v1/core/sched/scheduler.py). The real defect is that the per-request
cap is disabled by default (long_prefill_token_threshold=0), so one long
prompt takes the WHOLE budget, token_budget hits 0, the `while ... and
token_budget > 0` loop exits, and every other request is skipped.

The cap is applied before the budget clamp, so threshold >= budget never
binds. Replaying the loop: 8192/8192 serves 1 request per step (still
starved); 8192/2048 serves 4, and two long prompts advance together, which
is the issue's actual acceptance criterion.

  --max-num-batched-tokens=8192      (was unset -> default 2048)
  --long-prefill-token-threshold=2048 (was unset -> 0, cap disabled)
  --max-num-seqs=5, --gpu-memory-utilization=0.95

The ~166k KV figure is marked as a projection; the live reading is
num_gpu_blocks=9527 x block_size=16 = 152,432 tokens and is authoritative
post-boot. A preemption alert covers the accept-and-watch risk.

Observability for the companion admission gate (homelab-frontend#64, now
deployed at 64f2bfc1):
  - four dashboard panels: parked, rejected by reason, engine state stale,
    and gateway-view vs engine-truth. Each carries a description, because
    an unexplained blank panel reads as a broken dashboard.
  - alerts on observer_degraded and engine_state_stale_total, both with an
    absent() guard: the gate fails OPEN, so a missing or renamed metric must
    alert rather than look healthy. Verified live that the guards do not
    false-fire (both series present, 3 each).
  - LLMPreemptions and LLMPrefillStarvation: the acceptance metric and the
    regression direction this change introduces.
  - max(vllm_requests_running), not sum(): one engine-global gauge is read
    by all three gateway replicas, so sum() read 3x and could never agree.

Also corrects comments that were factually wrong. An earlier revision of
this branch claimed llm_ttft_seconds/llm_itl_seconds/llm_tokens_total were
dead code; homelab-frontend#63 wired them, verified live as 3 series each.
request_duration_seconds is the genuinely dead one (0 samples, zero
non-test call sites) and is now marked as such instead of being listed as
working.

Closes #51

k8s/apps/gotify/deployment.yaml

@@ -23,7 +23,7 @@ spec:
       containers:
         # --- Gotify server ---
         - name: gotify
-          image: ghcr.io/gotify/server:2.6.1
+          image: ghcr.io/gotify/server:3.1.0
           command: ["/bin/sh", "-c"]
           args:
             - |
@@ -84,7 +84,7 @@ spec:
         # Watches Gotify WebSocket stream, forwards messages as email.
         # https://github.com/eternal-flame-AD/gotify-broadcast
         - name: smtp-emailer
-          image: ghcr.io/gotify/server:2.6.1
+          image: ghcr.io/gotify/server:3.1.0
           command:
             - /bin/sh
             - -c

k8s/apps/gotify/igotify-deployment.yaml

@@ -0,0 +1,59 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: igotify
+  namespace: notifications
+  labels:
+    app: igotify
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: igotify
+  template:
+    metadata:
+      labels:
+        app: igotify
+    spec:
+      containers:
+        - name: igotify
+          image: ghcr.io/androidseb25/igotify-notification-assist:v1.6.0.1
+          ports:
+            - containerPort: 8080
+              protocol: TCP
+          env:
+            - name: ENABLE_CONSOLE_LOG
+              value: "true"
+            - name: ENABLE_SCALAR_UI
+              value: "true"
+            # GOTIFY_URLS / GOTIFY_CLIENT_TOKENS / SECNTFY_TOKENS intentionally
+            # left unset — not required when Gotify is reachable via public domain
+            # (gotify.riotpiao.com). iGotify iOS app registers itself at runtime.
+          volumeMounts:
+            - name: data
+              mountPath: /app/data
+          livenessProbe:
+            httpGet:
+              path: /Version
+              port: 8080
+            initialDelaySeconds: 10
+            periodSeconds: 30
+          readinessProbe:
+            httpGet:
+              path: /Version
+              port: 8080
+            initialDelaySeconds: 5
+            periodSeconds: 10
+          resources:
+            requests:
+              cpu: 50m
+              memory: 64Mi
+            limits:
+              cpu: 200m
+              memory: 128Mi
+      volumes:
+        - name: data
+          persistentVolumeClaim:
+            claimName: igotify-data

k8s/apps/gotify/igotify-ingress.yaml

@@ -0,0 +1,22 @@
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: igotify
+  namespace: notifications
+  annotations:
+    nginx.ingress.kubernetes.io/proxy-read-timeout: "600"
+    nginx.ingress.kubernetes.io/proxy-send-timeout: "600"
+    nginx.ingress.kubernetes.io/proxy-http-version: "1.1"
+spec:
+  ingressClassName: nginx
+  rules:
+    - host: igotify.riotpiao.com
+      http:
+        paths:
+          - path: /
+            pathType: Prefix
+            backend:
+              service:
+                name: igotify
+                port:
+                  number: 80

k8s/apps/gotify/igotify-pvc.yaml

@@ -0,0 +1,11 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: igotify-data
+  namespace: notifications
+spec:
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 1Gi

k8s/apps/gotify/igotify-service.yaml

@@ -0,0 +1,14 @@
+apiVersion: v1
+kind: Service
+metadata:
+  name: igotify
+  namespace: notifications
+  labels:
+    app: igotify
+spec:
+  selector:
+    app: igotify
+  ports:
+    - port: 80
+      targetPort: 8080
+      protocol: TCP

k8s/apps/gotify/ingress.yaml

@@ -6,11 +6,9 @@ metadata:
   annotations:
     nginx.ingress.kubernetes.io/proxy-read-timeout: "600"
     nginx.ingress.kubernetes.io/proxy-send-timeout: "600"
-    # WebSocket support for Gotify client connections
+    # WebSocket support for Gotify client connections (/stream endpoint)
+    # nginx-ingress handles Upgrade/Connection headers natively with http/1.1
     nginx.ingress.kubernetes.io/proxy-http-version: "1.1"
-    nginx.ingress.kubernetes.io/configuration-snippet: |
-      proxy_set_header Upgrade $http_upgrade;
-      proxy_set_header Connection "upgrade";
 spec:
   ingressClassName: nginx
   rules:

k8s/apps/gotify/kustomization.yaml

@@ -6,3 +6,7 @@ resources:
   - deployment.yaml
   - service.yaml
   - ingress.yaml
+  - igotify-pvc.yaml
+  - igotify-deployment.yaml
+  - igotify-service.yaml
+  - igotify-ingress.yaml

k8s/apps/llm-serving/ornith.yaml

@@ -44,7 +44,7 @@ spec:
       - name: OLLAMA_MODELS
         value: /mnt/models/ollama
       - name: OLLAMA_CONTEXT_LENGTH
-        value: '32768'
+        value: '65536'
       - name: OLLAMA_KEEP_ALIVE
         value: '-1'
       - name: OLLAMA_NUM_PARALLEL

k8s/apps/llm-serving/reasoning.yaml

@@ -28,9 +28,14 @@ spec:
       # KV cache math: 256KB/token total (64 layers, 8 KV heads, 128
       # head_dim, fp16), PP=2 splits both weights and KV load ~evenly, so
       # each GPU carries ~9.67GB weights + ~128KB/token KV. At
-      # gpu-memory-utilization=0.90 (28.8GB/GPU usable), that leaves
+      # gpu-memory-utilization=0.95 (30.4GB/GPU usable), that leaves
-      # ~19.1GB/GPU for KV cache -> ~156k tokens/GPU capacity, comfortably
+      # ~20.7GB/GPU for KV cache. NOTE this is a projection, not a
-      # above the 131072 target with room to spare -- the old
+      # measurement: the live reading at 0.90 is num_gpu_blocks=9527 x
+      # block_size=16 = 152,432 tokens (vllm:cache_config_info). Re-read that
+      # metric after the rollout -- max_num_batched_tokens=8192 also raises
+      # activation memory out of the same budget, so the post-boot
+      # num_gpu_blocks is the authoritative number, and --max-num-seqs=5 is
+      # only valid if it still covers 2x65k + 3x8k. The old
       # OffloadingConnector CPU-DRAM spillover (tuned for the previous
       # model's much smaller 16384 context) is no longer needed and is
       # dropped. Staying on PP=2 and vLLM 0.11.0 (no version bump needed,
@@ -54,10 +59,37 @@ spec:
       - --tensor-parallel-size=1
       - --pipeline-parallel-size=2
       - --max-model-len=131072
-      - --gpu-memory-utilization=0.90
+      - --gpu-memory-utilization=0.95
-      - --max-num-seqs=4
+      - --max-num-seqs=5
+      # Prefill scheduling -- see the measured starvation this fixes.
+      # V1's scheduler has no prefill/decode phase: it walks self.running and
+      # hands out tokens from one shared budget (vllm/v1/core/sched/scheduler.py:
+      # "There's no decoding phase nor prefill phase in the scheduler"). The cap
+      # per request is applied BEFORE the budget clamp:
+      #     if 0 < long_prefill_token_threshold < num_new_tokens: take = threshold
+      #     take = min(take, token_budget)
+      # The loop is `while ... and token_budget > 0`. With the threshold at its
+      # default of 0 the cap is disabled, so one long prompt takes the ENTIRE
+      # budget, token_budget hits 0, the loop exits and every other request is
+      # skipped for that step -- measured as 262 tok/s contended prefill and
+      # 0.0-0.1 tok/s generation for everything else, Running: 2-4 reqs.
+      # Hence threshold must be strictly BELOW the budget or it never binds:
+      # at 8192/8192 a single request is served per step (still starved), while
+      # 8192/2048 serves four. Verified by replaying the loop.
+      - --max-num-batched-tokens=8192
+      - --long-prefill-token-threshold=2048
       - --enable-chunked-prefill
       - --enable-prefix-caching
+      #
+      # NOT set, deliberately: --max-num-partial-prefills and
+      # --max-long-partial-prefills. Those are V0-only knobs. vLLM 0.11.0 ships
+      # V1 only (vllm/core/ absent), and EngineArgs._is_v1_supported_oracle
+      # rejects any value other than the default for either one, calling
+      # _raise_or_fallback("Concurrent Partial Prefill") -> "Falling back to V0"
+      # -> AssertionError, because there is no V0 engine to fall back to.
+      # The Deployment uses strategy Recreate, so setting them is an outage.
+      # V1 does concurrent chunked prefill natively; it needs no flag, only the
+      # budget/threshold ratio above.
       # qwen3 is vLLM's dedicated reasoning parser for this family's <think>
       # blocks.
       - --reasoning-parser=qwen3

k8s/apps/paperless/configmap.yaml

@@ -20,3 +20,7 @@ data:
   # wire up allauth's confirm-email view, so signup 500s with NoReverseMatch
   # on 'account_confirm_email' without this.
   PAPERLESS_ACCOUNT_EMAIL_VERIFICATION: "none"
+  # Auto-connect social accounts to existing users with matching email
+  PAPERLESS_SOCIAL_AUTO_SIGNUP: "true"
+  # Allow automatic linking of social accounts to existing users
+  PAPERLESS_ACCOUNT_ALLOW_SIGNUPS: "true"

k8s/apps/paperless/kustomization.yaml

@@ -4,6 +4,7 @@ namespace: paperless
 resources:
   - pvc.yaml
   - configmap.yaml
+  - paperless-ai-config.yaml
   - redis.yaml
   - deployment.yaml
   - service.yaml

k8s/apps/paperless/paperless-ai-config.yaml

@@ -0,0 +1,49 @@
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: paperless-ai-env
+  namespace: paperless
+data:
+  paperless-ai-config.env: |
+    # Paperless-NGX API configuration
+    PAPERLESS_URL=http://paperless.paperless.svc.cluster.local:8000
+    PAPERLESS_API_TOKEN=7b89463e04c141f4172fbcddf78d623547d327ec
+
+    # AI Processing settings
+    ENABLE_AI_PROCESSING=yes
+    AI_MODEL=reasoning
+    LLM_API_URL=https://api.riotpiao.com/v1
+    LLM_API_TOKEN_FILE=/app/data/llm_token.txt
+
+    # Auto-tagging configuration
+    ENABLE_AUTO_TAGGING=yes
+    ENABLE_CORRESPONDENT_EXTRACTION=yes
+    ENABLE_DOCUMENT_TYPE_EXTRACTION=yes
+    ENABLE_TITLE_GENERATION=yes
+    
+    # Advanced tagging with AI
+    ENABLE_INTELLIGENT_TAGGING=yes
+    ENABLE_DATE_EXTRACTION=yes
+    ENABLE_TIMELINE_TAGGING=yes
+    AUTO_CREATE_TAGS=yes
+    
+    # Tagging prompts for custom extraction
+    CUSTOM_EXTRACTION_PROMPT=Extract document date, time period, financial year, and temporal context. Create tags like YYYY, Q1-YYYY, Last-Month, Current-Year, Historic
+    TAG_EXTRACTION_MODEL=reasoning
+
+    # Scanning behavior
+    SCAN_INTERVAL_SECONDS=60
+    PROCESS_EXISTING_DOCUMENTS=no
+    ADD_AI_TAG=yes
+    AI_TAG_NAME=ai-processed
+
+    # Date/Timeline tagging configuration
+    DATE_EXTRACTION_FORMAT=iso8601
+    TIMELINE_TAG_GRANULARITY=year,quarter,month
+    AUTO_TAG_HISTORICAL=yes
+    AUTO_TAG_RECENT=yes
+    
+    # Performance
+    MAX_CONCURRENT_REQUESTS=2
+    REQUEST_TIMEOUT=60
+    DATE_EXTRACTION_TIMEOUT=30

