feat(monitoring): canary the CephPoolLostRedundancy join (REN-93)

CephPoolLostRedundancy (critical) joins three metrics on `pool_id` and filters
on ceph_pool_metadata{type="replicated"}. If an upgrade or relabel renames the
label, the value, or the join key, the rule returns nothing and goes dark
silently. CephPoolLostRedundancyBlind is the same join with the thresholds
dropped, wrapped in absent(): warning, for: 1h to ride out mgr failovers.

Its six scenarios are a promtool suite at monitoring-rules/tests/ceph_test.yaml,
with five negative controls in mutations.txt, run by the existing `promtool`
step (#1026). The CI half of the original #989 is dropped: #1026 covers it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.woodpecker/alert-monitor.yml

@@ -101,7 +101,7 @@ steps:
         bash scripts/ci/make-diag-kubeconfig.sh
 
   - name: alert-triage-agent
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     environment:
       CLAUDE_CODE_OAUTH_TOKEN:
         from_secret: claude_code_oauth_token

.woodpecker/build-ci-images.yml

@@ -350,17 +350,17 @@ steps:
           --destination gitea.renner.uno/gitea_admin/heating-can:$TAG \
           --destination gitea.renner.uno/gitea_admin/heating-can:latest
 
-  # netbird-proxy-services: the controller that reconciles NetBird reverse-proxy
+  # netbird-service-sync: the controller that reconciles NetBird reverse-proxy
   # services (public Ingress hosts -> TLS passthrough -> Traefik ClusterIP).
   # Runtime artifact like heating-can; same registry so the audit tracks it.
-  - name: build-netbird-proxy-services
+  - name: build-netbird-service-sync
     image: registry.gitlab.com/gitlab-ci-utils/container-images/kaniko:v1.25.18-debug
-    directory: ci/images/netbird-proxy-services
+    directory: ci/images/netbird-service-sync
     depends_on: []
     when:
       - event: deployment
       - event: push
-        path: ["ci/images/netbird-proxy-services/**", ".woodpecker/build-ci-images.yml"]
+        path: ["ci/images/netbird-service-sync/**", ".woodpecker/build-ci-images.yml"]
     environment:
       GITEA_USER: gitea_admin
       GITEA_TOKEN:
@@ -374,39 +374,8 @@ steps:
         /kaniko/executor \
           --context . \
           --dockerfile Containerfile \
-          --destination gitea.renner.uno/gitea_admin/netbird-proxy-services:$TAG \
-          --destination gitea.renner.uno/gitea_admin/netbird-proxy-services:latest
-
-  # wiki-index: embedding index over the Notes repo's wiki/, fed from Gitea and
-  # queried by ~/Notes/bin/wiki_rank.py. Runtime artifact like netbird-proxy-services.
-  - name: build-wiki-index
-    image: registry.gitlab.com/gitlab-ci-utils/container-images/kaniko:v1.25.18-debug
-    directory: ci/images/wiki-index
-    depends_on: []
-    when:
-      - event: deployment
-      - event: push
-        path: ["ci/images/wiki-index/**", ".woodpecker/build-ci-images.yml"]
-    environment:
-      GITEA_USER: gitea_admin
-      GITEA_TOKEN:
-        from_secret: ci_images_token
-    commands:
-      - mkdir -p /kaniko/.docker
-      - AUTH=$(printf '%s:%s' "$GITEA_USER" "$GITEA_TOKEN" | base64 | tr -d '\n')
-      - printf '{"auths":{"gitea.renner.uno":{"auth":"%s"}}}' "$AUTH" > /kaniko/.docker/config.json
-      - TAG=$(date +%Y-%m-%d)
-      # Also a per-commit tag: several builds a day (branch pushes and the
-      # merge) all rewrite the dated tag, and the deployment pin raced them
-      # on 2026-09-20. The commit tag is immutable and needs no digest lookup.
-      - SHORT=$(printf '%s' "$CI_COMMIT_SHA" | cut -c1-8)
-      - |
-        /kaniko/executor \
-          --context . \
-          --dockerfile Containerfile \
-          --destination gitea.renner.uno/gitea_admin/wiki-index:$TAG \
-          --destination gitea.renner.uno/gitea_admin/wiki-index:$TAG-$SHORT \
-          --destination gitea.renner.uno/gitea_admin/wiki-index:latest
+          --destination gitea.renner.uno/gitea_admin/netbird-service-sync:$TAG \
+          --destination gitea.renner.uno/gitea_admin/netbird-service-sync:latest
 
   # bird: BGP daemon for the pod-network mesh over pod0 (#576). Runtime
   # artifact like heating-can; same registry so the audit tracks it.

.woodpecker/ci-claude-refresh.yml

@@ -14,8 +14,8 @@
 #                 and  ci-claude:latest.  The Containerfile makes this layer
 #                 byte-reproducible, so rebuilding an unchanged version dedups
 #                 in the registry instead of orphaning a fresh ~260 MB copy.
-#   3. bump-pr  — (only when the version actually changed) rewrite the 5
-#                 consumer pins to  ci-claude:<version>  and open a PR. A human
+#   3. bump-pr  — (only when any pin is off the latest version) rewrite every
+#                 consumer pin to  ci-claude:<version>  and open a PR. A human
 #                 reviews and merges it; nothing auto-merges.
 #
 # TAG SCHEME — deliberately version tags (ci-claude:<claude-version>), NOT the dated
@@ -107,31 +107,22 @@ steps:
           [0-9]*.[0-9]*.[0-9]*) : ;;
           *) echo "npm returned an implausible version '$$NEW' — aborting."; exit 1 ;;
         esac
-        # The tag every consumer currently references. They are kept identical,
-        # so this must resolve to EXACTLY ONE non-empty tag; anything else
-        # (no match, or a partial/disagreeing set) aborts rather than risk a
-        # corrupting sed later. Counting with `grep -c .` (non-empty lines)
-        # instead of `wc -l` is deliberate: an empty result yields 0 here but
-        # 1 under wc -l, which would slip past the guard.
-        CUR=$$(grep -hoE 'ci-claude:[0-9A-Za-z._-]+' $$CONSUMERS | sed 's/^ci-claude://' | sort -u)
-        NTAGS=$$(printf '%s' "$$CUR" | grep -c . || true)
-        if [ -z "$$CUR" ] || [ "$$NTAGS" -ne 1 ]; then
-          echo "Expected exactly one ci-claude tag across the consumers, got [$$CUR]."
-          echo "Refusing to guess — fix the pins by hand."; exit 1
-        fi
-        echo "latest=$$NEW  current=$$CUR  event=$$CI_PIPELINE_EVENT"
+        # Distinct tags the consumers currently reference. Divergent pins are
+        # normal (a pipeline bumped on its own); the bump rewrites all of them.
+        CUR=$$(grep -hoE 'ci-claude:[0-9][0-9A-Za-z._-]*' $$CONSUMERS | sed 's/^ci-claude://' | sort -u | tr '\n' ' ' | sed 's/ $$//')
+        [ -n "$$CUR" ] || { echo "No ci-claude pin found in the consumers — aborting."; exit 1; }
+        echo "latest=$$NEW  current=[$$CUR]  event=$$CI_PIPELINE_EVENT"
         printf '%s' "$$NEW" > "$$CI_WORKSPACE/.ci-refresh/version"
         printf '%s' "$$CUR" > "$$CI_WORKSPACE/.ci-refresh/current"
-        # Build when the version changed, or on a push (Containerfile edit must
-        # ship even if the version is unchanged). Bump the pins only when the
-        # version the consumers reference is not the latest.
+        # Build when any pin is off latest, or on a push (Containerfile edit must
+        # ship even if the version is unchanged). Bump when any pin is off latest.
         if [ "$$CUR" != "$$NEW" ] || [ "$$CI_PIPELINE_EVENT" = "push" ]; then
           touch "$$CI_WORKSPACE/.ci-refresh/do-build"; echo "will build $$NEW"
         else
           echo "already on $$NEW — nothing to build"
         fi
         if [ "$$CUR" != "$$NEW" ]; then
-          touch "$$CI_WORKSPACE/.ci-refresh/do-bump"; echo "will open bump PR $$CUR -> $$NEW"
+          touch "$$CI_WORKSPACE/.ci-refresh/do-bump"; echo "will open bump PR [$$CUR] -> $$NEW"
         else
           echo "consumers already pin $$NEW — no PR"
         fi
@@ -175,6 +166,9 @@ steps:
       - |
         [ -f "$$CI_WORKSPACE/.ci-refresh/do-bump" ] || { echo "no version change — no PR"; exit 0; }
         set -eu
+        # Secret values can carry a trailing newline; it breaks the push URL and the auth header.
+        PUSH_TOKEN=$$(printf '%s' "$$PUSH_TOKEN" | tr -d '\r\n')
+        [ -n "$$PUSH_TOKEN" ] || { echo "PUSH_TOKEN is empty — aborting"; exit 1; }
         NEW=$$(cat "$$CI_WORKSPACE/.ci-refresh/version")
         CUR=$$(cat "$$CI_WORKSPACE/.ci-refresh/current")
         # Both were validated non-empty by the decide step; re-assert before a
@@ -190,9 +184,11 @@ steps:
         # is not a maintained list.
         CONSUMERS=$$(grep -rlE 'ci-claude:[0-9]' .woodpecker)
         [ -n "$$CONSUMERS" ] || { echo "No ci-claude consumers found — refusing to continue."; exit 1; }
-        # Exact-string swap of the current tag for the new one, consumers only.
+        # NEW goes into a sed replacement: allow only a plain x.y.z version.
+        printf '%s' "$$NEW" | grep -qE '^[0-9]+\.[0-9]+\.[0-9]+$$' || { echo "refusing non-x.y.z version '$$NEW'"; exit 1; }
+        # Every version-pinned tag becomes NEW, whatever it was; consumers only.
         for f in $$CONSUMERS; do
-          sed -i "s#ci-claude:$$CUR#ci-claude:$$NEW#g" "$$f"
+          sed -i "s#ci-claude:[0-9][0-9A-Za-z._-]*#ci-claude:$$NEW#g" "$$f"
         done
         # Author is claude-bot (the PAT's account); the commit subject makes
         # the automated-bump nature clear.

.woodpecker/ci-manifests-test.yml

@@ -17,30 +17,44 @@
 # here is a static check that either passes or fails on the checkout alone, so it
 # can gate PRs without becoming a bottleneck or a flake source.
 when:
-  # Path-filtered on purpose, unlike ci-scripts-test.yml. That pipeline is one
-  # step on one already-cached image, so running it everywhere is free; this one
-  # pulls three images and builds every kustomization in the tree, which is real
+  # Path-filtered on purpose, unlike ci-scripts-test.yml. That pipeline is two
+  # steps on already-cached images, so running it everywhere is free; this one
+  # pulls four images and builds every kustomization in the tree, which is real
   # cost to put on a docs-only PR. The filter covers everything the steps actually read:
-  # kubernetes/** for the manifests and scripts, .woodpecker/** for the lint (and
-  # so an edit to this file re-runs its own checks).
+  # kubernetes/** for the manifests, scripts and promtool suites, .woodpecker/**
+  # for the lint (and so an edit to this file re-runs its own checks), and the
+  # promtool driver script by name.
+  #
+  # That last entry is NOT covered by kubernetes/**: promtool-rules.sh lives
+  # under scripts/, so without it a PR that only edits the script would never run
+  # the check it drives. It is listed as one file rather than scripts/ci/**
+  # because the rest of that directory is CI-agent helpers with nothing to do
+  # with this pipeline, and pulling four images for them is the cost this filter
+  # exists to avoid. Their own tests run in ci-scripts-test.yml on every PR.
   - event: pull_request
     path:
       - "kubernetes/**"
       - ".woodpecker/**"
+      - "scripts/ci/promtool-rules.sh"
+      - "scripts/ci/yaml-citations.sh"
 
-# ADDING A THIRD DISCOVERY CHECK: kustomize-build and shellcheck both hand-roll
-# the same shape — `find`, a floor guard against empty discovery, then a
-# file-fed `while read`. Two copies is deliberate: the two traps being commented
-# around (BusyBox find lacking `-printf`, and a piped loop running in a subshell
-# where `fail` cannot propagate) are different, and both are worth spelling out
-# where they bite. A third instance is the point to stop copying and factor it
-# into a shared script under `scripts/ci/` — which is already covered by
-# ci-scripts-test.yml, so it would gain unit tests in the process.
+# THE THIRD DISCOVERY CHECK WENT TO A SCRIPT. kustomize-build and shellcheck
+# both hand-roll the same shape — `find`, a floor guard against empty discovery,
+# then a file-fed `while read`. Two copies was deliberate: the two traps being
+# commented around (BusyBox find lacking `-printf`, and a piped loop running in
+# a subshell where `fail` cannot propagate) are different, and both are worth
+# spelling out where they bite. An earlier revision of this comment said a third
+# instance would be the point to stop copying and factor it into `scripts/ci/`,
+# where it would gain unit tests — and that is what `promtool` below does. Its
+# discovery, floors and guards live in scripts/ci/promtool-rules.sh and are
+# tested by scripts/ci/test-promtool-rules.sh on every PR via
+# ci-scripts-test.yml. A FOURTH inline copy should go the same way; do not add
+# one here.
 #
-# Every step carries `depends_on: []` so all four run in PARALLEL and every
+# Every step carries `depends_on: []` so all six run in PARALLEL and every
 # failure shows in one run. Woodpecker's default is sequential, which would stop
 # at the first failure and make a contributor fix-and-repush to discover the
-# next — and these four are independent, cheap, offline checks with no ordering
+# next — and these six are independent, cheap, offline checks with no ordering
 # between them. Keep it on every step: `depends_on` is per-step, so dropping it
 # from one silently re-serialises that step behind the others.
 steps:
@@ -258,3 +272,72 @@ steps:
           exit 1
         fi
         echo "ok    no digest pins under openbao/rotation/"
+
+  # Machine-check the alert rules: parse every PrometheusRule, run the committed
+  # promtool unit suites, then damage the rules on purpose and require the suites
+  # to notice. Everything it does lives in scripts/ci/promtool-rules.sh; read
+  # that file and kubernetes/apps/monitoring-rules/tests/README.md for the why.
+  #
+  # The gap it closes: the other four steps prove a PrometheusRule assembles into
+  # valid YAML and nothing more — not that an expression parses, not that a label
+  # join matches anything, not that a `for` window does what its comment claims.
+  # On REN-129 the CNPG surge rule was written `for: 2m`, which fires on six of
+  # seven measured nights and would have mailed the owner twice a night at 04:16
+  # and 04:22 local for a backlog that clears itself. Nothing here would have
+  # caught it, and the arithmetic that did catch it was hand-written from scratch
+  # twice, in two runs, and thrown away both times because there was nowhere in
+  # the repo to commit it. REN-152 retired that rule; its `for` margin is now
+  # guarded by mutations.txt:backlog-for-12m.
+  #
+  # Offline, like the rest of this pipeline: no cluster, no registry, no secrets,
+  # no network. promtool reads files and exits.
+  #
+  # IMAGE — prom/prometheus, not a registry-local ci-* image, and pinned to an
+  # exact upstream version rather than a dated tag. The dated-tag convention
+  # guarded by no-digest-pins above is about OUR images, whose digests are
+  # garbage-collected out from under a pin; this is upstream and immutable.
+  #
+  # The version must be 3.x, and that is not a preference. The cluster runs
+  # whatever kube-prometheus-stack >= 72.0.0 ships, which is Prometheus 3, and
+  # alert `for` evaluation is version-sensitive — a 2.x promtool would be
+  # checking a different evaluator than production, so the margins the suites
+  # assert would be claims about the wrong program. promtool-rules.sh asserts the
+  # major version it was handed and fails on 2.x, so a careless bump here is a
+  # red step rather than a quietly wrong check. Keep this at the latest stable
+  # 3.x, per the standing rule that any software uses the latest stable version.
+  #
+  # This image is NOT covered by image-automation (same reason as
+  # woodpecker-lint above: a Woodpecker pipeline file has no apiVersion/kind, so
+  # an $imagepolicy marker here is silently ignored — fluxcd/flux2#2826). The
+  # bump is manual.
+  #
+  # `prom/prometheus` is busybox-based and ships /bin/sh, so it can run a
+  # `commands:` block — unlike the plain woodpecker-cli tag, which cannot. Its
+  # find/grep/awk/sed are BusyBox, which is why promtool-rules.sh avoids
+  # `find -printf` and `grep -r --include`; it was verified against BusyBox
+  # 1.35.0 as well as GNU. The image also declares `USER nobody`, so the script
+  # writes its stripped mirror under mktemp and never into the workspace.
+  - name: promtool
+    depends_on: [] # see note above `steps:`
+    image: prom/prometheus:v3.15.0
+    commands:
+      - sh scripts/ci/promtool-rules.sh
+
+  # Comments cite other YAML blocks by anchor name, never by line number
+  # (REN-183; the exact syntax is in CONTRIBUTING.md, not repeated here because
+  # an example written in this comment would itself be checked). Line numbers drift on any edit
+  # above them and nothing noticed: three drift fixes were merged by hand
+  # before this step existed. This fails on a new line-number citation and on
+  # a citation whose file or anchor marker does not exist.
+  # Existence only — it never judges whether the anchored block still says
+  # what the citing comment claims, so ordinary edits cannot turn it red.
+  #
+  # It scans every YAML file in the checkout, but this pipeline only runs on
+  # the paths in `when:` above, so an edit confined to ansible/ or esphome/ is
+  # checked on the next PR that does touch kubernetes/ or .woodpecker/.
+  # Self-test: scripts/ci/test-yaml-citations.sh, in ci-scripts-test.yml.
+  - name: yaml-citations
+    depends_on: [] # see note above `steps:`
+    image: gitea.renner.uno/gitea_admin/ci-k8s:2026-09-08
+    commands:
+      - sh scripts/ci/yaml-citations.sh

.woodpecker/ci-scripts-test.yml

@@ -8,15 +8,45 @@
 # against adversarial PR input. Running it here fails the offending PR's checks
 # before it can merge.
 #
+# scripts/ci/test-promtool-rules.sh is the same idea for the alert-rule gate.
+# What it tests is not whether promtool evaluates PromQL — ci-manifests-test.yml
+# does that with the real binary against the real rules — but whether
+# promtool-rules.sh REFUSES TO REPORT SUCCESS when it has checked nothing: zero
+# manifests discovered, zero suites discovered, a `rule_files` path that does not
+# resolve (promtool warns, loads zero rules, and passes every assertion), a
+# mutation the suites no longer catch. Those are the states a green pipeline
+# cannot tell apart from real coverage, so they need a test that asserts the
+# failure. It stubs promtool, so it needs neither the binary nor the heavy image.
+#
+# Both are unfiltered by path on purpose: this pipeline is two steps on one
+# already-cached image, so running it everywhere is free, and a guard whose test
+# only runs when someone remembers to touch the right directory is not a guard.
+#
 # Only in homelab (scripts/ci/ lives here); the file is not deployed to the
-# mirrored repos, so it does not run on their PRs. The test is offline (no token,
-# no network, no cluster) — the ci-claude image just provides sh + jq, and it is
-# already pulled by pr-review.yml on the same event, so the runner has it cached.
+# mirrored repos, so it does not run on their PRs. All tests here are offline (no
+# token, no network, no cluster) — the ci-claude image just provides sh + jq, and
+# it is already pulled by pr-review.yml on the same event, so the runner has it
+# cached.
 when:
   - event: pull_request
 
+# `depends_on: []` on every step so they run in parallel and one failure does not hide
+# the other. Same reasoning as the note above `steps:` in ci-manifests-test.yml.
 steps:
   - name: model-selection-tests
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    depends_on: []
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     commands:
       - sh scripts/ci/test-model-selection.sh
+
+  - name: promtool-rules-guard-tests
+    depends_on: []
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    commands:
+      - sh scripts/ci/test-promtool-rules.sh
+
+  - name: yaml-citations-guard-tests
+    depends_on: []
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    commands:
+      - sh scripts/ci/test-yaml-citations.sh

.woodpecker/deploy-netbird-server.yml

@@ -4,9 +4,6 @@ when:
     branch: master
     path:
       - "ansible/roles/netbird-server/**"
-      # install-netbird-server.yml also applies the netbird-proxy-watchdog role,
-      # so changes to it must redeploy this host too.
-      - "ansible/roles/netbird-proxy-watchdog/**"
       - "ansible/plays/install-netbird-server.yml"
       - ".woodpecker/deploy-netbird-server.yml"
   - event: deployment

.woodpecker/e2e.yml

@@ -26,7 +26,7 @@ steps:
     # fetch-models.sh response mapping. No cluster, no secrets, no network — the
     # ci-claude image just provides sh + jq. Keeps the flag-injection regression
     # test from rotting (it exercises the exact injection shapes).
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     commands:
       - sh scripts/ci/test-model-selection.sh
 

.woodpecker/issue-fix.yml

@@ -62,7 +62,7 @@ steps:
       - bash scripts/ci/make-diag-kubeconfig.sh
 
   - name: issue-fix-agent
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     environment:
       CLAUDE_CODE_OAUTH_TOKEN:
         from_secret: claude_code_oauth_token

.woodpecker/pr-review.yml

@@ -65,7 +65,7 @@ when:
 
 steps:
   - name: pr-review-agent
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     environment:
       CLAUDE_CODE_OAUTH_TOKEN:
         from_secret: claude_code_oauth_token

.woodpecker/security-scan.yml

@@ -70,7 +70,7 @@ steps:
     # this step is skipped — otherwise we'd fork-bomb PRs.
     when:
       - branch: master
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     environment:
       CLAUDE_CODE_OAUTH_TOKEN:
         from_secret: claude_code_oauth_token

.woodpecker/update-audit.yml

@@ -30,7 +30,7 @@ when:
 
 steps:
   - name: update-audit-agent
-    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.250
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
     environment:
       CLAUDE_CODE_OAUTH_TOKEN:
         from_secret: claude_code_oauth_token