k8s/apps/paperless/paperless-ai-new.yaml

@@ -0,0 +1,127 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: paperless-ai-data
+  namespace: paperless
+spec:
+  accessModes:
+    - ReadWriteOnce
+  storageClassName: longhorn
+  resources:
+    requests:
+      storage: 2Gi
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: paperless-ai
+  namespace: paperless
+  labels:
+    app.kubernetes.io/name: paperless-ai
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: paperless-ai
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: paperless-ai
+    spec:
+      tolerations:
+        - key: node-role.kubernetes.io/control-plane
+          operator: Exists
+          effect: NoSchedule
+      
+      volumes:
+        - name: paperless-ai-data
+          persistentVolumeClaim:
+            claimName: paperless-ai-data
+      
+      initContainers:
+        - name: fetch-llm-token
+          image: curlimages/curl:8.12.0
+          command:
+            - sh
+            - -c
+            - |
+              set -e
+              echo "[init] Fetching JWT token from Authentik for paperless-ai-agent..."
+              
+              # Get JWT token via client_credentials grant
+              TOKEN_RESPONSE=$(curl -s -X POST https://authentik.riotpiao.com/application/o/token/ \
+                -d "grant_type=client_credentials" \
+                -d "client_id=paperless-ai-agent" \
+                -d "client_secret=29skAhE0tBzWn0qrPvmPTQMva0joFjcCr7msDQr_VXQ" \
+                -d "scope=openid llm:inference" 2>/dev/null)
+              
+              # Extract token
+              TOKEN=$(echo "$TOKEN_RESPONSE" | grep -o '"access_token":"[^"]*' | cut -d'"' -f4)
+              
+              if [ -z "$TOKEN" ]; then
+                echo "[error] Failed to get token. Response: $TOKEN_RESPONSE"
+                exit 1
+              fi
+              
+              # Store token in file for main container to read
+              mkdir -p /data
+              echo "$TOKEN" > /data/llm_token.txt
+              echo "[init] Token fetched and stored successfully"
+          volumeMounts:
+            - name: paperless-ai-data
+              mountPath: /data
+      
+      containers:
+        - name: paperless-ai
+          image: clusterzx/paperless-ai:latest
+          imagePullPolicy: Always
+          env:
+            # Paperless-ngx connection
+            - name: PAPERLESS_API_URL
+              value: "http://paperless.paperless.svc.cluster.local:8000"
+            - name: PAPERLESS_API_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: paperless-ai-config
+                  key: PAPERLESS_API_TOKEN
+            - name: PAPERLESS_USERNAME
+              value: "admin"
+            
+            # LLM API configuration
+            - name: AI_PROVIDER
+              value: "custom"
+            - name: CUSTOM_BASE_URL
+              value: "https://api.riotpiao.com/v1"
+            # Token file created by init container with JWT from Authentik
+            - name: CUSTOM_API_KEY_FILE
+              value: "/app/data/llm_token.txt"
+            - name: CUSTOM_MODEL
+              value: "qwen2.5:3b-instruct"
+            
+            # Scanning behavior
+            - name: SCAN_INTERVAL
+              value: "60"
+            - name: PROCESS_PREDEFINED_DOCUMENTS
+              value: "no"
+            
+            # Tagging configuration
+            - name: ADD_AI_TAG
+              value: "yes"
+            - name: AI_TAG_NAME
+              value: "ai-processed"
+            - name: ADD_AI_DESCRIPTIONS
+              value: "yes"
+            - name: USE_PROMPT_TAGS
+              value: "yes"
+          
+          volumeMounts:
+            - name: paperless-ai-data
+              mountPath: /app/data
+          
+          resources:
+            requests:
+              cpu: 100m
+              memory: 512Mi
+            limits:
+              cpu: "1"
+              memory: 2Gi

k8s/apps/paperless/paperless-ai.yaml

@@ -25,9 +25,15 @@ spec:
         - name: paperless-ai-data
           persistentVolumeClaim:
             claimName: paperless-ai-data
+        - name: paperless-ai-env-config
+          configMap:
+            name: paperless-ai-env
       initContainers:
         - name: fetch-llm-token
           image: curlimages/curl:8.12.0
+          securityContext:
+            runAsUser: 0
+            fsGroup: 0
           command:
             - sh
             - -c
@@ -43,8 +49,8 @@ spec:
                 -d "client_secret=${LLM_AUTH_CLIENT_SECRET}" \
                 -d "scope=openid llm:inference" 2>/dev/null)
               
-              # Extract token
+              # Extract token from JSON response using sed (handles spaces after colons)
-              TOKEN=$(echo "$TOKEN_RESPONSE" | grep -o '"access_token":"[^"]*' | cut -d'"' -f4)
+              TOKEN=$(echo "$TOKEN_RESPONSE" | sed -n 's/.*"access_token"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p')
               
               if [ -z "$TOKEN" ]; then
                 echo "[error] Failed to get token. Response: $TOKEN_RESPONSE"
@@ -52,9 +58,13 @@ spec:
               fi
               
               # Store token in file for main container to read
-              mkdir -p /data
+              # Write to both locations for compatibility
-              echo "$TOKEN" > /data/llm_token.txt
+              mkdir -p /tmp/llm-token /app/data 2>/dev/null || true
-              echo "[init] Token fetched and stored successfully"
+              echo "$TOKEN" | tee /tmp/llm-token/llm_token.txt > /dev/null 2>&1
+              echo "$TOKEN" > /app/data/llm_token.txt 2>/dev/null || true
+              echo "[init] Token fetched and stored"
+              [ -f /tmp/llm-token/llm_token.txt ] && echo "  -> /tmp/llm-token/llm_token.txt"
+              [ -f /app/data/llm_token.txt ] && echo "  -> /app/data/llm_token.txt"
           env:
             - name: LLM_AUTH_CLIENT_SECRET
               valueFrom:
@@ -63,7 +73,7 @@ spec:
                   key: LLM_AUTH_CLIENT_SECRET
           volumeMounts:
             - name: paperless-ai-data
-              mountPath: /data
+              mountPath: /app/data
       containers:
         - name: paperless-ai
           image: clusterzx/paperless-ai:latest
@@ -105,6 +115,9 @@ spec:
           volumeMounts:
             - name: paperless-ai-data
               mountPath: /app/data
+            - name: paperless-ai-env-config
+              mountPath: /app/data/.env
+              subPath: paperless-ai-config.env
           resources:
             requests:
               cpu: 100m

k8s/apps/poimen-frontend/ingress.yaml

@@ -0,0 +1,24 @@
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: poimen-frontend
+  namespace: poimen
+  annotations:
+    # Next.js SSE + LLM streaming responses need long timeouts + no buffering.
+    nginx.ingress.kubernetes.io/proxy-read-timeout: "3600"
+    nginx.ingress.kubernetes.io/proxy-send-timeout: "3600"
+    nginx.ingress.kubernetes.io/proxy-buffering: "off"
+    nginx.ingress.kubernetes.io/proxy-body-size: "10m"
+spec:
+  ingressClassName: nginx
+  rules:
+    - host: poimen.riotpiao.com
+      http:
+        paths:
+          - path: /
+            pathType: Prefix
+            backend:
+              service:
+                name: poimen-frontend
+                port:
+                  number: 3000

k8s/apps/poimen-frontend/kustomization.yaml

@@ -0,0 +1,8 @@
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+# Edge Ingress for poimen.riotpiao.com.
+# Same pattern as k8s/apps/api/ (api-gw ingress).
+# App manifests (Deployment, Service, NetworkPolicy) live in
+# riotpiao-poimen/poimen-frontend.git k8s/ — each repo owns its own k8s.
+resources:
+  - ingress.yaml

k8s/argocd/apps/01-image-updater.yaml

@@ -1,33 +0,0 @@
-# ArgoCD Image Updater - auto-updates Application images from registry
-# Watches forgejo.riotpiao.com for new image tags and updates Applications
-apiVersion: argoproj.io/v1alpha1
-kind: Application
-metadata:
-  name: argocd-image-updater
-  namespace: argocd
-  finalizers:
-    - resources-finalizer.argocd.argoproj.io
-  annotations:
-    argocd.argoproj.io/sync-wave: "1"
-spec:
-  project: homelab
-  revisionHistoryLimit: 3
-  sources:
-    - repoURL: https://argoproj.github.io/argo-helm
-      chart: argocd-image-updater
-      targetRevision: "0.11.2"
-      helm:
-        valueFiles:
-          - $values/k8s/infra/argocd-image-updater/values.yaml
-    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/homelab.git
-      targetRevision: main
-      ref: values
-  destination:
-    server: https://kubernetes.default.svc
-    namespace: argocd
-  syncPolicy:
-    automated:
-      prune: true
-      selfHeal: true
-    syncOptions:
-      - CreateNamespace=false

k8s/argocd/apps/30-security.yaml

@@ -152,13 +152,8 @@ metadata:
   namespace: argocd
   annotations:
     argocd.argoproj.io/sync-wave: "3"
-    argocd-image-updater.argoproj.io/image-list: runner=forgejo.riotpiao.com/rock/forgejo-runner-golang
+    # Image promotion: Forgejo package webhook → api-gw.
-    argocd-image-updater.argoproj.io/runner.update-strategy: newest-build
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-    argocd-image-updater.argoproj.io/runner.allow-tags: regexp:^[0-9a-f]{7}$|^latest$|^v[0-9]+$
-    argocd-image-updater.argoproj.io/runner.helm.image-name: runner.image.repository
-    argocd-image-updater.argoproj.io/runner.helm.image-tag: runner.image.tag
-    argocd-image-updater.argoproj.io/write-back-method: git
-    argocd-image-updater.argoproj.io/git-branch: main
 spec:
   project: homelab
   source:
@@ -180,13 +175,8 @@ metadata:
   namespace: argocd
   annotations:
     argocd.argoproj.io/sync-wave: "3"
-    argocd-image-updater.argoproj.io/image-list: runner=forgejo.riotpiao.com/rock/forgejo-runner-node
+    # Image promotion: Forgejo package webhook → api-gw.
-    argocd-image-updater.argoproj.io/runner.update-strategy: newest-build
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-    argocd-image-updater.argoproj.io/runner.allow-tags: regexp:^[0-9a-f]{7}$|^latest$|^v[0-9]+$
-    argocd-image-updater.argoproj.io/runner.helm.image-name: runner.image.repository
-    argocd-image-updater.argoproj.io/runner.helm.image-tag: runner.image.tag
-    argocd-image-updater.argoproj.io/write-back-method: git
-    argocd-image-updater.argoproj.io/git-branch: main
 spec:
   project: homelab
   source:
@@ -211,13 +201,8 @@ metadata:
   namespace: argocd
   annotations:
     argocd.argoproj.io/sync-wave: "3"
-    argocd-image-updater.argoproj.io/image-list: runner=forgejo.riotpiao.com/rock/forgejo-runner-rust
+    # Image promotion: Forgejo package webhook → api-gw.
-    argocd-image-updater.argoproj.io/runner.update-strategy: newest-build
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-    argocd-image-updater.argoproj.io/runner.allow-tags: regexp:^[0-9a-f]{7}$|^latest$|^v[0-9]+$
-    argocd-image-updater.argoproj.io/runner.helm.image-name: runner.image.repository
-    argocd-image-updater.argoproj.io/runner.helm.image-tag: runner.image.tag
-    argocd-image-updater.argoproj.io/write-back-method: git
-    argocd-image-updater.argoproj.io/git-branch: main
 spec:
   project: homelab
   source:

k8s/argocd/apps/55-api-gateway.yaml

@@ -36,11 +36,9 @@ metadata:
     app.kubernetes.io/component: gateway
   annotations:
     argocd.argoproj.io/sync-wave: "7"
-    # ArgoCD Image Updater - auto-update on new image push
+    # Image promotion: Forgejo package webhook → api-gw /v1/webhooks/forgejo
-    argocd-image-updater.argoproj.io/image-list: gw=forgejo.riotpiao.com/rock/api-gateway
+    # → patches this Application's kustomize.images via k8s API.
-    argocd-image-updater.argoproj.io/gw.update-strategy: digest
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-    argocd-image-updater.argoproj.io/gw.allow-tags: regexp:^latest$
-    argocd-image-updater.argoproj.io/write-back-method: argocd
 spec:
   project: homelab
   revisionHistoryLimit: 3

k8s/argocd/apps/60-applications.yaml

@@ -246,11 +246,8 @@ metadata:
   namespace: argocd
   annotations:
     argocd.argoproj.io/sync-wave: "8"
-    # ArgoCD Image Updater - auto-update on new image push
+    # Image promotion: Forgejo package webhook → api-gw.
-    argocd-image-updater.argoproj.io/image-list: app=forgejo.riotpiao.com/rock/portfolio
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-    argocd-image-updater.argoproj.io/app.update-strategy: digest
-    argocd-image-updater.argoproj.io/app.allow-tags: regexp:^latest$
-    argocd-image-updater.argoproj.io/write-back-method: argocd
 spec:
   project: homelab
   source:

k8s/argocd/apps/70-poimen.yaml

@@ -1,37 +1,81 @@
+# Poimen — unified ArgoCD Application (wave 7).
+#
+# Consolidates what was previously five separate apps:
+#   poimen            → source 1 (poimen-workflows workers)
+#   poimen-workflows  → duplicate of source 1 (deleted)
+#   poimen-memory     → source 2 (memory service + redis)
+#   poimen-memory-app → duplicate of source 2 (deleted)
+#   poimen-memory-database → source 3 (CNPG cluster, moved to k8s/infra/poimen)
+#
+# poimen-toolkit is intentionally kept as a separate Application because it
+# manages cluster-scoped CRDs (ToolDefinition). A bad CRD can affect the
+# whole cluster; keeping it separate means it has its own sync history,
+# independent rollback, and can be paused without touching app workloads.
+#
+# Sync ordering via argocd sync-wave annotations:
+#   wave -1 → DB (CNPG Cluster) — must be Ready before app pods start
+#   wave  0 → all application Deployments (default)
+#
+# Image Updater tracks all four images. Write-back via argocd (in-cluster
+# annotation patch — no git commit per image update).
 apiVersion: argoproj.io/v1alpha1
 kind: Application
 metadata:
   name: poimen
   namespace: argocd
+  labels:
+    app.kubernetes.io/part-of: poimen
   annotations:
     argocd.argoproj.io/sync-wave: "7"
-    # Image Updater: auto-update on new image push (SHA tag filter)
+    # ── Image promotion ──────────────────────────────────────────────────────
-    argocd-image-updater.argoproj.io/image-list: |
+    # Forgejo package webhook → api-gw patches kustomize.images per source.
-      memory=forgejo.riotpiao.com/rock/poimen-memory
+    # Source index mapping: 0=workflows, 1=memory, 3=frontend.
-      workflows=forgejo.riotpiao.com/rock/poimen-workflows
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
-      frontend=forgejo.riotpiao.com/rock/poimen-frontend
-    argocd-image-updater.argoproj.io/memory.update-strategy: digest
-    argocd-image-updater.argoproj.io/memory.allow-tags: regexp:^latest$
-    argocd-image-updater.argoproj.io/workflows.update-strategy: digest
-    argocd-image-updater.argoproj.io/workflows.allow-tags: regexp:^latest$
-    argocd-image-updater.argoproj.io/frontend.update-strategy: digest
-    argocd-image-updater.argoproj.io/frontend.allow-tags: regexp:^latest$
-    argocd-image-updater.argoproj.io/write-back-method: argocd
 spec:
   project: homelab
-  source:
+  # Multi-source: each repo owns its own k8s manifests.
-    repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/poimen-workflows.git
+  # ArgoCD syncs all sources in parallel; sync-wave annotations on individual
+  # resources control ordering within the sync operation.
+  sources:
+    # 1. Workers — Temporal activity + workflow workers, pi-worker pool
+    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/poimen-workflows.git
       targetRevision: main
       path: k8s
+
+    # 2. Memory service — poimen-memory API, Redis cache, NetworkPolicies
+    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/poimen-memory.git
+      targetRevision: main
+      path: k8s/app
+
+    # 3. Database infrastructure — CNPG Postgres cluster (wave -1)
+    #    Moved from k8s/infra/databases to isolate poimen DB from other tenants.
+    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/homelab.git
+      targetRevision: main
+      path: k8s/infra/poimen
+
+    # 4. Frontend app manifests — Deployment, Service, NetworkPolicy.
+    #    Each repo owns its own k8s/ (same pattern as homelab-frontend for api-gw).
+    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/poimen-frontend.git
+      targetRevision: main
+      path: k8s
+
+    # 5. Frontend ingress — poimen.riotpiao.com edge route.
+    #    Lives in homelab (same pattern as k8s/apps/api/ for api-gw).
+    - repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/homelab.git
+      targetRevision: main
+      path: k8s/apps/poimen-frontend
+
   destination:
     server: https://kubernetes.default.svc
     namespace: poimen
+
   syncPolicy:
     automated:
       prune: true
       selfHeal: true
     syncOptions:
       - CreateNamespace=true
+      - SkipDryRunOnMissingResource=true
     retry:
       limit: 5
       backoff:

k8s/argocd/apps/71-poimen-toolkit.yaml

@@ -0,0 +1,32 @@
+apiVersion: argoproj.io/v1alpha1
+kind: Application
+metadata:
+  name: poimen-toolkit
+  namespace: argocd
+  annotations:
+    argocd.argoproj.io/sync-wave: "7"
+    # Image promotion: Forgejo package webhook → api-gw.
+    # Tracks: poimen-toolkit-controller, pi-worker.
+    # Replaced ArgoCD Image Updater (removed 2026-09-10).
+spec:
+  project: homelab
+  source:
+    repoURL: https://forgejo.riotpiao.com/riotpiao-poimen/poimen-toolkit.git
+    targetRevision: main
+    path: k8s
+  destination:
+    server: https://kubernetes.default.svc
+    namespace: poimen
+  syncPolicy:
+    automated:
+      prune: true
+      selfHeal: true
+    syncOptions:
+      - CreateNamespace=true
+      - ServerSideApply=true
+    retry:
+      limit: 5
+      backoff:
+        duration: 5s
+        factor: 2
+        maxDuration: 3m

k8s/argocd/projects/homelab-project.yaml

@@ -24,6 +24,7 @@ spec:
     - https://forgejo.riotpiao.com/riotpiao-poimen/poimen-memory.git
     - https://forgejo.riotpiao.com/riotpiao-poimen/poimen-workflows.git
     - https://forgejo.riotpiao.com/riotpiao-poimen/poimen-frontend.git
+    - https://forgejo.riotpiao.com/riotpiao-poimen/poimen-toolkit.git
     - https://forgejo.riotpiao.com/rock/riotpiao.com.git
     # Public Helm chart repos referenced by k8s/argocd/apps/* and bootstrap/*
     - https://cloudnative-pg.github.io/charts

k8s/argocd/secrets/forgejo-registry-pat-api.enc.yaml

@@ -0,0 +1,23 @@
+apiVersion: v1
+kind: Secret
+metadata:
+    name: forgejo-registry-pat
+    namespace: api
+type: Opaque
+stringData:
+    pat: ENC[AES256_GCM,data:alYPvHBVqnf13NZu6jtgA6WdFM4x46s5mIY8uEnfWdq0Vp+7oKrY1A==,iv:JbUjes5ZB9eKQ1nkV02++7F/BaWIROfT6gjShzajXro=,tag:ApjawJold0QApnX1GUXGhw==,type:str]
+sops:
+    age:
+        - enc: |
+            -----BEGIN AGE ENCRYPTED FILE-----
+            YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNTUxOSB6bFZhU3JLSUZEVWhYdEdZ
+            MURoSWNlZjY0RkxyaWlwU2ZTVjJ5V2FnMVVZCis5YlBwQnZrWnVJNFh4UnNldmx6
+            Y2JjdnpkRTVGSjRSZ0U0THdDZHIrNHcKLS0tIHZDd1BDVGRDSHJEQThuTExORWs2
+            b3BxenY4WHhpRkRTZHJGMkNYMm14RnMKH//Xpqbf/ZBNFQ43D0eFUWwRcQWqA3CC
+            RG8U9wyxC6YrlruU2hY6eB2x0OjsVic34ZGaR752Bwyp0Pz49C/k3g==
+            -----END AGE ENCRYPTED FILE-----
+          recipient: age1e5fq3hwxy78psus2nfvmtmua36g0u3suk78ephw6246l974d2utsvn0hla
+    encrypted_regex: ^(data|stringData)$
+    lastmodified: "2026-09-28T03:10:55Z"
+    mac: ENC[AES256_GCM,data:nLZueHygckw+Gddlea90tRU8h8nEGpe21JUi9nCdpc0ctkSlTDqcR3Q7JFPwbEGHGFU5UJ2xnkBZEDIVmExJryqPBMXT7hQd3ySeQgPwAPGt6zEhlRc5jXhqmmrpJiiN5XUL52AZVDwcWACg7kVzBXIClzihFRW+UqjJZhdtELM=,iv:C0H4AFfdfpsBvP1z2chVvW/VTejxYxwNn/KuA50LeH0=,tag:ZEkTCTKvoZH7DjrcwfL3dw==,type:str]
+    version: 3.13.2

k8s/argocd/secrets/forgejo-registry-token.enc.yaml

@@ -0,0 +1,24 @@
+apiVersion: v1
+kind: Secret
+type: Opaque
+metadata:
+    name: forgejo-registry-token
+    namespace: cicd
+data:
+    username: ENC[AES256_GCM,data:oRNecs0a7rs=,iv:/5EFdNqpSjoxe7XDmAEDMd6lXd0SMsRiMmBM9SIwDHY=,tag:Wi+ggoP+9/vgz9rHrxB9nA==,type:str]
+    password: ENC[AES256_GCM,data:HbmwnWk3ST5SDWXwIhu+M5f/FOfFbJHaG0Q/IWrO63HRXmJTmLMpbjtWIXC9IMFoHeDZ1MNAV8A=,iv:ztJZgpaS//yYCiyr6/ubs/UiXkbGMnokvy4Pe45jQMk=,tag:WMpawc/SVAV4mxsCd6XqAA==,type:str]
+sops:
+    age:
+        - enc: |
+            -----BEGIN AGE ENCRYPTED FILE-----
+            YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNTUxOSBjQmhxN2E3ZnNTdkR6U2c1
+            ZStsQktkeWUwMEdmdFIyMkRYRVk1ZGppWlZ3CmUzR2JubWw2VjdWUGt6S2lpNmQ3
+            cDdSWjJiYzRIWS9UVERpVDFJenFnaDAKLS0tIGhDSmNMSjZwcDVPR3U0dUcwNk9Q
+            WmdGS1lyS3dkSXIvOFhyRklGRll2UlUK8NGts6a6Fv10c6gxIIPqIIqfgzQ3EEyk
+            SFrMThwbdaOqLY4AMeDrg+k0Y2t1Xj/XOvdNbRCRictDwZUsptvO7A==
+            -----END AGE ENCRYPTED FILE-----
+          recipient: age1e5fq3hwxy78psus2nfvmtmua36g0u3suk78ephw6246l974d2utsvn0hla
+    encrypted_regex: ^(data|stringData)$
+    lastmodified: "2026-10-06T02:01:24Z"
+    mac: ENC[AES256_GCM,data:imebczVzG9YWd42ER4a+w+Bi3dxbN0HFm/1igxzJt6Z0HG/q6i0s59swpGVruk9dQJWrlO8KKqX739ATp2KGC5szrEuRFvPemF594XFCg6ykW+ZeP5VtTvpJg2nHv7/Pu+71ljKKi4gqOJ4SJDn8owAbSN+su6PiT323wax930A=,iv:iG/G2+qQkvbBU95w/GG2PrbAHDOaamuvjqMDBxSW6zI=,tag:wy9Vn72HBDHqCyzTcV15/Q==,type:str]
+    version: 3.13.2

k8s/argocd/secrets/memory-agent-oidc.enc.yaml

@@ -0,0 +1,26 @@
+apiVersion: v1
+kind: Secret
+type: Opaque
+metadata:
+    name: memory-agent-oidc
+    namespace: poimen
+data:
+    CLIENT_ID: ENC[AES256_GCM,data:c/RvtV6anofX8h1BmlKu9w==,iv:3lQ1BR8kVvF1RHVfKstmb0Y7u/uX5OrvmGiCi/xTpjw=,tag:9JC2WkfQFWaiZroAEmPenw==,type:str]
+    CLIENT_SECRET: ENC[AES256_GCM,data:I258/6O43PccyllYN/wfoKvTgNgZONTivM/KFIr8TGqSl+2XQkwBXvK2VQc5Xgc5cpwlrgnjk8pjQbqr,iv:MZyA+V5z9Po2VoxqvOSlWh1fZp1lrkcbYMx7VcSIaJs=,tag:gXH0BpA8KOZiVd+4RnhRzg==,type:str]
+    ISSUER: ENC[AES256_GCM,data:GX/uQPDoeJWXXQZlzD6oOIvxo7H4LRAv4MrPAiCgmv6WxW59+KKkymuDT/T9MiDm4bNhxVPgB7nHbDcMHfPvU2zCNM2TWYUsNOgGwxNAqM8=,iv:5OEsdrUZvMLO2C8Ea2i5alsyFGtmEk3othTemk8fOk4=,tag:8CMlRY2JCe4z4fv4WuWjRQ==,type:str]
+    TOKEN_URL: ENC[AES256_GCM,data:KeU1pQzvQ2/1lQhOqNtcR7t5TGVx1J7NY8ogxCXS6IuHhVe0zWYIbC4BYFc2H9PAjsMJwCCg75N1FrIiCYBvpfK8ln0=,iv:NbPiJC/B7W1Zm+JwnEGDikE2cb3L4l6TBeQTKEb2boM=,tag:FRmllFLZzQW03lsCu9mAUA==,type:str]
+sops:
+    age:
+        - enc: |
+            -----BEGIN AGE ENCRYPTED FILE-----
+            YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNTUxOSBmK0dPSmhKT01FSitDdUNW
+            Um1XQU9FVnBUeDdoWXpCcXRzVXh0VXkvNGxNClZaQ3R5U2NkYTJjaFdHRWhUVmFS
+            K0tHZVRFRDZGREVRZzZqcWJ6dnRpUUEKLS0tIE15UUNtMStKbWwzb3dEVGEwN1VQ
+            cktlVTM2SUhpOGJZMXJOOG84RnhCQzAKOdM0pThEaN37SJtRrKIivncMu8Z91Btq
+            es6NOlJdBUhzIhnEJWfDzZ9aA6I/fYzugKbC+AfUwdGgThAgHx9crg==
+            -----END AGE ENCRYPTED FILE-----
+          recipient: age1e5fq3hwxy78psus2nfvmtmua36g0u3suk78ephw6246l974d2utsvn0hla
+    encrypted_regex: ^(data|stringData)$
+    lastmodified: "2026-10-06T02:01:24Z"
+    mac: ENC[AES256_GCM,data:/BTQ984D+hsk0LmnjD3hxopB65NAR54LhoAPNcATWA7MCTJXfdnYRu3gNxUzyiyCBd+rDaPp77TBnak4jVjMnOiYJsYLj1R7hrrhTZopz8pT+YNMupperwsirKZwK8it4Mu+fnpCyox2mtGED0hk+Nh/b6IteTRt5lS7noumo2M=,iv:RWIrVkNYj+8mqMSPiLueToexabzAtVbIR525m6bUrlg=,tag:5+wcNoNYsjGKpFlML+N70Q==,type:str]
+    version: 3.13.2