.woodpecker/wiki-maintenance.yml

@@ -0,0 +1,198 @@
+# Nightly wiki maintenance for the Notes vault (k.renner/notes), replacing the
+# workstation's wiki-maintenance and wiki-consolidate timers: distill the raw/
+# backlog into wiki/ (wiki-distill agent), audit wiki/ with the vault's own
+# script (structural checks, no model), then consolidate (wiki-consolidate
+# agent: verifies a cohort of notes against sources, repos and the live
+# cluster and edits them). The pipeline commits and pushes as Claude Bot; the
+# push webhook then refreshes the wiki service.
+#
+# The agents address the vault as ~/Notes and the repos as ~/source/<repo>, so
+# each agent step links those into the workspace instead of rewriting them.
+when:
+  - event: cron
+    cron: wiki-maintenance
+  # Hand-started runs: `woodpecker pipeline create k.renner/homelab --branch master`.
+  # The typesafe_api_key secret is scoped to cron and manual, not deployment.
+  - event: manual
+
+steps:
+  - name: checkout-notes
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    pull: true
+    environment:
+      WIKI_BOT_TOKEN:
+        from_secret: wiki_bot_token
+    commands:
+      - |
+        set -eu
+        # Token only on the wire, never in .git/config. Secrets set from files carry
+        # a trailing newline (run 10220: "could not read Username"); strip CR/LF.
+        TOKEN=$$(printf '%s' "$$WIKI_BOT_TOKEN" | tr -d '\r\n')
+        git -c http.extraheader="AUTHORIZATION: token $$TOKEN" \
+          clone --depth 50 --branch main https://gitea.renner.uno/k.renner/notes.git notes
+        echo "raw backlog: $$(ls notes/raw/*.md | grep -v _index.md | wc -l) file(s)"
+        # Full history for the consolidate agent's `git log` checks; the
+        # workspace clone is shallow and tree-filtered.
+        mkdir -p sources
+        git clone --quiet https://gitea.renner.uno/k.renner/homelab.git sources/homelab
+        git -c http.extraheader="AUTHORIZATION: token $$TOKEN" \
+          clone --quiet https://gitea.renner.uno/k.renner/homeassistant.git sources/homeassistant
+
+  - name: fetch-diag
+    # Read-only live-cluster access for the consolidate agent, same mechanism
+    # as issue-fix.yml: ./.ci-diag/kubectl plus a 1h ci-diagnostics token.
+    image: gitea.renner.uno/gitea_admin/ci-k8s:2026-09-08
+    environment:
+      SOPS_AGE_KEY:
+        from_secret: sops_age_key
+      S3_ACCESS_KEY_ID:
+        from_secret: hetzner_s3_access_key
+      S3_SECRET_ACCESS_KEY:
+        from_secret: hetzner_s3_secret_key
+    commands:
+      - bash scripts/ci/fetch-kubeconfig.sh
+      - bash scripts/ci/make-diag-kubeconfig.sh
+
+  - name: distill
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    pull: true
+    environment:
+      CLAUDE_CODE_OAUTH_TOKEN:
+        from_secret: claude_code_oauth_token
+      TYPESAFE_API_KEY:
+        from_secret: typesafe_api_key
+      WIKI_INDEX_BASIC_AUTH:
+        from_secret: wiki_index_basic_auth
+    commands:
+      - |
+        set -eu
+        # An empty queue must not start the agent: its startup context alone
+        # costs more than the whole audit.
+        BACKLOG=$$(ls notes/raw/*.md | grep -v _index.md | wc -l)
+        if [ "$$BACKLOG" -eq 0 ]; then
+          echo "distill: raw/ is empty, skipping the agent"
+          exit 0
+        fi
+        ln -sfn "$$PWD/notes" "$$HOME/Notes"
+        mkdir -p "$$HOME/.config/wiki-index"
+        printf '%s' "$$WIKI_INDEX_BASIC_AUTH" | tr -d '\r\n' > "$$HOME/.config/wiki-index/basic-auth"
+        chmod 600 "$$HOME/.config/wiki-index/basic-auth"
+        "$$HOME/Notes/bin/wiki_rank.py" --probe-index
+        MODEL=$$(sh scripts/ci/pick-model.sh "distill raw source notes into atomic wiki notes: read sources, check the wiki for existing coverage, write or merge notes; careful writing, moderate reasoning")
+        if [ -n "$$MODEL" ]; then echo "distill: selected model $$MODEL"; else echo "distill: using the CLI default model"; fi
+        cd notes
+        cat ../ci/agents/wiki-maintenance/distill.md | claude $${MODEL:+--model $$MODEL} \
+          --agent wiki-distill \
+          --permission-mode acceptEdits \
+          --allowedTools "Read,Edit,Write,Glob,Grep,Bash(/root/Notes/bin/wiki_rank.py:*),Bash(/root/Notes/bin/wiki_ask.py:*),Bash(mv /root/Notes/raw/:*),Bash(mkdir -p /root/Notes/:*),Bash(~/Notes/bin/wiki_rank.py:*),Bash(~/Notes/bin/wiki_ask.py:*),Bash(mv ~/Notes/raw/:*),Bash(mkdir -p ~/Notes/:*)" \
+          --settings ../ci/agents/wiki-maintenance/settings.json \
+          -p
+
+  - name: audit
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    pull: true
+    commands:
+      # Findings are report lines in output/, not a pipeline failure.
+      - python3 -B notes/bin/wiki_audit.py --vault "$$PWD/notes"
+
+  - name: consolidate
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    pull: true
+    environment:
+      CLAUDE_CODE_OAUTH_TOKEN:
+        from_secret: claude_code_oauth_token
+      TYPESAFE_API_KEY:
+        from_secret: typesafe_api_key
+      WIKI_INDEX_BASIC_AUTH:
+        from_secret: wiki_index_basic_auth
+    commands:
+      - |
+        set -eu
+        ln -sfn "$$PWD/notes" "$$HOME/Notes"
+        ln -sfn "$$PWD/sources" "$$HOME/source"
+        mkdir -p "$$HOME/.config/wiki-index"
+        printf '%s' "$$WIKI_INDEX_BASIC_AUTH" | tr -d '\r\n' > "$$HOME/.config/wiki-index/basic-auth"
+        chmod 600 "$$HOME/.config/wiki-index/basic-auth"
+        if [ -f .ci-diag/kubeconfig ]; then
+          ln -sf "$$PWD/.ci-diag/kubectl" /usr/local/bin/kubectl
+          export KUBECONFIG="$$PWD/.ci-diag/kubeconfig"
+        else
+          echo "consolidate: no ci-diag kubeconfig, live cluster checks unavailable"
+        fi
+        cd notes
+        # JSON output: the agent's report and the run's cost come out of one
+        # record at the end instead of a 45-minute silent log.
+        # One source for the Bash allowlist: the settings file (run 10608: the
+        # agent writes ~/Notes, the rules carried only /root/Notes).
+        BASH_RULES=$$(jq -r '[.permissions.allow[] | select(startswith("Bash("))] | join(",")' ../ci/agents/wiki-maintenance/consolidate-settings.json)
+        rc=0
+        cat ../ci/agents/wiki-maintenance/consolidate.md | claude \
+          --agent wiki-consolidate \
+          --permission-mode acceptEdits \
+          --allowedTools "Read,Edit,Write,Glob,Grep,Agent,$$BASH_RULES" \
+          --settings ../ci/agents/wiki-maintenance/consolidate-settings.json \
+          --output-format json \
+          -p > /tmp/consolidate.json || rc=$$?
+        # A killed or crashed claude leaves no valid JSON; report rc, not a jq error.
+        if [ -s /tmp/consolidate.json ] && jq -e . /tmp/consolidate.json > /dev/null; then
+          jq -r '.result // empty' /tmp/consolidate.json
+          jq -c '{duration_ms, num_turns, total_cost_usd, is_error}' /tmp/consolidate.json
+        else
+          echo "consolidate: no result record (exit $$rc)"; head -c 2000 /tmp/consolidate.json || true
+        fi
+        echo "consolidate: git log --oneline of the run:"
+        git log --oneline origin/main..HEAD
+        exit $$rc
+
+  - name: push-notes
+    image: gitea.renner.uno/gitea_admin/ci-claude:2.1.280
+    pull: true
+    environment:
+      WIKI_BOT_TOKEN:
+        from_secret: wiki_bot_token
+    commands:
+      - |
+        set -eu
+        cd notes
+        git config user.email "claude-bot@renner.uno"
+        git config user.name "Claude Bot"
+        # Only the vault content is committed; anything an agent leaves elsewhere
+        # in the checkout (run 10362: scratch scripts in the root) stays behind.
+        # The consolidate agent commits its own snapshots, so HEAD may already be
+        # ahead of origin/main with a clean tree.
+        git add -A -- wiki raw output .wiki-consolidation-state.json
+        git diff --cached --quiet || git commit -q -m "wiki maintenance $$(date -u +%F): distill + audit + consolidate (CI run $$CI_PIPELINE_NUMBER)"
+        if [ "$$(git rev-list --count origin/main..HEAD)" -eq 0 ]; then
+          echo "wiki-maintenance: no changes"
+          exit 0
+        fi
+        # The workstation pushes main on its own (notes-push.path), so main moves
+        # while the agents run (run 10241: non-fast-forward). The run's work goes
+        # to a branch and a PR that this step merges through Gitea; a conflict
+        # leaves the PR open for a human and fails the run.
+        BRANCH="wiki-maintenance/$$(date -u +%F)-$$CI_PIPELINE_NUMBER"
+        git checkout -q -b "$$BRANCH"
+        TOKEN=$$(printf '%s' "$$WIKI_BOT_TOKEN" | tr -d '\r\n')
+        git -c http.extraheader="AUTHORIZATION: token $$TOKEN" push origin "HEAD:$$BRANCH"
+        API="https://gitea.renner.uno/api/v1/repos/k.renner/notes"
+        jq -n --arg title "wiki maintenance $$(date -u +%F) (CI run $$CI_PIPELINE_NUMBER)" \
+              --arg body "Nightly distill + audit + consolidate. Merged by the pipeline; open only if the merge failed." \
+              --arg head "$$BRANCH" '{title: $$title, body: $$body, head: $$head, base: "main"}' > /tmp/pr.json
+        code=$$(curl -s -o /tmp/pr-resp.json -w '%{http_code}' -X POST \
+          -H "Authorization: token $$TOKEN" -H "Content-Type: application/json" --data @/tmp/pr.json "$$API/pulls")
+        echo "POST pull -> HTTP $$code"
+        [ "$$code" -ge 200 ] && [ "$$code" -lt 300 ] || { cat /tmp/pr-resp.json; exit 1; }
+        NUM=$$(jq -r '.number' /tmp/pr-resp.json)
+        # Merge without delete_branch_after_merge: run 10254 got HTTP 500 from that
+        # call although the merge itself had landed. Verify the PR state instead of
+        # trusting the status code, then drop the branch as a separate, best-effort step.
+        code=$$(curl -s -o /tmp/merge-resp.json -w '%{http_code}' -X POST \
+          -H "Authorization: token $$TOKEN" -H "Content-Type: application/json" \
+          --data '{"Do":"merge"}' "$$API/pulls/$$NUM/merge")
+        echo "POST merge #$$NUM -> HTTP $$code"
+        MERGED=$$(curl -s -H "Authorization: token $$TOKEN" "$$API/pulls/$$NUM" | jq -r '.merged')
+        echo "PR #$$NUM merged: $$MERGED"
+        [ "$$MERGED" = "true" ] || { cat /tmp/merge-resp.json; echo "PR #$$NUM left open"; exit 1; }
+        code=$$(curl -s -o /tmp/del-resp.json -w '%{http_code}' -X DELETE -H "Authorization: token $$TOKEN" \
+          "$$API/branches/$$(printf '%s' "$$BRANCH" | sed 's|/|%2F|g')")
+        echo "DELETE branch $$BRANCH -> HTTP $$code (best effort)"

CONTRIBUTING.md

@@ -0,0 +1,7 @@
+# Contributing
+
+## Citing other YAML from a comment
+
+- Mark the block you cite with a comment line `# anchor: <name>` (kebab-case, unique in that file). If the block is inside a `|` block scalar, put the marker just above the key, never inside the scalar.
+- Cite it as `path.yaml#<name>`, with the path relative to the citing file's directory or to the repo root. Never cite `path.yaml:<line>`, because line numbers go stale on any edit above them.
+- `scripts/ci/yaml-citations.sh` enforces this in `ci-manifests-test.yml`: it fails on line-number citations and on citations whose file or anchor does not exist.

README.md

@@ -0,0 +1,48 @@
+# homelab
+
+GitOps source for the Renner homelab cluster. Flux applies `master` to
+production, so every merge is a production change. Changes go in by pull
+request only; see [CONTRIBUTING.md](CONTRIBUTING.md) for YAML conventions.
+
+## Change rules that protect Home Assistant
+
+Owner decision (REN-240, REN-243). Home Assistant runs the house, so it
+gets two rules on top of normal review.
+
+### 1. Disruptive work runs only 02:00–05:00 Europe/Berlin
+
+Planned work that can restart Home Assistant or disrupt the cluster is
+merged and applied only between **02:00 and 05:00 Europe/Berlin**. That
+covers:
+
+- node drain, node reboot, Talos upgrade
+- CNI (Cilium, Multus) and cluster-wide network policy
+- storage (Rook/Ceph, CSI, snapshot controller, CNPG operator)
+- Flux core (`kubernetes/clusters/*/flux-system/`, Flux controllers)
+- Authentik
+- OpenBao, including secret rotation that restarts consumers
+- anything in `kubernetes/apps/home-assistant/`
+
+A PR in this list states in its description that it waits for the window.
+Flux applies on merge, so "wait for the window" means **merge** in the
+window, not just apply. If unsure whether a change is disruptive, treat it
+as disruptive.
+
+Not covered: Home Assistant image bumps by Flux image automation. They stay
+fully automatic by owner decision. If one breaks Home Assistant, roll it
+back with [RUNBOOK-home-assistant-image-rollback.md](kubernetes/RUNBOOK-home-assistant-image-rollback.md).
+An incident rollback is not planned work; the owner decides its timing.
+
+### 2. HA-impact review before merge
+
+Every PR that touches `kubernetes/apps/home-assistant/`, or makes a
+cluster-wide disruptive change from the list above, needs an **HA-impact
+review** before merge: a named reviewer states in the PR what Home
+Assistant will notice (restart, lost integration, lost history, none) and
+approves.
+
+- Reviewer: Hestia, the Home Assistant maintainer.
+- Until Hestia exists: the owner does the HA-impact review.
+
+This is in addition to the normal review and merge by Argus. Argus does
+not merge such a PR without the HA-impact approval.

ansible/plays/install-netbird-proxy.yml

@@ -1,5 +1,5 @@
 ---
-# Deploy a NetBird reverse-proxy HA cluster member (a second proxy host).
+# Deploy the NetBird reverse-proxy HA cluster members (the public edge).
 # Usage: ansible-playbook -i inventory/ plays/install-netbird-proxy.yml
 #
 # Prerequisites:
@@ -7,12 +7,11 @@
 #   listed under [netbird_proxy] in inventory/pve_hosts.yml. Docker is
 #   installed by the (Arch-aware) docker-install role.
 # - DNS A/AAAA for the host's own name (e.g. proxy-1.renner.uno) -> its IP.
-# - Uses the existing vault_netbird_proxy_token (shared with VPS-1) — no new
-#   token to generate; a shared token + same NB_PROXY_DOMAIN forms the cluster.
-# - AFTER deploy + verify: add this host's IP to the proxy domain so traffic
-#   round-robins across the cluster. Repoint proxy.renner.uno away from its
-#   CNAME to netbird.renner.uno (management host) to A/AAAA records listing
-#   BOTH proxy IPs (VPS-1 + this host).
+# - Uses the existing vault_netbird_proxy_token (shared by every member) — no
+#   new token to generate; a shared token + same NB_PROXY_DOMAIN forms the
+#   cluster.
+# - AFTER deploy + verify: add this host's A/AAAA to proxy.renner.uno so
+#   traffic round-robins across the cluster.
 
 - name: Deploy NetBird reverse-proxy cluster member
   hosts: netbird_proxy

ansible/plays/install-netbird-server.yml

@@ -18,4 +18,3 @@
     - common
     - docker-install
     - netbird-server
-    - netbird-proxy-watchdog

ansible/plays/rotate-netbird-secrets.yml

@@ -43,10 +43,10 @@
 #
 # WHAT IS NOT, and must not be added:
 #
-#   vault_netbird_proxy_token    LOOKS symmetric — it is in proxy.env on both
-#                                the server and the proxies, all Ansible
-#                                templates — but the management server does not
-#                                validate it from any of them. It is registered
+#   vault_netbird_proxy_token    LOOKS symmetric — it is in proxy.env on the
+#                                proxies, an Ansible template — but the
+#                                management server does not validate it from
+#                                any file. It is registered
 #                                in the management DATABASE. Rotating it changes
 #                                only what the proxies present, so every proxy
 #                                is locked out: "proxy auth failed: too many

ansible/roles/netbird-proxy-watchdog/defaults/main.yml

@@ -1,5 +1,5 @@
 ---
-# Self-healing watchdog for the NetBird reverse proxy (VPS-1 and proxy-1).
+# Self-healing watchdog for the NetBird reverse proxy on the proxy hosts.
 #
 # Why it exists: the proxy's embedded NetBird client pins each proxied
 # domain's route to a specific routing-peer identity. The netbird-ingress
@@ -20,9 +20,7 @@ netbird_watchdog_probe_hosts:
   - home.renner.uno
   - immich.renner.uno
 
-# Address the probe connects to. 127.0.0.1 reaches the local proxy on both
-# members: VPS-1 routes it via Traefik's HostSNI(*) TCP passthrough, proxy-1
-# binds :443 directly.
+# Address the probe connects to. The proxy binds :443 on the host directly.
 netbird_watchdog_probe_address: 127.0.0.1
 
 netbird_watchdog_container: netbird-proxy

ansible/roles/netbird-proxy-watchdog/tasks/main.yml

@@ -1,7 +1,7 @@
 ---
 # Deploys the reverse-proxy watchdog (see defaults/main.yml for why).
 
-- name: Ensure curl is installed (probe dependency; proxy-1 is Arch, VPS-1 Debian)
+- name: Ensure curl is installed (probe dependency)
   ansible.builtin.package:
     name: curl
     state: present

ansible/roles/netbird-proxy/defaults/main.yml

@@ -1,8 +1,8 @@
 ---
 # NetBird reverse-proxy HA cluster member (proxy-only host, e.g. proxy-1).
-# Joins the same proxy cluster as the VPS-1 proxy: identical NB_PROXY_DOMAIN
-# + the same token. NetBird load-balances/fails the cluster over via DNS
-# round-robin on the proxy domain (see plays/install-netbird-proxy.yml).
+# Every member uses the same NB_PROXY_DOMAIN + token. NetBird load-balances/
+# fails the cluster over via DNS round-robin on the proxy domain (see
+# plays/install-netbird-proxy.yml).
 
 # MUST match the netbird-server role's netbird_version — the reverse-proxy
 # release line has to track the management server it enrolls against.
@@ -11,11 +11,13 @@ netbird_version: "0.73.2"
 # Management server this proxy enrolls against (public API, TLS).
 netbird_server_host: netbird.renner.uno
 
-# Proxy domain — MUST be identical across every cluster member.
+# Proxy domain — MUST be identical across every cluster member, and must NOT
+# be a subdomain of the management domain (netbird.renner.uno) per NetBird
+# docs (TLS/routing conflicts).
 netbird_proxy_domain: "proxy.renner.uno"
 
-# Reuse VPS-1's proxy token — a shared NB_PROXY_TOKEN across members with the
-# same NB_PROXY_DOMAIN is a valid HA cluster, so no separate token is needed.
+# A shared NB_PROXY_TOKEN across members with the same NB_PROXY_DOMAIN is a
+# valid HA cluster, so one token serves every member.
 netbird_proxy_token: "{{ vault_netbird_proxy_token }}"
 
 # Relay auth secret — MUST be identical across every relay instance and the

ansible/roles/netbird-proxy/templates/docker-compose.yml.j2

@@ -1,8 +1,7 @@
 # NetBird reverse-proxy — HA cluster member on a proxy-only host.
-# Just the reverse-proxy container, bound directly to :443. No Traefik:
-# unlike VPS-1 (where the proxy shares :443 with management/signal/relay via
-# SNI routing), this host runs nothing else, so the proxy owns the port and
-# terminates TLS + answers its own tls-alpn-01 ACME challenges directly.
+# Just the reverse-proxy container, bound directly to :443. No Traefik: the
+# proxy owns the port and terminates TLS + answers its own tls-alpn-01 ACME
+# challenges directly.
 services:
   proxy:
     image: netbirdio/reverse-proxy:{{ netbird_version }}

ansible/roles/netbird-server/defaults/main.yml

@@ -23,9 +23,3 @@ netbird_datastore_key: "{{ vault_netbird_datastore_key }}"
 
 # Relay secret for WebSocket relay authentication
 netbird_relay_secret: "{{ vault_netbird_relay_secret }}"
-
-# Reverse proxy configuration
-# IMPORTANT: Proxy domain must NOT be a subdomain of the management domain
-# (netbird.renner.uno) per NetBird docs - causes TLS/routing conflicts
-netbird_proxy_domain: "proxy.renner.uno"
-netbird_proxy_token: "{{ vault_netbird_proxy_token }}"

ansible/roles/netbird-server/handlers/main.yml

@@ -16,7 +16,3 @@
 - name: Restart netbird coturn
   ansible.builtin.shell: |
     cd /opt/netbird && docker compose up -d --force-recreate coturn
-
-- name: Restart netbird proxy
-  ansible.builtin.shell: |
-    cd /opt/netbird && docker compose up -d --force-recreate proxy

ansible/roles/netbird-server/tasks/main.yml

@@ -36,7 +36,6 @@
     - /opt/netbird/management
     - /opt/netbird/signal
     - /opt/netbird/letsencrypt
-    - /opt/netbird/proxy/certs
 
 - name: Deploy docker-compose.yml
   ansible.builtin.template:
@@ -59,23 +58,83 @@
     mode: "0644"
   notify: Restart netbird coturn
 
-- name: Deploy proxy.env
-  ansible.builtin.template:
-    src: proxy.env.j2
-    dest: /opt/netbird/proxy.env
-    mode: "0644"
-  notify: Restart netbird proxy
+# The public edge is the proxy hosts (plays/install-netbird-proxy.yml); this
+# host runs no reverse proxy. The tasks below remove the one it used to run,
+# its watchdogs first so they cannot restart it mid-removal, and can go once
+# the role has run against every host that ever carried it.
+- name: Check for reverse-proxy watchdog units
+  ansible.builtin.stat:
+    path: "/etc/systemd/system/{{ item }}"
+  loop:
+    - netbird-proxy-watchdog.timer
+    - netbird-proxy-flood-watchdog.service
+  register: _proxy_watchdog_units
+
+- name: Stop and disable reverse-proxy watchdogs
+  ansible.builtin.systemd:
+    name: "{{ item.item }}"
+    state: stopped
+    enabled: false
+  loop: "{{ _proxy_watchdog_units.results }}"
+  loop_control:
+    label: "{{ item.item }}"
+  when: item.stat.exists
+
+- name: Remove reverse-proxy watchdog units and scripts
+  ansible.builtin.file:
+    path: "{{ item }}"
+    state: absent
+  loop:
+    - /etc/systemd/system/netbird-proxy-watchdog.timer
+    - /etc/systemd/system/netbird-proxy-watchdog.service
+    - /etc/systemd/system/netbird-proxy-flood-watchdog.service
+    - /usr/local/bin/netbird-proxy-watchdog
+    - /usr/local/bin/netbird-proxy-flood-watchdog
+    - /run/netbird-proxy-watchdog
+  register: _proxy_watchdog_files
+
+- name: Reload systemd after removing the watchdog units
+  ansible.builtin.systemd:
+    daemon_reload: true
+  when: _proxy_watchdog_files.changed
 