k8s/argocd/secrets/oauth2-credentials.enc.yaml

@@ -0,0 +1,27 @@
+apiVersion: v1
+kind: Secret
+type: Opaque
+metadata:
+    name: oauth2-credentials
+    namespace: iam
+data:
+    api-gw-client-secret: ENC[AES256_GCM,data:YiUCk61cCsTJE8gTLZWhqVPsCSu3b4OnG+Oh456A+mnxip0FjLTTqV+kYBdfZivIkn9ua4RcVW+rYILu,iv:K08m1EaB7/SIVNWZ3FfeRP4tbCAo6KeiNC0NpE9VO1o=,tag:ZXV9YRWUOPJ/vdoqMtvQeA==,type:str]
+    minio-client-secret: ENC[AES256_GCM,data:vcwLo1cJqbZr2MwkAKO09xPjb09ss7+s6HIvEfaLxt8ka5PUAlXQ37pTkynuktlKq/t0DvSk+x1+5G7QqVD1JJlwyDGM6yuOhBQ3MwucLNl+T4RDKtgGuA==,iv:sCdd+UNX2TfTAM6Zslrh8KrvBSzqpe67F3YMGrl4YDU=,tag:9Tpzm9NEXNuXyCd7l7GeHw==,type:str]
+    poimen-client-secret: ENC[AES256_GCM,data:dYtCiMGZaOFokihfGD2SXJDfXrhfwEQcTOPNu7JNHDUnx0hqwAdYAPAJTEN21p0OpHYDINMs5lwUpMwY,iv:ll8USuzs0881ryHImVjHGO0iFxUzqg5XtI6Ekd+EFXE=,tag:I/4ZfB9Ss4zEP57sRtRTXQ==,type:str]
+    paperless-client-secret: ENC[AES256_GCM,data:XMj9rHkI20lI37vChq9wS5lUq7l95SisRX0LChuj09O5w76Zt9NdG6OEUYQFf1z8vCfMJxBInIEEIq0e,iv:UpYy2ZEZ4AgBACEyrlxEcr13V+8Ksu4p8JHYApmyYHo=,tag:o9Z+/b6T9POgowIsks56kw==,type:str]
+    grafana-client-secret: ENC[AES256_GCM,data:V/j0+P3HoDLqYjRalWK4v1hvJTuVmzxVra47HsF1/BZvWZjIyEIH3Pt8BekT3O1dUFynxjcaxdZMBVYg,iv:9pMp+vIsp/sA7GP9wqNDf5DdZVLDow46OURW0mMRXIc=,tag:i7VJFA5o0hToZEJ5JPWIvw==,type:str]
+sops:
+    age:
+        - enc: |
+            -----BEGIN AGE ENCRYPTED FILE-----
+            YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNTUxOSB2amZPendIOU9PeStOcmJG
+            c1dzOFVvd3RoT3dXQlJUNkdCSytEZUVGVEJJCklMUkNjUytBMXh0OVUrNUlpMWw3
+            aHRFT1hBZTZHNHVtaGJCTEFyT2ZWNWsKLS0tIENxR3FKMEplV0V5Y0dsZzBkODVG
+            U0ZlT2R0VFdEQmpsakVvZ1FJRkRQYTAKESQcSlMAuCG0/tLWWZaeoPayf1EKxSaf
+            e4o8TmjmSgPEhepQ/LmNXl6obyuqUxztkh4Mm6pocqDp+gfzt7X/1g==
+            -----END AGE ENCRYPTED FILE-----
+          recipient: age1e5fq3hwxy78psus2nfvmtmua36g0u3suk78ephw6246l974d2utsvn0hla
+    encrypted_regex: ^(data|stringData)$
+    lastmodified: "2026-10-06T02:01:24Z"
+    mac: ENC[AES256_GCM,data:AI/znHiUa/Rwg0YRz65qUZ+b6V73oi3ZfAG/V5uZZvrGFrA0EISzD5LaP5Kl4GUU1svo4G5EI4aftPDgpdPptz21CCa8iYkXGe9pTDmtaE5ZICQAUbRcVhGMTfuWFhrgEhMfCJdOpfHaIQVEsQJQJgrnkXQVuUe9aT75LQz7pnY=,iv:KbIaYT0VeBT2WeSU5V5ZB4e+04JICYaY+egGN1Pv5bM=,tag:pPbhDxeLUx6V1XC/LPwxTg==,type:str]
+    version: 3.13.2

k8s/argocd/secrets/secret-generator.yaml

@@ -12,6 +12,7 @@ files:
   - authentik-secrets.enc.yaml
   - cloudflare-secrets.enc.yaml
   - forgejo-registry-pat.enc.yaml
+  - forgejo-registry-pat-api.enc.yaml
   - forgejo-registry-pull.enc.yaml
   - forgejo-runner-token.enc.yaml
   - forgejo-secrets.enc.yaml
@@ -32,3 +33,6 @@ files:
   - gotify-smtp-secrets.enc.yaml
   - forgejo-smtp-secrets.enc.yaml
   - paperless-ai-secrets.enc.yaml
+    - memory-agent-oidc.enc.yaml
+    - oauth2-credentials.enc.yaml
+    - forgejo-registry-token.enc.yaml

k8s/bootstrap/phase3-forgejo/forgejo-values.yaml

@@ -46,6 +46,12 @@ gitea:
       SSH_PORT: 2222
       SSH_LISTEN_PORT: 2222
 
+    # Self-signup was open (GET /user/sign_up returned 200), so anyone reaching
+    # forgejo.riotpiao.com could create an account. Accounts are provisioned by
+    # the admin only; admin create-user via UI/API is unaffected by this flag.
+    service:
+      DISABLE_REGISTRATION: true
+
     database:
       DB_TYPE: postgres
       HOST: forgejo-db-rw.cicd.svc.cluster.local:5432
@@ -166,3 +172,8 @@ deployment:
         secretKeyRef:
           name: forgejo-smtp
           key: password
+    # Allow webhooks to reach api-gw through the cluster ingress (resolves to
+    # internal ClusterIP 10.101.x.x which is blocked by default ALLOWED_HOST_LIST).
+    # api.riotpiao.com is our own gateway — safe to allowlist.
+    - name: GITEA__WEBHOOK__ALLOWED_HOST_LIST
+      value: "external,api.riotpiao.com"

k8s/infra/argocd-image-updater/values.yaml

@@ -1,70 +0,0 @@
-# ArgoCD Image Updater configuration
-# Watches Forgejo registry and updates ArgoCD Applications with new image tags
-
-config:
-  # Registry configuration - Forgejo allows anonymous pulls
-  registries:
-    - name: forgejo
-      api_url: https://forgejo.riotpiao.com
-      prefix: forgejo.riotpiao.com
-      default: true
-      insecure: false
-
-  # Log level
-  logLevel: debug
-
-  # ArgoCD API server
-  argocd:
-    grpcWeb: true
-    serverAddress: argocd-server.argocd.svc.cluster.local
-    insecure: true
-    plaintext: true
-
-  # Git write-back configuration (for multi-source Applications)
-  git:
-    # Commit author for image updates
-    user:
-      name: "ArgoCD Image Updater"
-      email: "[email protected]"
-    # Use SSH keys from ArgoCD's known hosts + credentials
-    # Image Updater inherits ArgoCD's git credentials (mounted via ArgoCD secret)
-
-# Mount ArgoCD's git credentials for write-back
-extraVolumes:
-  - name: argocd-ssh-known-hosts-cm
-    configMap:
-      name: argocd-ssh-known-hosts-cm
-      defaultMode: 0644
-  - name: argocd-gpg-keys-cm
-    configMap:
-      name: argocd-gpg-keys-cm
-      optional: true
-      defaultMode: 0644
-  - name: argocd-gpg-pubring
-    configMap:
-      name: argocd-gpg-pubring-cm
-      optional: true
-      defaultMode: 0644
-
-extraVolumeMounts:
-  - name: argocd-ssh-known-hosts-cm
-    mountPath: /etc/ssh/ssh_known_hosts.d/argocd-ssh-known-hosts
-    subPath: ssh_known_hosts
-  - name: argocd-gpg-keys-cm
-    mountPath: /etc/gpg/source
-  - name: argocd-gpg-pubring
-    mountPath: /etc/gpg/pubring
-
-# Extra environment variables
-extraEnv:
-  - name: GIT_SSH_KNOWN_HOSTS_CONFIG_MAP_ENABLED
-    value: "true"
-
-# Resources
-resources:
-  requests:
-    cpu: 50m
-    memory: 64Mi
-  limits:
-    cpu: 200m
-    memory: 128Mi

k8s/infra/databases/kustomization.yaml

@@ -8,5 +8,4 @@ resources:
   - namespaces.yaml
   - authentik-db.yaml
   - temporal-db.yaml
-  - memory-db.yaml
   - paperless-db.yaml

k8s/infra/logging/grafana-values.yaml

@@ -73,7 +73,7 @@ grafana.ini:
   auth.generic_oauth:
     enabled: true
     name: Authentik
-    allow_sign_up: true
+    allow_sign_up: true  # Allow sync to create/update users from OAuth claims
     client_id: grafana
     scopes: openid email profile groups
     auth_url: https://authentik.riotpiao.com/application/o/authorize/
@@ -84,13 +84,13 @@ grafana.ini:
     # Authentik doesn't serve — request 404s with "Error getting email address"
     # and the whole OAuth login fails.
     email_attribute_path: email
-    login_attribute_path: preferred_username
+    login_attribute_path: preferred_username  # Use preferred_username — OpenID standard claim, always present
     name_attribute_path: name
     role_attribute_path: "preferred_username == 'akadmin' && 'GrafanaAdmin' || contains(groups[*], 'homelab-admins') && 'Admin' || 'Viewer'"
     allow_assign_grafana_admin: true
     use_pkce: false
     use_refresh_token: false
-    skip_org_role_sync: false
+    skip_org_role_sync: false  # Sync org roles from OAuth groups
     tls_skip_verify_insecure: true   # Authentik uses self-signed cert; verify in prod
 
 # GF_AUTH_GENERIC_OAUTH_CLIENT_SECRET is injected from the grafana-oidc K8s

k8s/infra/longhorn/longhorn-add-disks-job.yaml

@@ -117,6 +117,22 @@ spec:
             echo "  ✓ talos-cp-3 already has $CP3_DISKS disks configured"
           fi
 
+          # worker-1: enable scheduling, reserve 400Gi for LLM weights,
+          # clear eviction flag that was set when scheduling was disabled.
+          echo "Configuring worker-1 disk..."
+          kubectl -n longhorn-system patch nodes.longhorn.io worker-1 --type merge -p '{
+            "spec": {
+              "disks": {
+                "default-disk-1030400000000": {
+                  "allowScheduling": true,
+                  "evictionRequested": false,
+                  "storageReserved": 429496729600
+                }
+              }
+            }
+          }'
+          echo "  ✓ worker-1 disk configured (400Gi reserved)"
+
           echo
           echo "Waiting for disks to be ready..."
           sleep 10

k8s/infra/longhorn/longhorn-nodes.yaml

@@ -1,33 +1,25 @@
-# Longhorn Node CRDs for cp-2 and cp-3.
+# Longhorn Node CRDs — enable scheduling on nodes that need explicit opt-in.
-# These nodes have the control-plane taint, so Longhorn doesn't auto-discover them.
-# Explicit Node CRDs + the taint-toleration setting enable storage across all 3 nodes.
 #
-# `spec.disks` is deliberately absent. Longhorn owns disk identity: it names the
+# Control-plane nodes (cp-2, cp-3) have taints so Longhorn doesn't auto-discover.
-# entry itself (`default-disk-080400000000`, not `default-disk`) and writes
+# worker-1 was manually disabled; re-enabled 2026-09-28 to relieve degraded-volume
-# `storageReserved`, `diskType` and `evictionRequested` into it. Declaring a
+# pressure (400Gi reserved for LLM weights via the add-disks-job, ~300Gi available
-# `default-disk` key here never matched the live one, so the Application sat
+# for general replica scheduling).
-# OutOfSync and selfHeal kept trying to add a SECOND disk record pointing at the
-# same /var/lib/longhorn path — which is worse than the drift it was fixing.
 #
-# What these objects are actually for is `allowScheduling: true` on tainted
+# `spec.disks` deliberately absent — Longhorn owns disk identity and writes
-# control-plane nodes. That is all they need to declare.
+# storageReserved/diskType/evictionRequested itself. Disk-level config lives in
+# longhorn-add-disks-job.yaml. These objects only set `allowScheduling: true`.
 ---
 apiVersion: longhorn.io/v1beta2
 kind: Node
-metadata:
+metadata: { name: talos-cp-2, namespace: longhorn-system }
-  name: talos-cp-2
+spec: { name: talos-cp-2, allowScheduling: true, tags: [] }
-  namespace: longhorn-system
-spec:
-  name: talos-cp-2
-  allowScheduling: true
-  tags: []
 ---
 apiVersion: longhorn.io/v1beta2
 kind: Node
-metadata:
+metadata: { name: talos-cp-3, namespace: longhorn-system }
-  name: talos-cp-3
+spec: { name: talos-cp-3, allowScheduling: true, tags: [] }
-  namespace: longhorn-system
+---
-spec:
+apiVersion: longhorn.io/v1beta2
-  name: talos-cp-3
+kind: Node
-  allowScheduling: true
+metadata: { name: worker-1, namespace: longhorn-system }
-  tags: []
+spec: { name: worker-1, allowScheduling: true, tags: [] }

k8s/infra/minio/minio-tenant.yaml

@@ -11,7 +11,12 @@ metadata:
     prometheus.io/port: "9000"
     prometheus.io/scrape: "true"
 spec:
-  image: minio/minio:RELEASE.2025-07-23T15-54-02Z
+  # Was minio/minio:RELEASE.2025-07-23T15-54-02Z — that Docker Hub repository no longer
+  # exists (404 for every tag), so the tenant sat in ImagePullBackOff and Vault's S3
+  # backend could not resolve minio-cluster-hl. Official images now need registry auth
+  # (quay.io/minio/minio returns 401), so this pins a public mirror of the same upstream
+  # release; swap it for the official image once MinIO registry credentials exist.
+  image: coollabsio/minio:RELEASE.2025-10-15T17-29-55Z
 