 - name: Pull NetBird container images
   ansible.builtin.shell: |
     cd /opt/netbird && docker compose pull
   changed_when: false
 
+# --remove-orphans removes containers of services no longer in the compose
+# file (the reverse proxy).
 - name: Start NetBird containers
   ansible.builtin.shell: |
-    cd /opt/netbird && docker compose up -d
+    cd /opt/netbird && docker compose up -d --remove-orphans
   register: docker_compose_result
-  changed_when: "'Started' in docker_compose_result.stdout or 'Creating' in docker_compose_result.stdout"
+  changed_when: >-
+    'Started' in docker_compose_result.stdout or
+    'Creating' in docker_compose_result.stdout or
+    'Removed' in docker_compose_result.stdout or
+    'Removed' in docker_compose_result.stderr
+
+- name: Remove reverse-proxy volumes
+  ansible.builtin.shell: |
+    set -eu
+    for v in netbird_netbird_proxy_certs netbird_netbird_proxy_data netbird_netbird_proxy_state; do
+      if [ -n "$(docker volume ls -q --filter "name=^${v}$")" ]; then
+        docker volume rm "$v"
+        echo "removed $v"
+      fi
+    done
+  register: _proxy_volumes
+  changed_when: "'removed' in _proxy_volumes.stdout"
+
+- name: Remove reverse-proxy config and state
+  ansible.builtin.file:
+    path: "{{ item }}"
+    state: absent
+  loop:
+    - /opt/netbird/proxy.env
+    - /opt/netbird/proxy
+    - /opt/netbird/proxy-state
 
 # The management store has no cluster-side backup. Pulling one put a root
 # key on this host that the cluster held, and this server's own state is

ansible/roles/netbird-server/templates/docker-compose.yml.j2

@@ -187,10 +187,9 @@ services:
       - "traefik.http.routers.netbird-management.service=netbird-management"
       - "traefik.http.services.netbird-management.loadbalancer.server.port=33073"
       - "traefik.http.services.netbird-management.loadbalancer.server.scheme=h2c"
-      # gRPC reverse-proxy service — remote reverse-proxy instances (e.g.
-      # proxy-1) connect here via NB_PROXY_MANAGEMENT_ADDRESS=https://host:443.
-      # The local proxy uses management:33073 directly, so it never needed
-      # this route; remote ones get a 404 without it. Same backend (h2c 33073).
+      # gRPC reverse-proxy service — the proxy hosts (proxy-1, proxy-2)
+      # connect here via NB_PROXY_MANAGEMENT_ADDRESS=https://host:443 and get
+      # a 404 without it. Same backend (h2c 33073).
       - "traefik.http.routers.netbird-proxyservice.rule=Host(`{{ netbird_server_host }}`) && PathPrefix(`/management.ProxyService/`)"
       - "traefik.http.routers.netbird-proxyservice.entrypoints=websecure"
       - "traefik.http.routers.netbird-proxyservice.tls=true"
@@ -221,37 +220,6 @@ services:
         max-size: "500m"
         max-file: "2"
 
-  # Reverse Proxy Service (per docs.netbird.io/selfhosted/migration/enable-reverse-proxy)
-  proxy:
-    image: netbirdio/reverse-proxy:{{ netbird_version }}
-    container_name: netbird-proxy
-    restart: unless-stopped
-    networks:
-      - netbird
-    depends_on:
-      - management
-    env_file:
-      - ./proxy.env
-    volumes:
-      - netbird_proxy_certs:/certs
-      - /opt/netbird/proxy-state:/var/lib/netbird
-    labels:
-      - "traefik.enable=true"
-      - "traefik.tcp.routers.proxy-passthrough.entrypoints=websecure"
-      - "traefik.tcp.routers.proxy-passthrough.rule=HostSNI(`*`)"
-      - "traefik.tcp.routers.proxy-passthrough.tls.passthrough=true"
-      - "traefik.tcp.routers.proxy-passthrough.service=proxy-tls"
-      - "traefik.tcp.routers.proxy-passthrough.priority=1"
-      - "traefik.tcp.services.proxy-tls.loadbalancer.server.port=8443"
-    logging:
-      driver: "json-file"
-      options:
-        max-size: "500m"
-        max-file: "2"
-
 networks:
   netbird:
     driver: bridge
-
-volumes:
-  netbird_proxy_certs:

ansible/roles/netbird-server/templates/proxy.env.j2

@@ -1,8 +0,0 @@
-NB_PROXY_DOMAIN={{ netbird_proxy_domain }}
-NB_PROXY_TOKEN={{ netbird_proxy_token }}
-NB_PROXY_MANAGEMENT_ADDRESS=http://management:33073
-NB_PROXY_ALLOW_INSECURE=true
-NB_PROXY_ADDRESS=:8443
-NB_PROXY_ACME_CERTIFICATES=true
-NB_PROXY_ACME_CHALLENGE_TYPE=tls-alpn-01
-NB_PROXY_CERTIFICATE_DIRECTORY=/certs

ci/agents/wiki-maintenance/consolidate-settings.json

@@ -0,0 +1,120 @@
+{
+  "//": "Permission scope for the wiki-consolidate agent in CI: read the vault and the",
+  "//2": "source repos cloned under /root/source, read the live cluster through the",
+  "//3": "ci-diagnostics kubeconfig, edit wiki/, output/ and the rotation state only.",
+  "//4": "Deny rules win; git and kubectl are read-only subcommands, no network tools.",
+  "permissions": {
+    "allow": [
+      "Read(//root/Notes/**)",
+      "Read(//root/source/**)",
+      "Edit(//root/Notes/wiki/**)",
+      "Edit(//root/Notes/output/**)",
+      "Edit(//root/Notes/.wiki-consolidation-state.json)",
+      "Bash(/root/Notes/bin/wiki-snapshot.sh:*)",
+      "Bash(/root/Notes/bin/wiki_rank.py:*)",
+      "Bash(/root/Notes/bin/wiki_ask.py:*)",
+      "Bash(~/Notes/bin/wiki-snapshot.sh:*)",
+      "Bash(~/Notes/bin/wiki_rank.py:*)",
+      "Bash(~/Notes/bin/wiki_ask.py:*)",
+      "Bash(~/Notes/bin/wiki-delete.sh:*)",
+      "Bash(git -C /root/Notes log:*)",
+      "Bash(git -C /root/Notes show:*)",
+      "Bash(git -C /root/Notes diff:*)",
+      "Bash(git -C /root/Notes status:*)",
+      "Bash(git -C /root/Notes rev-parse:*)",
+      "Bash(git -C /root/Notes ls-files:*)",
+      "Bash(git -C ~/Notes log:*)",
+      "Bash(git -C ~/Notes show:*)",
+      "Bash(git -C ~/Notes diff:*)",
+      "Bash(git -C ~/Notes status:*)",
+      "Bash(git -C ~/Notes rev-parse:*)",
+      "Bash(git -C ~/Notes ls-files:*)",
+      "Bash(kubectl get:*)",
+      "Bash(kubectl describe:*)",
+      "Bash(kubectl version:*)",
+      "Bash(kubectl api-resources:*)",
+      "Bash(kubectl explain:*)",
+      "Bash(git log:*)",
+      "Bash(git show:*)",
+      "Bash(git status:*)",
+      "Bash(git diff:*)",
+      "Bash(git rev-parse:*)",
+      "Bash(git ls-files:*)",
+      "Bash(git -C /root/source/homelab log:*)",
+      "Bash(git -C /root/source/homelab show:*)",
+      "Bash(git -C /root/source/homelab diff:*)",
+      "Bash(git -C /root/source/homelab status:*)",
+      "Bash(git -C /root/source/homelab rev-parse:*)",
+      "Bash(git -C /root/source/homelab ls-files:*)",
+      "Bash(git -C /root/source/homeassistant log:*)",
+      "Bash(git -C /root/source/homeassistant show:*)",
+      "Bash(git -C /root/source/homeassistant diff:*)",
+      "Bash(git -C /root/source/homeassistant status:*)",
+      "Bash(git -C /root/source/homeassistant rev-parse:*)",
+      "Bash(git -C /root/source/homeassistant ls-files:*)",
+      "Bash(git -C ~/source/homelab log:*)",
+      "Bash(git -C ~/source/homelab show:*)",
+      "Bash(git -C ~/source/homelab diff:*)",
+      "Bash(git -C ~/source/homelab status:*)",
+      "Bash(git -C ~/source/homelab rev-parse:*)",
+      "Bash(git -C ~/source/homelab ls-files:*)",
+      "Bash(git -C ~/source/homeassistant log:*)",
+      "Bash(git -C ~/source/homeassistant show:*)",
+      "Bash(git -C ~/source/homeassistant diff:*)",
+      "Bash(git -C ~/source/homeassistant status:*)",
+      "Bash(git -C ~/source/homeassistant rev-parse:*)",
+      "Bash(git -C ~/source/homeassistant ls-files:*)",
+      "Bash(/root/Notes/bin/wiki-delete.sh:*)",
+      "Bash(ls:*)",
+      "Bash(date:*)",
+      "Bash(wc:*)",
+      "Bash(stat:*)",
+      "Bash(head:*)",
+      "Bash(tail:*)"
+    ],
+    "deny": [
+      "Read(//proc/**)",
+      "Read(//etc/**)",
+      "Read(//root/.config/**)",
+      "Read(//root/.kube/**)",
+      "Read(//woodpecker/**/.ci-diag/**)",
+      "Read(//root/Notes/.git/**)",
+      "Read(//root/source/*/.git/**)",
+      "Read(//root/Notes/_archive/**)",
+      "Edit(//root/Notes/raw/**)",
+      "Edit(//root/Notes/_archive/**)",
+      "Edit(//root/source/**)",
+      "Bash(kubectl apply:*)",
+      "Bash(kubectl delete:*)",
+      "Bash(kubectl patch:*)",
+      "Bash(kubectl edit:*)",
+      "Bash(kubectl scale:*)",
+      "Bash(kubectl rollout:*)",
+      "Bash(kubectl exec:*)",
+      "Bash(kubectl cp:*)",
+      "Bash(kubectl drain:*)",
+      "Bash(kubectl cordon:*)",
+      "Bash(kubectl label:*)",
+      "Bash(kubectl annotate:*)",
+      "Bash(git push:*)",
+      "Bash(git reset:*)",
+      "Bash(git checkout:*)",
+      "Bash(git clean:*)",
+      "Bash(git rebase:*)",
+      "Bash(git config:*)",
+      "Bash(git -C /root/Notes config:*)",
+      "Bash(git -C ~/Notes config:*)",
+      "Bash(git -C /root/source/homelab config:*)",
+      "Bash(git -C /root/source/homeassistant config:*)",
+      "Bash(git -C ~/source/homelab config:*)",
+      "Bash(git -C ~/source/homeassistant config:*)",
+      "Bash(rm:*)",
+      "Bash(rm -rf:*)",
+      "Bash(curl:*)",
+      "Bash(wget:*)",
+      "Bash(ssh:*)",
+      "Bash(tea:*)",
+      "Bash(woodpecker:*)"
+    ]
+  }
+}

ci/agents/wiki-maintenance/consolidate.md

@@ -0,0 +1,5 @@
+You are running as the scheduled wiki-maintenance job in CI. The Notes vault is checked out at ~/Notes (the repository k.renner/notes); the homelab and homeassistant repositories are checked out under ~/source/. `kubectl` reads the live cluster with a read-only identity.
+
+Run the wiki-consolidate agent now for today's scheduled consolidation pass. Follow its instructions exactly: snapshot the vault to git first, abort if that fails, select today's rotating cohort of 20 from the state file, verify claims against both their sources and the live systems, act on the verdicts, write the dated report to ~/Notes/output/, update the state file, and snapshot again at the end. Do not push; the pipeline pushes the run's commits.
+
+Report the counts and the coverage progress when you are done.

ci/agents/wiki-maintenance/distill.md

@@ -0,0 +1,7 @@
+You are running as the scheduled wiki-maintenance job in CI. The Notes vault is checked out at ~/Notes (the repository k.renner/notes). Do exactly this, nothing else:
+
+Glob ~/Notes/raw/*.md (top level only, never ~/Notes/raw/_processed/). Distill every file you find into ~/Notes/wiki/ following your process, including the ranker check for every concept. Move each fully processed source to ~/Notes/raw/_processed/. If raw/ has no top-level files besides _index.md, say so and stop.
+
+Do not commit or push; the pipeline does that after you finish. Do not touch anything outside ~/Notes/wiki/ and ~/Notes/raw/.
+
+End with one line per file: which wiki notes it produced or merged into, or "nothing distillable".

ci/agents/wiki-maintenance/settings.json

@@ -0,0 +1,41 @@
+{
+  "//": "Permission scope for the wiki-maintenance agent (wiki-distill) in CI.",
+  "//2": "The step's environment holds a Gitea push token, a TypeSafe key and the wiki-index",
+  "//3": "client credential, so the Read tool is kept away from /proc (env dumps), /root outside",
+  "//4": "the vault symlink target, /etc and the checkout's .git/. The vault itself is the work.",
+  "//5": "Mirrors ci/agents/issue-fix/settings.json; absolute globs because the vault is reached",
+  "//6": "through the ~/Notes symlink the pipeline creates.",
+  "permissions": {
+    "allow": [
+      "Read(//root/Notes/**)",
+      "Edit(//root/Notes/wiki/**)",
+      "Edit(//root/Notes/output/**)",
+      "Edit(//root/Notes/raw/**)",
+      "Bash(/root/Notes/bin/wiki_rank.py:*)",
+      "Bash(/root/Notes/bin/wiki_ask.py:*)",
+      "Bash(mv /root/Notes/raw/:*)",
+      "Bash(mkdir -p /root/Notes/:*)",
+      "Bash(~/Notes/bin/wiki_rank.py:*)",
+      "Bash(~/Notes/bin/wiki_ask.py:*)",
+      "Bash(mv ~/Notes/raw/:*)",
+      "Bash(mkdir -p ~/Notes/:*)",
+      "Bash(ls:*)",
+      "Bash(date:*)",
+      "Bash(wc:*)",
+      "Bash(head:*)",
+      "Bash(tail:*)"
+    ],
+    "deny": [
+      "Read(//proc/**)",
+      "Read(//etc/**)",
+      "Read(//root/.config/**)",
+      "Read(//root/Notes/.git/**)",
+      "Read(//root/Notes/_archive/**)",
+      "Edit(//root/Notes/_archive/**)",
+      "Bash(rm:*)",
+      "Bash(git:*)",
+      "Bash(curl:*)",
+      "Bash(wget:*)"
+    ]
+  }
+}

ci/images/claude/Containerfile

@@ -13,7 +13,9 @@ RUN apk add --no-cache \
       git \
       jq \
       openssh-client \
-      openssl
+      openssl \
+      python3 \
+      py3-yaml
 
 # Install Claude Code, then make the resulting layer byte-stable so that a
 # rebuild of the SAME version DEDUPS in the Gitea registry instead of orphaning

ci/images/k8s/Containerfile

@@ -1,6 +1,6 @@
 FROM alpine:3.21@sha256:48b0309ca019d89d40f670aa1bc06e426dc0931948452e8491e3d65087abc07d
 
-ARG TALOS_VERSION=v1.13.9
+ARG TALOS_VERSION=v1.13.10
 ARG VELERO_VERSION=v1.16.0
 ARG KUBECTL_VERSION=v1.36.3
 ARG FLUX_VERSION=2.9.5
@@ -18,7 +18,7 @@ RUN apk add --no-cache \
 
 RUN curl -sfL -o /usr/local/bin/talosctl \
       "https://github.com/siderolabs/talos/releases/download/${TALOS_VERSION}/talosctl-linux-amd64" \
- && echo "7e1d4b7d5846964bdcf63a794e3c8161bb6ef2983d5ace58ea5322f3bf32a27e  /usr/local/bin/talosctl" | sha256sum -c - \
+ && echo "28e9ccd9aebd690fa03d8cf2582221d87ea38b2a1704f411d75a03589c58f2aa  /usr/local/bin/talosctl" | sha256sum -c - \
  && chmod +x /usr/local/bin/talosctl
 
 RUN curl -sfL -o /usr/local/bin/kubectl \

ci/images/netbird-service-sync/Containerfile

@@ -1,4 +1,4 @@
-# netbird-proxy-services: reconciles NetBird reverse-proxy services (public
+# netbird-service-sync: reconciles NetBird reverse-proxy services (public
 # domains -> TLS passthrough -> Traefik ClusterIP) from Ingresses carrying the
 # public label. Standard library only, so the image is the interpreter plus one
 # file; runs unprivileged as its own user.

ci/images/netbird-service-sync/controller.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""netbird-proxy-services: reconcile NetBird reverse-proxy services from Ingresses.
+"""netbird-service-sync: reconcile NetBird reverse-proxy services from Ingresses.
 
 Source of truth: every Ingress carrying the public label (dns.renner.uno/public).
 For each host on such an Ingress the NetBird management gets one reverse-proxy
@@ -30,11 +30,11 @@ import urllib.error
 import urllib.parse
 import urllib.request
 
-LOG = logging.getLogger("netbird-proxy-services")
+LOG = logging.getLogger("netbird-service-sync")
 
 NB = os.environ.get("NB_MANAGEMENT_URL", "https://netbird.renner.uno").rstrip("/")
 BAO_ADDR = os.environ.get("BAO_ADDR", "http://openbao.openbao.svc.cluster.local:8200").rstrip("/")
-BAO_ROLE = os.environ.get("BAO_ROLE", "netbird-proxy-services")
+BAO_ROLE = os.environ.get("BAO_ROLE", "netbird-service-sync")
 BAO_PAT_PATH = os.environ.get("BAO_PAT_PATH", "homelab/data/infrastructure/secret-rotators-bootstrap/netbird")
 BAO_PAT_KEY = os.environ.get("BAO_PAT_KEY", "pat")
 SERVICE_CIDR = os.environ.get("SERVICE_CIDR", "10.96.0.0/12")
@@ -156,7 +156,7 @@ def ensure_network(nb):
         if n["name"] == NETWORK_NAME:
             return n["id"]
     n = nb.post("networks", {"name": NETWORK_NAME,
-                             "description": "managed by netbird-proxy-services; public ingress via TLS passthrough"})
+                             "description": "managed by netbird-service-sync; public ingress via TLS passthrough"})
     LOG.info("created network %s (%s)", NETWORK_NAME, n["id"])
     return n["id"]
 
@@ -166,7 +166,7 @@ def ensure_resource(nb, net):
         if r.get("address") == SERVICE_CIDR:
             return r["id"]
     r = nb.post(f"networks/{net}/resources", {"name": "cluster service CIDR", "address": SERVICE_CIDR,
-                                              "description": "managed by netbird-proxy-services",
+                                              "description": "managed by netbird-service-sync",
                                               "enabled": True, "groups": []})
     LOG.info("created subnet resource %s (%s)", SERVICE_CIDR, r["id"])
     return r["id"]

ci/images/vmtest/Containerfile

@@ -14,7 +14,7 @@ FROM alpine:3.21@sha256:48b0309ca019d89d40f670aa1bc06e426dc0931948452e8491e3d650
 # kubernetes/talos/patches/cluster.yaml. The phase reads the version from that
 # file at run time and passes it to the provisioner; this pin only decides which
 # client talks to it.
-ARG TALOS_VERSION=v1.13.9
+ARG TALOS_VERSION=v1.13.10
 ARG KUBECTL_VERSION=v1.36.2
 ARG FLUX_VERSION=2.9.2
 ARG KOPIA_VERSION=0.23.1
@@ -49,7 +49,7 @@ RUN apk add --no-cache \
 
 RUN curl -sfL -o /usr/local/bin/talosctl \
       "https://github.com/siderolabs/talos/releases/download/${TALOS_VERSION}/talosctl-linux-amd64" \
- && echo "7e1d4b7d5846964bdcf63a794e3c8161bb6ef2983d5ace58ea5322f3bf32a27e  /usr/local/bin/talosctl" | sha256sum -c - \
+ && echo "28e9ccd9aebd690fa03d8cf2582221d87ea38b2a1704f411d75a03589c58f2aa  /usr/local/bin/talosctl" | sha256sum -c - \
  && chmod +x /usr/local/bin/talosctl
 
 RUN curl -sfL -o /usr/local/bin/kubectl \

ci/images/wiki-index/Containerfile

@@ -1,10 +0,0 @@
-# wiki-index: embedding index over the Notes repo's wiki/, fed from Gitea,
-# queried by ~/Notes/bin/wiki_rank.py. Standard library only; runs unprivileged.
-FROM python:3.13-slim
-
-RUN useradd --system --no-create-home --shell /usr/sbin/nologin wikiindex
-COPY wiki_index.py /opt/wiki_index.py
-
-ENV PYTHONUNBUFFERED=1
-USER wikiindex
-ENTRYPOINT ["python3", "/opt/wiki_index.py"]