   # Disable operator auto-TLS: MinIO serves plain HTTP internally on 9000.
   # External TLS is terminated at nginx ingress (wildcard riotpiao-com-tls cert).
@@ -112,10 +117,16 @@ spec:
           name: minio-oidc
           key: MINIO_IDENTITY_OPENID_CLIENT_SECRET
     - name: MINIO_IDENTITY_OPENID_CLAIM_NAME
-      value: "policy"
+      value: "groups"
+    - name: MINIO_IDENTITY_OPENID_CLAIM_PREFIX
+      value: ""
     - name: MINIO_IDENTITY_OPENID_REDIRECT_URI
       value: "https://minio.riotpiao.com/oauth_callback"
     - name: MINIO_IDENTITY_OPENID_DISPLAY_NAME
       value: "Authentik"
     - name: MINIO_IDENTITY_OPENID_SCOPES
-      value: "openid,profile,email,minio"
+      value: "openid,profile,email,groups"
+    - name: MINIO_BROWSER_LOGIN_ANIMATION
+      value: "off"
+    - name: MINIO_BROWSER_REDIRECT_URL
+      value: "https://minio.riotpiao.com"

k8s/infra/monitoring/alerts/api-gateway-alerts.yaml

@@ -158,3 +158,66 @@ spec:
             severity: warning
           annotations:
             summary: "Gateway pod {{ $labels.pod }} memory at {{ $value | humanizePercentage }} of limit"
+
+    - name: api-gateway.admission
+      # The admission gate fails OPEN when its view of the engine gauges goes
+      # stale, so a broken observer degrades silently to the pre-#64 behaviour
+      # of admitting everything. These two series are the only detection —
+      # the gate has no automated tests.
+      rules:
+        # The gate fails OPEN when it cannot read the engine, so "no data" must
+        # alert: otherwise a missing emitter, a renamed metric, or a broken
+        # scrape is indistinguishable from a healthy observer. Both expressions
+        # therefore OR in an absent() guard on the metric itself.
+        - alert: APIGatewayObserverDegraded
+          expr: observer_degraded == 1 or absent(observer_degraded)
+          for: 2m
+          labels:
+            severity: warning
+          annotations:
+            summary: "Gateway admission observer degraded or not reporting — gate is failing open"
+            description: "observer_degraded is either 1 or absent. Absent means the emitter is missing, renamed, or unscraped, which is the fail-open case."
+
+        # Engine gauges went stale, or are not being reported at all
+        - alert: GatewayEngineStateStale
+          expr: increase(engine_state_stale_total[15m]) > 0 or absent(engine_state_stale_total)
+          for: 2m
+          labels:
+            severity: warning
+          annotations:
+            summary: "Gateway engine state stale or not reporting — admission may be based on stale gauges"
+            description: "Non-zero means the observer lost the engine; absent means the metric is not being emitted at all."
+
+        # The acceptance metric for reasoning.yaml's prefill change. Preemption
+        # of a 65k prompt costs a full re-prefill (~250s at the measured
+        # 262 tok/s), so any sustained rate is a regression, not noise.
+        - alert: LLMPreemptions
+          # Guarded absent() is deliberately NOT used bare here: the engine
+          # restarts (strategy Recreate) and the series legitimately vanishes
+          # for the length of a boot, so a bare absent() would false-fire on
+          # every rollout. It only fires when the metric is missing WHILE the
+          # predictor is up, which is the real "renamed or unscraped" case.
+          expr: |
+            increase(vllm:num_preemptions_total[15m]) > 0
+            or (absent(vllm:num_preemptions_total) and on() up{job="reasoning-predictor"} == 1)
+          for: 5m
+          labels:
+            severity: warning
+          annotations:
+            summary: "vLLM preempting requests ({{ $value }} in 15m) — KV cache over-committed, re-prefill cost is severe"
+            description: "Any sustained preemption rate is a regression for the reasoning prefill change: evicting a 65k-token prompt costs a full re-prefill (~250s at the measured 262 tok/s). Orchestrator verified `up{job=\"reasoning-predictor\"}` is the live job label — if that label is wrong the guard is inert but harmless."
+
+        # This change trades single-request latency for fairness: capping a long
+        # prompt at 2048 tokens/step means it takes more steps to finish, so a
+        # lone large prompt can now be slower even while preemptions stay 0.
+        # Without this alert the change has no way to show a regression.
+        - alert: LLMPrefillStarvation
+          expr: |
+            (sum(vllm:num_requests_waiting) > 0 and max(vllm:kv_cache_usage_perc) < 0.5)
+            or histogram_quantile(0.95, sum(rate(vllm:time_to_first_token_seconds_bucket[10m])) by (le)) > 60
+          for: 10m
+          labels:
+            severity: warning
+          annotations:
+            summary: "reasoning queueing or TTFT p95 above 60s while KV is under 50% — scheduling, not capacity"
+            description: "Queue depth with spare KV means requests are being starved rather than blocked on memory; a high TTFT p95 with headroom points at the same. Both series verified live (vllm:num_requests_waiting and vllm:time_to_first_token_seconds exist)."

k8s/infra/monitoring/alerts/cag-kv-cache-rules.yaml

@@ -0,0 +1,76 @@
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: cag-kv-cache-rules
+  namespace: monitoring
+  labels:
+    release: prometheus
+spec:
+  groups:
+    # Recording rules: derive vLLM prefix cache hit rate from raw counters.
+    # Metric names verified against live vLLM 0.11.0 scrape (2026-09-29).
+    # The previous names (vllm:cache_query_hit_total / vllm:cache_query_miss_total)
+    # do not exist in this version — vLLM exposes prefix_cache_hits_total and
+    # prefix_cache_queries_total — so this rule produced no data and
+    # VLLMPrefixCacheHitRateLow could never fire.
+    - name: vllm.recording
+      rules:
+        - record: vllm:prefix_cache_hit_rate
+          expr: |
+            rate(vllm:prefix_cache_hits_total[5m])
+            /
+            clamp_min(
+              rate(vllm:prefix_cache_queries_total[5m]),
+              0.001
+            )
+          labels:
+            job: "reasoning-predictor"
+
+    # Alert rules: CAG pipeline health
+    - name: cag.alerts
+      rules:
+        # KV warmup consistently failing — cache never gets warm
+        - alert: KVWarmupFailureHigh
+          expr: |
+            rate(memory_kv_cache_warmup_error_total[5m]) > 0.01
+          for: 5m
+          labels:
+            severity: warning
+            team: poimen
+          annotations:
+            summary: "KV warmup errors elevated ({{ $value | humanize }}/s)"
+            description: |
+              The KV cache warmer is failing to reach vLLM at a rate > 0.01/s
+              for 5 consecutive minutes. The prefix KV cache is not being warmed;
+              agents will see full prefill latency on every call.
+              Check reasoning-predictor health and poimen-memory NetworkPolicy.
+
+        # Warmup misses: vLLM reachable but not caching
+        - alert: KVWarmupMissHigh
+          expr: |
+            rate(memory_kv_cache_warmup_miss_total[5m]) > 0.01
+          for: 5m
+          labels:
+            severity: warning
+            team: poimen
+          annotations:
+            summary: "KV warmup cache misses elevated ({{ $value | humanize }}/s)"
+            description: |
+              vLLM is reachable but cached_tokens=0 on warmup responses.
+              Possible causes: prefix caching disabled in vLLM, max-num-seqs
+              saturated during warmup, or context format changed.
+
+        # Prefix cache hit rate low after warmup should be established
+        - alert: VLLMPrefixCacheHitRateLow
+          expr: |
+            vllm:prefix_cache_hit_rate < 0.70
+          for: 10m
+          labels:
+            severity: warning
+            team: poimen
+          annotations:
+            summary: "vLLM prefix cache hit rate low ({{ $value | humanizePercentage }})"
+            description: |
+              vLLM prefix cache hit rate below 70% for 10 minutes.
+              Expected >= 85% in steady state with CAG warming active.
+              Check KVWarmupFailureHigh / KVWarmupMissHigh alerts for root cause.

k8s/infra/monitoring/dashboards/generate-dashboards.py

@@ -26,20 +26,23 @@ def stat_panel(id, title, expr, x, y, w=4, h=4, unit="short", mappings=None, thr
     return p
 
 
-def ts_panel(id, title, exprs, x, y, w=8, h=8, unit="short"):
+def ts_panel(id, title, exprs, x, y, w=8, h=8, unit="short", description=None):
     targets = []
     for e in exprs:
         if isinstance(e, tuple):
             targets.append({"expr": e[0], "legendFormat": e[1]})
         else:
             targets.append({"expr": e, "legendFormat": "{{pod}}"})
-    return {
+    p = {
         "id": id, "title": title, "type": "timeseries",
         "gridPos": {"h": h, "w": w, "x": x, "y": y},
         "datasource": DS_PROM,
         "fieldConfig": {"defaults": {"unit": unit}},
         "targets": targets,
     }
+    if description:
+        p["description"] = description
+    return p
 
 
 def table_panel(id, title, expr, x, y, w=12, h=8):
@@ -265,6 +268,237 @@ def build_api_gateway():
     }
 