ci/images/wiki-index/test_wiki_index.py

@@ -1,178 +0,0 @@
-#!/usr/bin/env python3
-"""Unit tests for wiki_index. Run: python3 -m unittest test_wiki_index"""
-
-import hashlib
-import hmac
-import json
-import tempfile
-import unittest
-from pathlib import Path
-
-import wiki_index as wi
-
-NOTE = """---
-title: 'UDM intercepts DNS'
-type: wiki
-tags: [dns, unifi]
-aliases: [udm dns interception, dig 1.1.1.1 wrong]
----
-
-# UDM intercepts DNS
-
-The UDM answers every UDP/53 query itself.
-
-## Details
-
-More.
-"""
-
-
-class ParseTests(unittest.TestCase):
-    def test_note_text(self):
-        title, text = wi.note_text("udm-dns", NOTE)
-        self.assertEqual(title, "UDM intercepts DNS")
-        self.assertIn("udm dns UDM intercepts DNS udm dns interception", text)
-        self.assertIn("dns unifi", text)
-        self.assertTrue(text.endswith("The UDM answers every UDP/53 query itself."))
-        self.assertLessEqual(len(text), wi.EMBED_CHARS)
-
-    def test_credentials_are_stripped(self):
-        original = wi.http
-        wi.http = lambda method, url, body=None, headers=None, timeout=60.0: (
-            {"auth": {"client_token": "t"}} if url.endswith("/login")
-            else {"data": {"data": {"gitea_token": "abc\n", "webhook_secret": " s3cret\n"}}})
-        original_sa = wi.SA_TOKEN
-        try:
-            with tempfile.TemporaryDirectory() as d:
-                wi.SA_TOKEN = Path(d) / "token"
-                wi.SA_TOKEN.write_text("jwt\n")
-                self.assertEqual(wi.Credentials().fetch(), {"gitea_token": "abc", "webhook_secret": "s3cret"})
-        finally:
-            wi.http, wi.SA_TOKEN = original, original_sa
-
-    def test_tree_paging_stops_on_short_last_page(self):
-        # Gitea: truncated=true on every page including the last, "tree": null past it.
-        pages = {
-            1: {"tree": [{"path": f"wiki/n{i}.md", "type": "blob", "sha": f"s{i}"} for i in range(1000)], "truncated": True, "total_count": 1192},
-            2: {"tree": [{"path": f"wiki/n{i}.md", "type": "blob", "sha": f"s{i}"} for i in range(1000, 1192)] + [{"path": "raw/x.md", "type": "blob", "sha": "r"}, {"path": "wiki/sub/y.md", "type": "blob", "sha": "y"}], "truncated": True, "total_count": 1192},
-            3: {"tree": None, "truncated": False, "total_count": 1192},
-        }
-        calls = []
-        original = wi.http
-
-        def fake_http(method, url, body=None, headers=None, timeout=60.0):
-            page = int(url.rsplit("page=", 1)[1])
-            calls.append(page)
-            return pages[page]
-
-        wi.http = fake_http
-        try:
-            tree = wi.Gitea("t").tree("abc")
-        finally:
-            wi.http = original
-        self.assertEqual(calls, [1, 2])
-        self.assertEqual(len(tree), 1192)
-        self.assertNotIn("raw/x.md", tree)
-        self.assertNotIn("wiki/sub/y.md", tree)
-        self.assertEqual(tree["wiki/n1191.md"], "s1191")
-
-    def test_signature(self):
-        body = b'{"ref":"refs/heads/master"}'
-        sig = hmac.new(b"s3cret", body, hashlib.sha256).hexdigest()
-        self.assertTrue(wi.verify_signature("s3cret", body, sig))
-        self.assertFalse(wi.verify_signature("other", body, sig))
-        self.assertFalse(wi.verify_signature("s3cret", body, ""))
-
-
-class FakeGitea:
-    def __init__(self, head, tree, files):
-        self._head, self._tree, self.files = head, tree, files
-        self.raw_calls = []
-
-    def head(self):
-        return self._head
-
-    def tree(self, commit):
-        return self._tree
-
-    def raw(self, path, ref):
-        self.raw_calls.append(path)
-        return self.files[path]
-
-
-class IndexTests(unittest.TestCase):
-    def setUp(self):
-        self.tmp = tempfile.TemporaryDirectory()
-        self.path = Path(self.tmp.name) / "index.json"
-        self.original_embed = wi.embed
-        self.embed_calls = []
-
-        def fake_embed(texts, batch=2):
-            self.embed_calls.append(list(texts))
-            return [wi.unit([1.0, float(len(t) % 7)]) for t in texts]
-
-        wi.embed = fake_embed
-
-    def tearDown(self):
-        wi.embed = self.original_embed
-        self.tmp.cleanup()
-
-    def test_sync_is_incremental_and_persists(self):
-        files = {"wiki/udm-dns.md": NOTE, "wiki/ceph.md": "---\ntitle: Ceph\n---\n# Ceph\n\nOSD.\n", "wiki/_index.md": "---\ntitle: Wiki\n---\n"}
-        tree = {"wiki/udm-dns.md": "aaa", "wiki/ceph.md": "bbb", "wiki/_index.md": "ccc"}
-        g = FakeGitea("c1", tree, files)
-        idx = wi.Index(self.path)
-        s = idx.sync(g)
-        self.assertEqual((s["changed"], s["removed"], s["notes"]), (2, 0, 2))
-        self.assertEqual(set(idx.notes), {"udm-dns", "ceph"})
-        self.assertEqual(idx.notes["udm-dns"]["title"], "UDM intercepts DNS")
-        self.assertTrue(self.path.exists())
-
-        # head unchanged: nothing fetched
-        s = idx.sync(g)
-        self.assertEqual(s.get("skipped"), "head unchanged")
-        self.assertEqual(len(g.raw_calls), 2)
-
-        # one note changed, one removed, new head
-        g2 = FakeGitea("c2", {"wiki/udm-dns.md": "aaa2"}, {"wiki/udm-dns.md": NOTE.replace("itself", "itself!")})
-        s = idx.sync(g2)
-        self.assertEqual((s["changed"], s["removed"], s["notes"]), (1, 1, 1))
-        self.assertEqual(g2.raw_calls, ["wiki/udm-dns.md"])
-
-        # restart loads the persisted index
-        idx2 = wi.Index(self.path)
-        self.assertEqual(idx2.commit, "c2")
-        self.assertEqual(set(idx2.notes), {"udm-dns"})
-
-    def test_rank_orders_by_cosine(self):
-        idx = wi.Index(self.path)
-        idx.notes = {"a": {"sha": "1", "title": "A", "vector": wi.unit([1.0, 0.0])},
-                     "b": {"sha": "2", "title": "B", "vector": wi.unit([0.0, 1.0])}}
-        wi.embed = lambda texts, batch=2: [wi.unit([0.2, 1.0])]
-        rows = idx.rank("q", k=5)
-        self.assertEqual([r["slug"] for r in rows], ["b", "a"])
-        self.assertGreater(rows[0]["score"], rows[1]["score"])
-
-    def test_status_reports_age_and_errors(self):
-        idx = wi.Index(self.path)
-        self.assertFalse(idx.status()["ready"])
-        g = FakeGitea("c1", {"wiki/a.md": "x"}, {"wiki/a.md": "---\ntitle: A\n---\n# A\n\nt\n"})
-        idx.sync(g)
-        st = idx.status()
-        self.assertTrue(st["ready"])
-        self.assertEqual(st["notes"], 1)
-        self.assertEqual(st["commit"], "c1")
-
-        class Broken(FakeGitea):
-            def head(self):
-                raise wi.ApiError("gitea down")
-
-        with self.assertRaises(wi.ApiError):
-            idx.sync(Broken("c1", {}, {}))
-        self.assertIn("gitea down", idx.status()["last_error"])
-        self.assertEqual(idx.status()["notes"], 1)  # last good index kept
-        self.assertTrue(idx.status()["ready"])  # and still served
-
-
-if __name__ == "__main__":
-    unittest.main()