 
+# ============================================================================
+# Dashboard 4: LLM Observability (vLLM / TEI engines + gateway)
+# ============================================================================
+#
+# Two independent metric sources, deliberately both shown:
+#
+#   vLLM + TEI (llm-serving)  — engine-truth, but only for the three engines
+#                               that expose /metrics. Ollama (ornith:35b,
+#                               qwen2.5:3b-instruct) exposes nothing at all.
+#   api-gateway (/metrics)    — covers EVERY model including the two Ollama
+#                               ones, derived from the SSE byte stream: TTFT is
+#                               request-start to first byte, ITL is the gap
+#                               between SSE writes carrying "data: ", so it
+#                               counts SSE chunks rather than true tokens.
+#                               vLLM's own histograms remain authoritative for
+#                               `reasoning`.
+#
+# These three series ARE live as of 2026-09-29 — 3 series each, one per replica,
+# verified against the running gateway and Prometheus:
+#   llm_ttft_seconds, llm_itl_seconds, llm_tokens_total
+#
+# History, recorded so nobody re-derives it the hard way: on 2026-09-29 these
+# recorded nothing at all. internal/proxy/llm_metrics.go declared a capture
+# wrapper that was never constructed, so the exporter emitted HELP-only stubs.
+# homelab-frontend#63 fixed exactly that — the wrapper is now
+# ResponseWriterWrapper, constructed by NewResponseWriterWrapper
+# (internal/proxy/llm_metrics.go:38) and installed on the streaming proxy path.
+# The "dead code" note that used to live in this header is superseded.
+#
+# Panels backed by queue_full_requests_total, warmup_requests_delayed_total and
+# vllm_prefix_cache_hit_tokens_total also work (3 series, one per replica).
+#
+#
+# All metric names below were verified against a live scrape, not assumed:
+#   vLLM 0.11.0 exposes prefix_cache_hits_total / prefix_cache_queries_total,
+#   NOT cache_query_hit_total / cache_query_miss_total.
+
+def build_llm_observability():
+    def hq(metric, q, by="le"):
+        return (
+            f"histogram_quantile({q}, sum(rate({metric}_bucket[5m])) by ({by}, model_name))",
+            f"p{int(q * 100)}",
+        )
+
+    def hq_gw(metric, q):
+        return (
+            f"histogram_quantile({q}, sum(rate({metric}_bucket[5m])) by (le, model))",
+            f"p{int(q * 100)} · {{{{model}}}}",
+        )
+
+    def avg_of(metric):
+        # mean = rate(_sum) / rate(_count) — valid for cumulative histograms
+        return f"rate({metric}_sum[5m]) / clamp_min(rate({metric}_count[5m]), 0.001)"
+
+    panels = []
+
+    # ── TTFT ────────────────────────────────────────────────────────────────
+    panels.append(row(100, "Latency — Time To First Token", 0, [
+        stat_panel(101, "TTFT p95 (gateway)",
+                   "histogram_quantile(0.95, sum(rate(llm_ttft_seconds_bucket[5m])) by (le))",
+                   0, 1, w=4, h=4, unit="s"),
+        stat_panel(102, "TTFT p95 (vLLM)",
+                   "histogram_quantile(0.95, sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le))",
+                   4, 1, w=4, h=4, unit="s"),
+        stat_panel(103, "KV cache in use",
+                   "max(vllm:kv_cache_usage_perc)",
+                   8, 1, w=4, h=4, unit="percentunit"),
+        stat_panel(104, "Requests running",
+                   "sum(vllm:num_requests_running)",
+                   12, 1, w=4, h=4, unit="short"),
+        stat_panel(105, "Requests waiting",
+                   "sum(vllm:num_requests_waiting)",
+                   16, 1, w=4, h=4, unit="short"),
+        stat_panel(106, "Preemptions (5m)",
+                   "increase(vllm:num_preemptions_total[5m])",
+                   20, 1, w=4, h=4, unit="short"),
+        ts_panel(110, "TTFT percentile by model (gateway)",
+                 [hq_gw("llm_ttft_seconds", 0.50),
+                  hq_gw("llm_ttft_seconds", 0.95),
+                  hq_gw("llm_ttft_seconds", 0.99)],
+                 0, 2, w=12, h=8, unit="s"),
+        ts_panel(111, "TTFT percentile — vLLM reasoning",
+                 [hq("vllm:time_to_first_token_seconds", 0.50),
+                  hq("vllm:time_to_first_token_seconds", 0.95),
+                  hq("vllm:time_to_first_token_seconds", 0.99)],
+                 12, 2, w=12, h=8, unit="s"),
+        ts_panel(112, "Prefill vs decode (mean)",
+                 [(avg_of("vllm:request_prefill_time_seconds"), "prefill"),
+                  (avg_of("vllm:request_decode_time_seconds"), "decode")],
+                 0, 3, w=12, h=8, unit="s"),
+        ts_panel(113, "Queue wait before scheduling (mean)",
+                 [(avg_of("vllm:request_queue_time_seconds"), "queue")],
+                 12, 3, w=12, h=8, unit="s"),
+    ]))
+
+    # ── Inter-token latency ─────────────────────────────────────────────────
+    panels.append(row(120, "Latency — Inter-Token", 4, [
+        ts_panel(121, "Inter-token latency by model (gateway)",
+                 [hq_gw("llm_itl_seconds", 0.50),
+                  hq_gw("llm_itl_seconds", 0.95),
+                  hq_gw("llm_itl_seconds", 0.99)],
+                 0, 5, w=12, h=8, unit="s"),
+        ts_panel(122, "Inter-token latency — vLLM reasoning",
+                 [hq("vllm:inter_token_latency_seconds", 0.50),
+                  hq("vllm:inter_token_latency_seconds", 0.95),
+                  hq("vllm:inter_token_latency_seconds", 0.99)],
+                 12, 5, w=12, h=8, unit="s"),
+        ts_panel(123, "Decode throughput (tokens/s per request)",
+                 [(f"1 / clamp_min({avg_of('vllm:time_per_output_token_seconds')}, 0.0001)", "tok/s")],
+                 0, 6, w=12, h=8, unit="short"),
+        ts_panel(124, "End-to-end request latency p50/p95/p99",
+                 [hq("vllm:e2e_request_latency_seconds", 0.50),
+                  hq("vllm:e2e_request_latency_seconds", 0.95),
+                  hq("vllm:e2e_request_latency_seconds", 0.99)],
+                 12, 6, w=12, h=8, unit="s"),
+    ]))
+
+    # ── Batching / scheduling ───────────────────────────────────────────────
+    panels.append(row(130, "Batching & Scheduling", 7, [
+        ts_panel(131, "vLLM running vs waiting requests (batch depth)",
+                 [("sum(vllm:num_requests_running)", "running (in batch)"),
+                  ("sum(vllm:num_requests_waiting)", "waiting (queued)")],
+                 0, 8, w=12, h=8, unit="short"),
+        ts_panel(132, "Preemption rate (KV cache eviction pressure)",
+                 [("rate(vllm:num_preemptions_total[5m])", "preemptions/s")],
+                 12, 8, w=12, h=8, unit="short"),
+        ts_panel(133, "TEI batch size (embeddings / reranker)",
+                 [("avg(te_batch_next_size_sum / clamp_min(te_batch_next_size_count, 1)) by (app)",
+                   "{{app}}")],
+                 0, 9, w=12, h=8, unit="short"),
+        ts_panel(134, "TEI queue depth",
+                 [("max(te_queue_size) by (app)", "{{app}}")],
+                 12, 9, w=12, h=8, unit="short"),
+    ]))
+
+    # ── KV cache ────────────────────────────────────────────────────────────
+    panels.append(row(140, "KV Cache", 10, [
+        ts_panel(141, "KV cache utilisation (GPU memory pressure)",
+                 [("max(vllm:kv_cache_usage_perc) by (model_name)", "{{model_name}}")],
+                 0, 11, w=12, h=8, unit="percentunit"),
+        ts_panel(142, "Prefix cache hit rate (CAG effectiveness)",
+                 [("vllm:prefix_cache_hit_rate", "{{model_name}}")],
+                 12, 11, w=12, h=8, unit="percentunit"),
+        ts_panel(143, "Prefix cache hits vs queries",
+                 [("rate(vllm:prefix_cache_hits_total[5m])", "hits/s"),
+                  ("rate(vllm:prefix_cache_queries_total[5m])", "queries/s")],
+                 0, 12, w=12, h=8, unit="short"),
+        ts_panel(144, "Prefix cache hit tokens (gateway-observed, all models)",
+                 [("rate(vllm_prefix_cache_hit_tokens_total[5m])", "hit tokens/s")],
+                 12, 12, w=12, h=8, unit="short"),
+    ]))
+
+    # ── Throughput & health ─────────────────────────────────────────────────
+    panels.append(row(150, "Throughput & Health", 13, [
+        ts_panel(151, "Token throughput",
+                 [("sum(rate(vllm:prompt_tokens_total[5m]))", "prompt tok/s"),
+                  ("sum(rate(vllm:generation_tokens_total[5m]))", "generation tok/s"),
+                  ("rate(llm_tokens_total[5m])", "gateway SSE chunks/s (all models)")],
+                 0, 14, w=12, h=8, unit="short"),
+        ts_panel(152, "Successful requests/s",
+                 [("sum(rate(vllm:request_success_total[5m]))", "vLLM"),
+                  ("sum(rate(te_request_count[5m]))", "TEI")],
+                 12, 14, w=12, h=8, unit="short"),
+        ts_panel(153, "Backpressure — queue-full rejects",
+                 [("rate(queue_full_requests_total[5m])", "rejected/s")],
+                 0, 15, w=8, h=8, unit="short"),
+        ts_panel(154, "KV warmup requests delayed",
+                 [("rate(warmup_requests_delayed_total[5m])", "delayed/s")],
+                 8, 15, w=8, h=8, unit="short"),
+        ts_panel(155, "Embedding / rerank latency p95",
+                 [("histogram_quantile(0.95, sum(rate(te_embed_duration_bucket[5m])) by (le, app))",
+                   "{{app}} embed"),
+                  ("histogram_quantile(0.95, sum(rate(te_embed_queue_duration_bucket[5m])) by (le, app))",
+                   "{{app}} queue")],
+                 16, 15, w=8, h=8, unit="s"),
+    ]))
+
+    # ── Admission control ───────────────────────────────────────────────────
+    # These four panels read the gateway's admission gate, which is the
+    # companion change in homelab-frontend#64. That emitter is NOT deployed
+    # yet (verified against the live /metrics endpoint: 0 series matching
+    # ^(admission_|observer_|engine_state_)), so they render "No data" until
+    # #64 ships. That is expected, and each panel carries a description
+    # saying so — an unexplained blank panel would read as a broken
+    # dashboard, which is exactly the failure mode this row exists to catch.
+    panels.append(row(160, "Admission Control", 16, [
+        ts_panel(161, "Admission parked (5m)",
+                 [("increase(admission_parked_total[5m])", "parked")],
+                 0, 17, w=8, h=8, unit="short",
+                 description="Requests the gateway held back rather than admitting. "
+                             "Emitted by the admission gate (homelab-frontend#64); "
+                             "empty until that change is deployed."),
+        ts_panel(162, "Admission rejected by reason",
+                 [("sum by (reason) (increase(admission_rejected_total[5m]))", "{{reason}}")],
+                 8, 17, w=8, h=8, unit="short",
+                 description="Admission rejections split by the gate's own reason label. "
+                             "Emitted by homelab-frontend#64; empty until deployed."),
+        ts_panel(163, "Engine state stale",
+                 [("increase(engine_state_stale_total[5m])", "stale")],
+                 16, 17, w=8, h=8, unit="short",
+                 description="Times the gateway's view of the engine gauges went stale. "
+                             "The gate fails OPEN when this happens, so any non-zero rate "
+                             "means admission is silently degraded to admit-everything. "
+                             "Emitted by homelab-frontend#64; empty until deployed."),
+        ts_panel(164, "Gateway view vs engine truth",
+                 [("max(vllm_requests_running)", "gateway: running"),
+                  ("sum(vllm:num_requests_running)", "engine: running"),
+                  ("max(vllm_kv_usage_perc)", "gateway: kv (fraction)  [same as engine if healthy]"),
+                  ("max(vllm:kv_cache_usage_perc)", "engine: kv (fraction)")],
+                 0, 18, w=24, h=8, unit="short",
+                 description="The gateway's own reading of the engine against the engine's "
+                             "self-reported value. A disagreement between a pair is how a "
+                             "stale observer or a mistuned admission target shows itself, "
+                             "and nothing else in this dashboard reveals it. The two "
+                             "gateway series (vllm_requests_running, vllm_kv_usage_perc) "
+                             "ship with homelab-frontend#64 and are absent until it is "
+                             "deployed — until then this panel shows the engine's pair only."),
+    ]))
+
+    return {
+        "title": "LLM Observability",
+        "uid": "llm-observability",
+        "schemaVersion": 39,
+        "timezone": "browser",
+        "time": {"from": "now-6h", "to": "now"},
+        "refresh": "30s",
+        "tags": ["llm", "vllm", "tei", "gpu", "observability"],
+        "panels": panels,
+    }
+
+
 # ============================================================================
 # Generate
 # ============================================================================
@@ -277,4 +511,7 @@ write_dashboard("cluster-infrastructure.yaml", build_cluster_infrastructure(), "
 # Dashboard 3
 write_dashboard("api-gateway.yaml", build_api_gateway(), "API")
 
+# Dashboard 4
+write_dashboard("llm-observability.yaml", build_llm_observability(), "LLM")
+
 print("Done.")

k8s/infra/monitoring/dashboards/llm-observability.yaml

@@ -0,0 +1,95 @@
+apiVersion: v1
+data:
+  llm-observability.json: '{"title":"LLM Observability","uid":"llm-observability","schemaVersion":39,"timezone":"browser","time":{"from":"now-6h","to":"now"},"refresh":"30s","tags":["llm","vllm","tei","gpu","observability"],"panels":[{"id":100,"title":"Latency
+    \u2014 Time To First Token","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":0},"panels":[{"id":101,"title":"TTFT
+    p95 (gateway)","type":"stat","gridPos":{"h":4,"w":4,"x":0,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.95,
+    sum(rate(llm_ttft_seconds_bucket[5m])) by (le))"}]},{"id":102,"title":"TTFT p95
+    (vLLM)","type":"stat","gridPos":{"h":4,"w":4,"x":4,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.95,
+    sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le))"}]},{"id":103,"title":"KV
+    cache in use","type":"stat","gridPos":{"h":4,"w":4,"x":8,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"percentunit"}},"targets":[{"expr":"max(vllm:kv_cache_usage_perc)"}]},{"id":104,"title":"Requests
+    running","type":"stat","gridPos":{"h":4,"w":4,"x":12,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum(vllm:num_requests_running)"}]},{"id":105,"title":"Requests
+    waiting","type":"stat","gridPos":{"h":4,"w":4,"x":16,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum(vllm:num_requests_waiting)"}]},{"id":106,"title":"Preemptions
+    (5m)","type":"stat","gridPos":{"h":4,"w":4,"x":20,"y":1},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"increase(vllm:num_preemptions_total[5m])"}]},{"id":110,"title":"TTFT
+    percentile by model (gateway)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":2},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.5,
+    sum(rate(llm_ttft_seconds_bucket[5m])) by (le, model))","legendFormat":"p50 \u00b7
+    {{model}}"},{"expr":"histogram_quantile(0.95, sum(rate(llm_ttft_seconds_bucket[5m]))
+    by (le, model))","legendFormat":"p95 \u00b7 {{model}}"},{"expr":"histogram_quantile(0.99,
+    sum(rate(llm_ttft_seconds_bucket[5m])) by (le, model))","legendFormat":"p99 \u00b7
+    {{model}}"}]},{"id":111,"title":"TTFT percentile \u2014 vLLM reasoning","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":2},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.5,
+    sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p50"},{"expr":"histogram_quantile(0.95,
+    sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p95"},{"expr":"histogram_quantile(0.99,
+    sum(rate(vllm:time_to_first_token_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p99"}]},{"id":112,"title":"Prefill
+    vs decode (mean)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":3},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"rate(vllm:request_prefill_time_seconds_sum[5m])
+    / clamp_min(rate(vllm:request_prefill_time_seconds_count[5m]), 0.001)","legendFormat":"prefill"},{"expr":"rate(vllm:request_decode_time_seconds_sum[5m])
+    / clamp_min(rate(vllm:request_decode_time_seconds_count[5m]), 0.001)","legendFormat":"decode"}]},{"id":113,"title":"Queue
+    wait before scheduling (mean)","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":3},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"rate(vllm:request_queue_time_seconds_sum[5m])
+    / clamp_min(rate(vllm:request_queue_time_seconds_count[5m]), 0.001)","legendFormat":"queue"}]}]},{"id":120,"title":"Latency
+    \u2014 Inter-Token","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":4},"panels":[{"id":121,"title":"Inter-token
+    latency by model (gateway)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":5},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.5,
+    sum(rate(llm_itl_seconds_bucket[5m])) by (le, model))","legendFormat":"p50 \u00b7
+    {{model}}"},{"expr":"histogram_quantile(0.95, sum(rate(llm_itl_seconds_bucket[5m]))
+    by (le, model))","legendFormat":"p95 \u00b7 {{model}}"},{"expr":"histogram_quantile(0.99,
+    sum(rate(llm_itl_seconds_bucket[5m])) by (le, model))","legendFormat":"p99 \u00b7
+    {{model}}"}]},{"id":122,"title":"Inter-token latency \u2014 vLLM reasoning","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":5},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.5,
+    sum(rate(vllm:inter_token_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p50"},{"expr":"histogram_quantile(0.95,
+    sum(rate(vllm:inter_token_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p95"},{"expr":"histogram_quantile(0.99,
+    sum(rate(vllm:inter_token_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p99"}]},{"id":123,"title":"Decode
+    throughput (tokens/s per request)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":6},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"1
+    / clamp_min(rate(vllm:time_per_output_token_seconds_sum[5m]) / clamp_min(rate(vllm:time_per_output_token_seconds_count[5m]),
+    0.001), 0.0001)","legendFormat":"tok/s"}]},{"id":124,"title":"End-to-end request
+    latency p50/p95/p99","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":6},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.5,
+    sum(rate(vllm:e2e_request_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p50"},{"expr":"histogram_quantile(0.95,
+    sum(rate(vllm:e2e_request_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p95"},{"expr":"histogram_quantile(0.99,
+    sum(rate(vllm:e2e_request_latency_seconds_bucket[5m])) by (le, model_name))","legendFormat":"p99"}]}]},{"id":130,"title":"Batching
+    & Scheduling","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":7},"panels":[{"id":131,"title":"vLLM
+    running vs waiting requests (batch depth)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":8},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum(vllm:num_requests_running)","legendFormat":"running
+    (in batch)"},{"expr":"sum(vllm:num_requests_waiting)","legendFormat":"waiting
+    (queued)"}]},{"id":132,"title":"Preemption rate (KV cache eviction pressure)","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":8},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"rate(vllm:num_preemptions_total[5m])","legendFormat":"preemptions/s"}]},{"id":133,"title":"TEI
+    batch size (embeddings / reranker)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":9},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"avg(te_batch_next_size_sum
+    / clamp_min(te_batch_next_size_count, 1)) by (app)","legendFormat":"{{app}}"}]},{"id":134,"title":"TEI
+    queue depth","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":9},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"max(te_queue_size)
+    by (app)","legendFormat":"{{app}}"}]}]},{"id":140,"title":"KV Cache","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":10},"panels":[{"id":141,"title":"KV
+    cache utilisation (GPU memory pressure)","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":11},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"percentunit"}},"targets":[{"expr":"max(vllm:kv_cache_usage_perc)
+    by (model_name)","legendFormat":"{{model_name}}"}]},{"id":142,"title":"Prefix
+    cache hit rate (CAG effectiveness)","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":11},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"percentunit"}},"targets":[{"expr":"vllm:prefix_cache_hit_rate","legendFormat":"{{model_name}}"}]},{"id":143,"title":"Prefix
+    cache hits vs queries","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":12},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"rate(vllm:prefix_cache_hits_total[5m])","legendFormat":"hits/s"},{"expr":"rate(vllm:prefix_cache_queries_total[5m])","legendFormat":"queries/s"}]},{"id":144,"title":"Prefix
+    cache hit tokens (gateway-observed, all models)","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":12},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"rate(vllm_prefix_cache_hit_tokens_total[5m])","legendFormat":"hit
+    tokens/s"}]}]},{"id":150,"title":"Throughput & Health","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":13},"panels":[{"id":151,"title":"Token
+    throughput","type":"timeseries","gridPos":{"h":8,"w":12,"x":0,"y":14},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum(rate(vllm:prompt_tokens_total[5m]))","legendFormat":"prompt
+    tok/s"},{"expr":"sum(rate(vllm:generation_tokens_total[5m]))","legendFormat":"generation
+    tok/s"},{"expr":"rate(llm_tokens_total[5m])","legendFormat":"gateway SSE chunks/s
+    (all models)"}]},{"id":152,"title":"Successful requests/s","type":"timeseries","gridPos":{"h":8,"w":12,"x":12,"y":14},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum(rate(vllm:request_success_total[5m]))","legendFormat":"vLLM"},{"expr":"sum(rate(te_request_count[5m]))","legendFormat":"TEI"}]},{"id":153,"title":"Backpressure
+    \u2014 queue-full rejects","type":"timeseries","gridPos":{"h":8,"w":8,"x":0,"y":15},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"rate(queue_full_requests_total[5m])","legendFormat":"rejected/s"}]},{"id":154,"title":"KV
+    warmup requests delayed","type":"timeseries","gridPos":{"h":8,"w":8,"x":8,"y":15},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"rate(warmup_requests_delayed_total[5m])","legendFormat":"delayed/s"}]},{"id":155,"title":"Embedding
+    / rerank latency p95","type":"timeseries","gridPos":{"h":8,"w":8,"x":16,"y":15},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"s"}},"targets":[{"expr":"histogram_quantile(0.95,
+    sum(rate(te_embed_duration_bucket[5m])) by (le, app))","legendFormat":"{{app}}
+    embed"},{"expr":"histogram_quantile(0.95, sum(rate(te_embed_queue_duration_bucket[5m]))
+    by (le, app))","legendFormat":"{{app}} queue"}]}]},{"id":160,"title":"Admission
+    Control","type":"row","collapsed":true,"gridPos":{"h":1,"w":24,"x":0,"y":16},"panels":[{"id":161,"title":"Admission
+    parked (5m)","type":"timeseries","gridPos":{"h":8,"w":8,"x":0,"y":17},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"increase(admission_parked_total[5m])","legendFormat":"parked"}],"description":"Requests
+    the gateway held back rather than admitting. Emitted by the admission gate (homelab-frontend#64);
+    empty until that change is deployed."},{"id":162,"title":"Admission rejected by
+    reason","type":"timeseries","gridPos":{"h":8,"w":8,"x":8,"y":17},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"sum
+    by (reason) (increase(admission_rejected_total[5m]))","legendFormat":"{{reason}}"}],"description":"Admission
+    rejections split by the gate''s own reason label. Emitted by homelab-frontend#64;
+    empty until deployed."},{"id":163,"title":"Engine state stale","type":"timeseries","gridPos":{"h":8,"w":8,"x":16,"y":17},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"increase(engine_state_stale_total[5m])","legendFormat":"stale"}],"description":"Times
+    the gateway''s view of the engine gauges went stale. The gate fails OPEN when
+    this happens, so any non-zero rate means admission is silently degraded to admit-everything.
+    Emitted by homelab-frontend#64; empty until deployed."},{"id":164,"title":"Gateway
+    view vs engine truth","type":"timeseries","gridPos":{"h":8,"w":24,"x":0,"y":18},"datasource":{"type":"prometheus","uid":"prometheus"},"fieldConfig":{"defaults":{"unit":"short"}},"targets":[{"expr":"max(vllm_requests_running)","legendFormat":"gateway:
+    running"},{"expr":"sum(vllm:num_requests_running)","legendFormat":"engine: running"},{"expr":"max(vllm_kv_usage_perc)","legendFormat":"gateway:
+    kv (fraction)  [same as engine if healthy]"},{"expr":"max(vllm:kv_cache_usage_perc)","legendFormat":"engine:
+    kv (fraction)"}],"description":"The gateway''s own reading of the engine against
+    the engine''s self-reported value. A disagreement between a pair is how a stale
+    observer or a mistuned admission target shows itself, and nothing else in this
+    dashboard reveals it. The two gateway series (vllm_requests_running, vllm_kv_usage_perc)
+    ship with homelab-frontend#64 and are absent until it is deployed \u2014 until
+    then this panel shows the engine''s pair only."}]}]}'
+kind: ConfigMap
+metadata:
+  annotations:
+    grafana_folder: LLM
+  labels:
+    grafana_dashboard: '1'
+  name: llm-observability-dashboard
+  namespace: logging

k8s/infra/monitoring/kustomization.yaml

@@ -14,10 +14,14 @@ resources:
   - alerts/svc-vault-rules.yaml
   - alerts/cluster-alerts.yaml
   - alerts/api-gateway-alerts.yaml
+  - alerts/cag-kv-cache-rules.yaml
   - servicemonitors/argocd.yaml
   - servicemonitors/authentik.yaml
   - servicemonitors/forgejo.yaml
   - servicemonitors/minio.yaml
   - servicemonitors/ingress-nginx.yaml
+  - servicemonitors/vllm.yaml
+  - servicemonitors/api-gateway.yaml
   - dashboards/cluster-infrastructure.yaml
   - dashboards/api-gateway.yaml
+  - dashboards/llm-observability.yaml

k8s/infra/monitoring/prometheus-values.yaml

@@ -54,8 +54,8 @@ alertmanager:
 # ── Prometheus ────────────────────────────────────────────────────────────────
 prometheus:
   prometheusSpec:
-    retention: 15d
+    retention: 30d
-    retentionSize: "18GB"
+    retentionSize: "95GB"
 
     # scrapeTimeout MUST be <= scrapeInterval or the operator refuses to generate
     # the Prometheus config ("scrapeTimeout greater than scrapeInterval") and no
@@ -64,10 +64,9 @@ prometheus:
     scrapeTimeout: 60s
     evaluationInterval: 30s
 
-    # Pin to az-a (talos-cp-1) — sole Longhorn node; else the RWO PVC can't
+    # Pin to worker-1 for more storage capacity
-    # attach on cp-2/cp-3 (CSINode lacks driver.longhorn.io).
     nodeSelector:
-      topology.kubernetes.io/zone: az-a
+      kubernetes.io/hostname: worker-1
 
     # Persistent storage — metrics survive node reboots and pod restarts.
     # Uses the default `longhorn` StorageClass; nodeSelector above already pins
@@ -79,7 +78,7 @@ prometheus:
           accessModes: ["ReadWriteOnce"]
           resources:
             requests:
-              storage: 20Gi
+              storage: 25Gi
 
     resources:
       requests:

k8s/infra/monitoring/servicemonitors/api-gateway.yaml

@@ -0,0 +1,62 @@
+# History, recorded so nobody re-derives it the hard way: on 2026-09-29 the
+# three latency series below recorded nothing. internal/proxy/llm_metrics.go
+# declared a capture wrapper (LLMMetricsCapture) that was never constructed
+# anywhere, so the exporter emitted HELP-only stubs and the dashboard panels
+# stayed empty.
+#
+# homelab-frontend#63 fixed exactly that. The wrapper is now
+# ResponseWriterWrapper, constructed by NewResponseWriterWrapper
+# (internal/proxy/llm_metrics.go:38) and installed on the streaming proxy
+# request path. Verified live on 2026-09-29:
+#   llm_ttft_seconds  count=7  sum=598.212s
+#   llm_itl_seconds   count=16116
+#   llm_tokens_total  3 series
+#
+# These derived-from-SSE series cover EVERY model including the two Ollama ones
+# (ornith:35b, qwen2.5:3b-instruct), which expose no metrics endpoint at all.
+# They are proxy-side: TTFT is request-start to first byte and ITL is the gap
+# between SSE writes, so both are granular to SSE write boundaries rather than
+# to the engine's own token cadence. llm_tokens_total counts `data:` events,
+# and the documented assumption in internal/proxy/llm_metrics.go is that these
+# upstreams emit one such event per token — so it is a token count only as far
+# as that assumption holds. vLLM's own histograms remain authoritative.
+#
+# Also recorded and working (3 series, one per replica):
+#   queue_full_requests_total           counter    backpressure rejects
+#   warmup_requests_delayed_total       counter    vLLM slot contention
+#   vllm_prefix_cache_hit_tokens_total  counter    prefix-cache hits
+#
+# NOT working, do not build on it: request_duration_seconds is declared and
+# exported but has 0 samples, and its only writer Metrics.RecordRequest has
+# zero call sites outside _test.go — verified by ast-grep, not assumed. It is
+# the same dead-code shape this branch already corrected for the latency
+# series, so it is listed here rather than left to be rediscovered.
+#
+# Also note the Deployment already carried prometheus.io/scrape annotations,
+# but kube-prometheus-stack ignores pod annotations (no additionalScrapeConfigs,
+# serviceMonitorSelectorNilUsesHelmValues: false) — so nothing at all was being
+# collected before this ServiceMonitor existed.
+apiVersion: monitoring.coreos.com/v1
+kind: ServiceMonitor
+metadata:
+  name: api-gateway
+  namespace: monitoring
+  labels:
+    app.kubernetes.io/name: api-gateway
+    app.kubernetes.io/part-of: api
+spec:
+  namespaceSelector:
+    matchNames:
+      - api
+  selector:
+    matchLabels:
+      app: api-gateway
+  endpoints:
+    # api-gateway Service port IS named `http` (unlike the KServe ones), so
+    # the name form works here.
+    - port: http
+      interval: 30s
+      scrapeTimeout: 10s
+      path: /metrics
+      scheme: http
+

k8s/infra/monitoring/servicemonitors/vllm.yaml

@@ -11,20 +11,39 @@ spec:
     matchNames:
       - llm-serving
   selector:
-    matchLabels:
+    # Only the three engines that actually serve /metrics.
-      app.kubernetes.io/part-of: llm-serving
+    # ornith-predictor and qwen-cpu are Ollama, which has no Prometheus
+    # endpoint (/metrics returns 404) — selecting them would add permanently
+    # DOWN targets. Ollama-model visibility comes from the api-gateway
+    # ServiceMonitor instead (llm_ttft_seconds / llm_itl_seconds are
+    # measured at the proxy, so they cover every model including Ollama).
+    matchExpressions:
+      - key: app.kubernetes.io/name
+        operator: In
+        values:
+          - llm-reasoning   # vLLM 0.11.0 — full engine metrics
+          - llm-embeddings  # TEI 1.8.2 — CPU
+          - llm-reranker    # TEI 1.8.2 — CPU
   endpoints:
-    - port: http
+    # targetPort, not port: the KServe-generated Services on llm-serving have
+    # UNNAMED ports, so `port: <name>` can never match. Naming them would mean
+    # editing the InferenceServices (predictor restart = full weights reload,
+    # tens of seconds on the vLLM slot), so scrape the pod port directly.
+    # prometheus-operator v0.94.1 supports targetPort on ServiceMonitor.
+    - targetPort: 8080
       interval: 30s
       scrapeTimeout: 10s
       path: /metrics
       scheme: http
       relabelings:
-        - sourceLabels: [__meta_kubernetes_namespace]
+        # Prometheus's Kubernetes service discovery already supplies the
-          targetLabel: namespace
+        # namespace, pod and service labels; the only one that needs deriving
-        - sourceLabels: [__meta_kubernetes_pod_name]
+        # is `app`, since the TEI panels aggregate by it (by (app)) and SD does
-          targetLabel: pod
+        # not promote arbitrary pod labels.
-        - sourceLabels: [__meta_kubernetes_service_name]
+        #
-          targetLabel: service
+        # `action: replace` is written explicitly. It is the API server's
-        - sourceLabels: [__meta_kubernetes_pod_label_app_kubernetes_io_name]
+        # default for relabeling rules, so omitting it leaves the live object
+        # diverged from this manifest and ArgoCD reports permanent OutOfSync.
+        - action: replace
+          sourceLabels: [__meta_kubernetes_pod_label_app_kubernetes_io_name]
           targetLabel: app

k8s/infra/poimen/kustomization.yaml

@@ -0,0 +1,11 @@
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+# Poimen infrastructure — CNPG Postgres for memory service.
+# Part of the `poimen` ArgoCD Application (multi-source, source 3).
+# Sync-wave -1 ensures the DB cluster is healthy before application
+# deployments (wave 0) are applied.
+resources:
+  - memory-db.yaml
+  - poimen-sandbox-namespace.yaml
+  - poimen-sandbox-quota.yaml
+  - pi-session-netpol.yaml

k8s/infra/poimen/memory-db.yaml

@@ -0,0 +1,37 @@
+# Dedicated CNPG Postgres for Poimen Memory (GitOps, wave 2).
+# Uses default longhorn storage class (3 replicas, dataLocality disabled).
+# CNPG generates secret `memory-db-app` + service `memory-db-rw` in ns poimen.
+apiVersion: postgresql.cnpg.io/v1
+kind: Cluster
+metadata:
+  name: memory-db
+  namespace: poimen
+  annotations:
+    argocd.argoproj.io/sync-wave: "-1"
+    argocd.argoproj.io/sync-options: SkipDryRunOnMissingResource=true
+spec:
+  instances: 3
+  imageName: ghcr.io/cloudnative-pg/postgresql:16.2
+  bootstrap:
+    initdb:
+      database: memory
+      owner: app
+      encoding: UTF8
+      localeCollate: C
+      localeCType: C
+      postInitApplicationSQL:
+        - "CREATE EXTENSION vector;"
+  enableSuperuserAccess: false
+  resources:
+    requests: { memory: "512Mi", cpu: "250m" }
+    limits: { memory: "2Gi", cpu: "1" }
+  storage:
+    size: 20Gi
+    storageClass: longhorn
+  affinity:
+    podAntiAffinityType: preferred
+    topologyKey: kubernetes.io/hostname
+    tolerations:
+      - key: node-role.kubernetes.io/control-plane
+        operator: Exists
+        effect: NoSchedule

k8s/infra/poimen/pi-session-netpol.yaml

@@ -0,0 +1,12 @@
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: pi-session-deny-kube-api
+  namespace: poimen-sandbox
+spec:
+  endpointSelector:
+    matchLabels:
+      app: pi-session
+  egressDeny:
+    - toEntities:
+        - kube-apiserver

k8s/infra/poimen/poimen-sandbox-namespace.yaml

@@ -0,0 +1,9 @@
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: poimen-sandbox
+  labels:
+    pod-security.kubernetes.io/enforce: restricted
+    pod-security.kubernetes.io/enforce-version: latest
+    pod-security.kubernetes.io/audit: restricted
+    pod-security.kubernetes.io/warn: restricted

k8s/infra/poimen/poimen-sandbox-quota.yaml

@@ -0,0 +1,12 @@
+apiVersion: v1
+kind: ResourceQuota
+metadata:
+  name: sandbox-quota
+  namespace: poimen-sandbox
+spec:
+  hard:
+    pods: "20"
+    requests.cpu: "10"
+    requests.memory: 20Gi
+    limits.cpu: "20"
+    limits.memory: 40Gi

k8s/infra/rbac/kustomization.yaml

@@ -16,6 +16,7 @@ resources:
   - temporal-operator-role.yaml
   - llm-serving-operator-role.yaml
   - poimen-memory-operator-role.yaml
+  - sandbox-observer-role.yaml
 # paperless's Role/RoleBinding lives in k8s/apps/paperless/rbac.yaml instead -
 # that app already has its own kustomization + namespace, no need to
 # duplicate it here. All of these stay inert (grant nothing) until

k8s/infra/rbac/sandbox-observer-role.yaml

@@ -0,0 +1,23 @@
+apiVersion: rbac.authorization.k8s.io/v1
+kind: Role
+metadata:
+  name: sandbox-observer
+  namespace: poimen-sandbox
+rules:
+  - apiGroups: [""]
+    resources: [pods, pods/log, services]
+    verbs: [get, list, watch]
+---
+apiVersion: rbac.authorization.k8s.io/v1
+kind: RoleBinding
+metadata:
+  name: poimen-devs-sandbox
+  namespace: poimen-sandbox
+subjects:
+  - kind: Group
+    apiGroup: rbac.authorization.k8s.io
+    name: "oidc:poimen-devs"
+roleRef:
+  kind: Role
+  name: sandbox-observer
+  apiGroup: rbac.authorization.k8s.io

k8s/infra/tracing/api-latency-dashboard.yaml

@@ -123,6 +123,156 @@ data:
               "custom": {"fillOpacity": 10}
             }
           }
+        },
+        {
+          "id": 7,
+          "title": "vLLM Prefix Cache Hit Rate",
+          "type": "timeseries",
+          "gridPos": {"h": 8, "w": 12, "x": 0, "y": 36},
+          "datasource": {"type": "prometheus", "uid": "prometheus"},
+          "targets": [
+            {
+              "expr": "vllm:prefix_cache_hit_rate",
+              "legendFormat": "prefix cache hit rate"
+            }
+          ],
+          "fieldConfig": {
+            "defaults": {
+              "unit": "percentunit",
+              "min": 0,
+              "max": 1,
+              "custom": {"fillOpacity": 15},
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {"color": "red", "value": 0},
+                  {"color": "yellow", "value": 0.7},
+                  {"color": "green", "value": 0.85}
+                ]
+              }
+            }
+          }
+        },
+        {
+          "id": 8,
+          "title": "KV Warmup Hit vs Miss",
+          "type": "timeseries",
+          "gridPos": {"h": 8, "w": 12, "x": 12, "y": 36},
+          "datasource": {"type": "prometheus", "uid": "prometheus"},
+          "targets": [
+            {
+              "expr": "rate(memory_kv_cache_warmup_hit_total[5m])",
+              "legendFormat": "warmup hit"
+            },
+            {
+              "expr": "rate(memory_kv_cache_warmup_miss_total[5m])",
+              "legendFormat": "warmup miss"
+            },
+            {
+              "expr": "rate(memory_kv_cache_warmup_error_total[5m])",
+              "legendFormat": "warmup error"
+            }
+          ],
+          "fieldConfig": {
+            "defaults": {
+              "unit": "ops",
+              "custom": {"fillOpacity": 10}
+            }
+          }
+        },
+        {
+          "id": 9,
+          "title": "KV Warmup Latency p50/p95 (ms)",
+          "type": "timeseries",
+          "gridPos": {"h": 8, "w": 12, "x": 0, "y": 44},
+          "datasource": {"type": "prometheus", "uid": "prometheus"},
+          "targets": [
+            {
+              "expr": "histogram_quantile(0.50, sum(rate(memory_kv_cache_warmup_latency_ms_bucket[5m])) by (le))",
+              "legendFormat": "p50 warmup latency"
+            },
+            {
+              "expr": "histogram_quantile(0.95, sum(rate(memory_kv_cache_warmup_latency_ms_bucket[5m])) by (le))",
+              "legendFormat": "p95 warmup latency"
+            }
+          ],
+          "fieldConfig": {
+            "defaults": {
+              "unit": "ms",
+              "custom": {"fillOpacity": 10}
+            }
+          }
+        },
+        {
+          "id": 10,
+          "title": "Standing Context Cache Tier Hits",
+          "type": "timeseries",
+          "gridPos": {"h": 8, "w": 12, "x": 12, "y": 44},
+          "datasource": {"type": "prometheus", "uid": "prometheus"},
+          "targets": [
+            {
+              "expr": "rate(memory_standing_ctx_hit_redis_total[5m])",
+              "legendFormat": "Redis L1 hit"
+            },
+            {
+              "expr": "rate(memory_standing_ctx_hit_db_total[5m])",
+              "legendFormat": "DB L2 hit"
+            },
+            {
+              "expr": "rate(memory_standing_ctx_recompute_total[5m])",
+              "legendFormat": "recompute (L3)"
+            },
+            {
+              "expr": "rate(memory_standing_ctx_304_total[5m])",
+              "legendFormat": "304 Not Modified"
+            }
+          ],
+          "fieldConfig": {
+            "defaults": {
+              "unit": "ops",
+              "custom": {"fillOpacity": 10}
+            }
+          }
+        },
+        {
+          "id": 11,
+          "title": "Standing Context Alert: Stale > 10min post-ingest",
+          "type": "timeseries",
+          "description": "Alert if standing context was not refreshed within 10min of an ingest event. Indicates staleness propagation failure.",
+          "gridPos": {"h": 8, "w": 24, "x": 0, "y": 52},
+          "datasource": {"type": "prometheus", "uid": "prometheus"},
+          "targets": [
+            {
+              "expr": "rate(memory_standing_ctx_recompute_total[5m]) > 0",
+              "legendFormat": "recompute rate (non-zero = fresh)"
+            },
+            {
+              "expr": "rate(memory_kv_cache_warmup_miss_total[5m])",
+              "legendFormat": "warmup miss rate (alert if > 3/5min)"
+            }
+          ],
+          "fieldConfig": {
+            "defaults": {
+              "unit": "ops",
+              "custom": {"fillOpacity": 10},
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {"color": "green", "value": 0},
+                  {"color": "red", "value": 0.01}
+                ]
+              }
+            },
+            "overrides": [
+              {
+                "matcher": {"id": "byName", "options": "warmup miss rate (alert if > 3/5min)"},
+                "properties": [
+                  {"id": "thresholds", "value": {"mode": "absolute", "steps": [{"color": "green", "value": 0}, {"color": "red", "value": 0.01}]}},
+                  {"id": "custom.thresholdsStyle", "value": {"mode": "area"}}
+                ]
+              }
+            ]
+          }
         }
       ]
     }

scripts/iam/provision-rbac.py

@@ -145,6 +145,7 @@ SERVICE_ACCOUNTS: Dict[str, Dict[str, Any]] = {
 SCOPE_MAPPINGS: Dict[str, Dict[str, str]] = {
     "roles": {"expression": 'return user.attributes.get("roles", [])'},
     "permissions": {"expression": 'return ["*"] if any(user.groups.filter(is_superuser=True)) else list(user.groups.values_list("name", flat=True))'},
+    "groups": {"expression": 'return {"groups": list(user.groups.values_list("name", flat=True))}'},
     "minio_buckets": {"expression": 'return user.attributes.get("minio_buckets", [])'},
     "paperless_doctypes": {"expression": 'return user.attributes.get("paperless_doctypes", [])'},
     "memory_projects": {"expression": 'return user.attributes.get("memory_projects", [])'},
@@ -246,6 +247,25 @@ for scope_name, scope_spec in SCOPE_MAPPINGS.items():
         else:
             print(f"  {scope_name}: FAILED {status} {res}")
 
+# Fix built-in 'openid' scope to include preferred_username (needed by Grafana, Forgejo, etc.)
+# Note: sub MUST be a string, not int, for JSON compatibility with Go struct unmarshaling
+if "openid" in existing_scopes:
+    openid_scope = existing_scopes["openid"]
+    openid_pk = openid_scope["pk"]
+    openid_expression = 'return {"sub": str(user.pk), "preferred_username": user.username}'
+    if openid_expression not in openid_scope.get("expression", ""):
+        status, res = api("PUT", f"/api/v3/propertymappings/provider/scope/{openid_pk}/", {
+            "name": openid_scope["name"],
+            "scope_name": "openid",
+            "expression": openid_expression,
+        })
+        if status in (200, 201):
+            print(f"  openid: updated (sub as string + preferred_username)")
+        else:
+            print(f"  openid: UPDATE FAILED {status} {res}")
+    else:
+        print(f"  openid: already correct (sub as string)")
+
 # ===========================================================================
 # Phase 4: Get flow UUIDs (needed for providers)
 # ===========================================================================

terraform/files/coredns/Corefile

@@ -38,6 +38,8 @@
     rewrite name api.riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
     rewrite name comfy.riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
     rewrite name comfyui.riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
+    rewrite name gotify.riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
+    rewrite name igotify.riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
     rewrite name riotpiao.com ingress-nginx-controller.ingress-nginx.svc.cluster.local
 
     kubernetes cluster.local in-addr.arpa ip6.arpa {