ci/images/wiki-index/wiki_index.py

@@ -1,408 +0,0 @@
-#!/usr/bin/env python3
-"""wiki-index: embedding index over the Notes repo's wiki/, fed from Gitea.
-
-Source of truth is the notes repository on Gitea. The index holds, per wiki
-note, the git blob sha, title and a bge-m3 vector of slug/title/aliases/tags/
-first paragraph. A sync fetches the repo tree, embeds every note whose blob sha
-changed (via wiki-embed, in-cluster) and drops notes that left the tree; it
-runs at start, on every push webhook from Gitea, and every RESYNC_SECONDS.
-
-HTTP:
-  POST /rank     {"query": str, "k": int}   -> {"rows": [{"slug","title","score"}], "index": {...}}
-  GET  /health   -> 200 always (liveness) with {"ready", "notes", "commit", "synced_at", "age_s", "syncing", "last_error"}
-  GET  /ready    -> 200 when an index exists, else 503 (readiness)
-  POST /webhook  Gitea push event, HMAC-SHA256 in X-Gitea-Signature -> {"queued": true}
-
-Credentials (Gitea token, webhook secret) are read from OpenBao through the
-Kubernetes auth role on every use; they never land in a Secret or a log line.
-Standard library only.
-"""
-
-from __future__ import annotations
-
-import hashlib
-import hmac
-import json
-import logging
-import os
-import re
-import sys
-import threading
-import time
-import urllib.error
-import urllib.parse
-import urllib.request
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
-from pathlib import Path
-
-LOG = logging.getLogger("wiki-index")
-
-GITEA_URL = os.environ.get("GITEA_URL", "https://gitea.renner.uno").rstrip("/")
-NOTES_REPO = os.environ.get("NOTES_REPO", "k.renner/notes")
-NOTES_BRANCH = os.environ.get("NOTES_BRANCH", "master")
-WIKI_DIR = os.environ.get("WIKI_DIR", "wiki").strip("/")
-EMBED_URL = os.environ.get("EMBED_URL", "http://wiki-embed.ai.svc.cluster.local:8080").rstrip("/")
-EMBED_MODEL = os.environ.get("EMBED_MODEL", "bge-m3")
-EMBED_CHARS = int(os.environ.get("EMBED_CHARS", "1500"))
-INDEX_PATH = Path(os.environ.get("INDEX_PATH", "/data/index.json"))
-RESYNC_SECONDS = int(os.environ.get("RESYNC_SECONDS", "1800"))
-LISTEN_PORT = int(os.environ.get("LISTEN_PORT", "8080"))
-BAO_ADDR = os.environ.get("BAO_ADDR", "http://openbao.openbao.svc.cluster.local:8200").rstrip("/")
-BAO_ROLE = os.environ.get("BAO_ROLE", "wiki-index")
-BAO_SECRET_PATH = os.environ.get("BAO_SECRET_PATH", "homelab/data/apps/ai/wiki-index")
-SA_TOKEN = Path(os.environ.get("SA_TOKEN", "/var/run/secrets/kubernetes.io/serviceaccount/token"))
-
-
-class ApiError(Exception):
-    pass
-
-
-# ---------------------------------------------------------------- http helpers
-
-
-def http(method: str, url: str, body=None, headers=None, timeout: float = 60.0):
-    data = None
-    hdrs = dict(headers or {})
-    if body is not None:
-        data = json.dumps(body).encode("utf-8")
-        hdrs.setdefault("Content-Type", "application/json")
-    req = urllib.request.Request(url, data=data, method=method, headers=hdrs)
-    try:
-        with urllib.request.urlopen(req, timeout=timeout) as resp:
-            raw = resp.read()
-            ctype = resp.headers.get("Content-Type", "")
-    except urllib.error.HTTPError as e:
-        raise ApiError(f"{method} {url}: HTTP {e.code}: {e.read().decode('utf-8', 'replace')[:300]}") from e
-    except urllib.error.URLError as e:
-        raise ApiError(f"{method} {url}: {e.reason}") from e
-    if "json" in ctype:
-        return json.loads(raw.decode("utf-8"))
-    return raw
-
-
-# ---------------------------------------------------------------- credentials
-
-
-class Credentials:
-    """Gitea token and webhook secret from OpenBao, fetched per use."""
-
-    def __init__(self, static: dict | None = None):
-        self.static = static  # tests / local runs
-
-    def fetch(self) -> dict:
-        if self.static is not None:
-            return self.static
-        jwt = SA_TOKEN.read_text(encoding="utf-8").strip()
-        login = http("POST", f"{BAO_ADDR}/v1/auth/kubernetes/login", {"role": BAO_ROLE, "jwt": jwt})
-        token = login["auth"]["client_token"]
-        kv = http("GET", f"{BAO_ADDR}/v1/{BAO_SECRET_PATH}", headers={"X-Vault-Token": token})
-        data = {k: str(v).strip() for k, v in kv["data"]["data"].items()}  # `bao kv put key=@file` keeps the file's newline
-        for key in ("gitea_token", "webhook_secret"):
-            if not data.get(key):
-                raise ApiError(f"OpenBao {BAO_SECRET_PATH} has no key {key}")
-        return data
-
-
-# ---------------------------------------------------------------- gitea
-
-
-class Gitea:
-    def __init__(self, token: str, base: str = GITEA_URL, repo: str = NOTES_REPO, branch: str = NOTES_BRANCH):
-        self.base, self.repo, self.branch = base, repo, branch
-        self.headers = {"Authorization": f"token {token}"}
-
-    def api(self, path: str, **params):
-        q = ("?" + urllib.parse.urlencode(params)) if params else ""
-        return http("GET", f"{self.base}/api/v1/repos/{self.repo}{path}{q}", headers=self.headers)
-
-    def head(self) -> str:
-        return self.api(f"/branches/{urllib.parse.quote(self.branch, safe='')}")["commit"]["id"]
-
-    def tree(self, commit: str) -> dict[str, str]:
-        """path -> blob sha for every wiki note at that commit."""
-        out: dict[str, str] = {}
-        page, per_page, seen = 1, 1000, 0
-        while True:
-            t = self.api(f"/git/trees/{commit}", recursive="true", per_page=per_page, page=page)
-            entries = t.get("tree") or []  # Gitea answers "tree": null past the last page
-            for e in entries:
-                p = e.get("path", "")
-                if e.get("type") == "blob" and p.startswith(WIKI_DIR + "/") and p.endswith(".md") and "/" not in p[len(WIKI_DIR) + 1:]:
-                    out[p] = e["sha"]
-            seen += len(entries)
-            total = int(t.get("total_count") or 0)
-            if not entries or len(entries) < per_page or (total and seen >= total):
-                break
-            page += 1
-            if page > 50:
-                raise ApiError("tree paging runaway")
-        return out
-
-    def raw(self, path: str, ref: str) -> str:
-        data = http("GET", f"{self.base}/api/v1/repos/{self.repo}/raw/{urllib.parse.quote(path)}?ref={urllib.parse.quote(ref, safe='')}",
-                    headers=self.headers)
-        return data.decode("utf-8") if isinstance(data, bytes) else json.dumps(data)
-
-
-# ---------------------------------------------------------------- notes
-
-
-def parse_frontmatter(text: str) -> tuple[dict, str]:
-    m = re.match(r"^---[ \t]*\n(.*?)\n---[ \t]*(?:\n|$)", text, re.DOTALL)
-    if not m:
-        return {}, text
-    fm: dict = {}
-    for line in m.group(1).splitlines():
-        mm = re.match(r"^([A-Za-z_][\w-]*):\s*(.*)$", line)
-        if mm:
-            fm[mm.group(1)] = mm.group(2).strip()
-    return fm, text[m.end():]
-
-
-def as_list(value: str) -> list[str]:
-    s = (value or "").strip()
-    if s.startswith("[") and s.endswith("]"):
-        s = s[1:-1]
-    return [p.strip().strip("'\"") for p in s.split(",") if p.strip()]
-
-
-def first_paragraph(body: str) -> str:
-    para: list[str] = []
-    seen_h1 = False
-    for line in body.splitlines():
-        if line.startswith("# "):
-            seen_h1 = True
-            continue
-        if not seen_h1:
-            continue
-        if line.strip() == "" or line.startswith("#"):
-            if para:
-                break
-            continue
-        para.append(line.strip())
-    return " ".join(para)
-
-
-def note_text(slug: str, text: str) -> tuple[str, str]:
-    """(title, embedding text) for one note file."""
-    fm, body = parse_frontmatter(text)
-    title = (fm.get("title") or slug).strip().strip("'\"")
-    parts = [slug.replace("-", " "), title, " ".join(as_list(fm.get("aliases", ""))),
-             " ".join(as_list(fm.get("tags", ""))), first_paragraph(body)]
-    return title, " ".join(parts)[:EMBED_CHARS]
-
-
-# ---------------------------------------------------------------- embeddings
-
-
-def embed(texts: list[str], batch: int = 2) -> list[list[float]]:
-    out: list[list[float]] = []
-    for i in range(0, len(texts), batch):
-        resp = http("POST", f"{EMBED_URL}/v1/embeddings", {"model": EMBED_MODEL, "input": texts[i:i + batch]}, timeout=120)
-        rows = sorted(resp["data"], key=lambda d: d["index"])
-        out.extend(unit(d["embedding"]) for d in rows)
-    return out
-
-
-def unit(v: list[float]) -> list[float]:
-    n = sum(x * x for x in v) ** 0.5 or 1.0
-    return [round(x / n, 5) for x in v]
-
-
-# ---------------------------------------------------------------- index
-
-
-class Index:
-    def __init__(self, path: Path):
-        self.path = path
-        self.lock = threading.Lock()
-        self.notes: dict[str, dict] = {}  # slug -> {sha, title, vector}
-        self.commit = ""
-        self.synced_at = 0.0
-        self.syncing = False
-        self.last_error = ""
-        if path.exists():
-            try:
-                d = json.loads(path.read_text(encoding="utf-8"))
-                self.notes, self.commit, self.synced_at = d.get("notes", {}), d.get("commit", ""), d.get("synced_at", 0.0)
-                LOG.info("loaded index: %d notes at %s", len(self.notes), self.commit[:8])
-            except (OSError, ValueError) as e:
-                LOG.warning("index unreadable (%s); starting empty", e)
-
-    def save(self) -> None:
-        self.path.parent.mkdir(parents=True, exist_ok=True)
-        tmp = self.path.with_suffix(".tmp")
-        tmp.write_text(json.dumps({"notes": self.notes, "commit": self.commit, "synced_at": self.synced_at}), encoding="utf-8")
-        tmp.replace(self.path)
-
-    def sync(self, gitea: Gitea) -> dict:
-        """Bring the index to the branch head; returns a summary."""
-        with self.lock:
-            self.syncing = True
-        try:
-            head = gitea.head()
-            if head == self.commit and self.notes:
-                summary = {"commit": head, "changed": 0, "removed": 0, "notes": len(self.notes), "skipped": "head unchanged"}
-                self.synced_at = time.time()
-                self.save()
-                return summary
-            tree = gitea.tree(head)
-            slugs = {p[len(WIKI_DIR) + 1:-3]: (p, sha) for p, sha in tree.items() if not p.endswith("/_index.md")}
-            todo = [s for s, (p, sha) in slugs.items() if self.notes.get(s, {}).get("sha") != sha]
-            removed = [s for s in list(self.notes) if s not in slugs]
-            LOG.info("sync %s: %d notes, %d to embed, %d removed", head[:8], len(slugs), len(todo), len(removed))
-            fresh: dict[str, dict] = {}
-            for i in range(0, len(todo), 20):
-                chunk = todo[i:i + 20]
-                texts, titles = [], {}
-                for s in chunk:
-                    title, text = note_text(s, gitea.raw(slugs[s][0], head))
-                    titles[s] = title
-                    texts.append(text)
-                for s, vec in zip(chunk, embed(texts)):
-                    fresh[s] = {"sha": slugs[s][1], "title": titles[s], "vector": vec}
-            with self.lock:
-                for s in removed:
-                    self.notes.pop(s, None)
-                self.notes.update(fresh)
-                self.commit = head
-                self.synced_at = time.time()
-                self.last_error = ""
-                self.save()
-            return {"commit": head, "changed": len(todo), "removed": len(removed), "notes": len(self.notes)}
-        except Exception as e:
-            with self.lock:
-                self.last_error = str(e)[:300]
-            raise
-        finally:
-            with self.lock:
-                self.syncing = False
-
-    def rank(self, query: str, k: int = 60) -> list[dict]:
-        q = embed([query])[0]
-        with self.lock:
-            rows = [(s, n["title"], sum(a * b for a, b in zip(q, n["vector"]))) for s, n in self.notes.items()]
-        rows.sort(key=lambda r: -r[2])
-        return [{"slug": s, "title": t, "score": round(sc, 4)} for s, t, sc in rows[:k]]
-
-    def status(self) -> dict:
-        """`ready` = an index exists and can answer /rank; a failed sync keeps the last good index."""
-        with self.lock:
-            return {"ready": bool(self.notes), "notes": len(self.notes), "commit": self.commit,
-                    "synced_at": self.synced_at, "age_s": round(time.time() - self.synced_at) if self.synced_at else None,
-                    "syncing": self.syncing, "last_error": self.last_error or None}
-
-
-# ---------------------------------------------------------------- syncer
-
-
-class Syncer(threading.Thread):
-    """One worker: syncs on demand (webhook) and every RESYNC_SECONDS."""
-
-    def __init__(self, index: Index, creds: Credentials, interval: int):
-        super().__init__(daemon=True, name="syncer")
-        self.index, self.creds, self.interval = index, creds, interval
-        self.wake = threading.Event()
-
-    def trigger(self) -> None:
-        self.wake.set()
-
-    RETRY_SECONDS = 60
-
-    def run(self) -> None:
-        while True:
-            try:
-                data = self.creds.fetch()
-                summary = self.index.sync(Gitea(data["gitea_token"]))
-                LOG.info("sync done: %s", summary)
-                wait = self.interval
-            except Exception as e:  # keep serving the last good index; retry soon, not next interval
-                LOG.error("sync failed: %s", e)
-                wait = self.RETRY_SECONDS
-            self.wake.wait(wait)
-            self.wake.clear()
-
-
-def verify_signature(secret: str, body: bytes, signature: str) -> bool:
-    digest = hmac.new(secret.encode("utf-8"), body, hashlib.sha256).hexdigest()
-    return hmac.compare_digest(digest, (signature or "").strip())
-
-
-# ---------------------------------------------------------------- server
-
-
-def make_handler(index: Index, syncer: Syncer, creds: Credentials):
-    class Handler(BaseHTTPRequestHandler):
-        def log_message(self, fmt, *args):
-            LOG.info("%s %s", self.address_string(), fmt % args)
-
-        def send_json(self, code: int, obj) -> None:
-            data = json.dumps(obj).encode("utf-8")
-            self.send_response(code)
-            self.send_header("Content-Type", "application/json")
-            self.send_header("Content-Length", str(len(data)))
-            self.end_headers()
-            self.wfile.write(data)
-
-        def body(self) -> bytes:
-            n = int(self.headers.get("Content-Length") or 0)
-            return self.rfile.read(n) if n else b""
-
-        def do_GET(self):
-            if self.path == "/health":  # liveness: the process serves
-                return self.send_json(200, index.status())
-            if self.path == "/ready":  # readiness: an index exists
-                st = index.status()
-                return self.send_json(200 if st["ready"] else 503, st)
-            self.send_json(404, {"error": "not found"})
-
-        def do_POST(self):
-            if self.path == "/rank":
-                try:
-                    req = json.loads(self.body() or b"{}")
-                    query = str(req.get("query", "")).strip()
-                    k = max(1, min(int(req.get("k", 60)), 500))
-                    if not query:
-                        return self.send_json(400, {"error": "query required"})
-                    rows = index.rank(query, k)
-                    return self.send_json(200, {"rows": rows, "index": index.status()})
-                except ApiError as e:
-                    return self.send_json(502, {"error": str(e)})
-                except (ValueError, TypeError) as e:
-                    return self.send_json(400, {"error": str(e)})
-            if self.path == "/webhook":
-                raw = self.body()
-                try:
-                    secret = creds.fetch()["webhook_secret"]
-                except ApiError as e:
-                    return self.send_json(503, {"error": str(e)})
-                if not verify_signature(secret, raw, self.headers.get("X-Gitea-Signature", "")):
-                    return self.send_json(403, {"error": "bad signature"})
-                try:
-                    event = json.loads(raw or b"{}")
-                except ValueError:
-                    return self.send_json(400, {"error": "bad json"})
-                ref = event.get("ref", "")
-                if ref and ref != f"refs/heads/{NOTES_BRANCH}":
-                    return self.send_json(200, {"queued": False, "ref": ref})
-                syncer.trigger()
-                return self.send_json(200, {"queued": True})
-            self.send_json(404, {"error": "not found"})
-
-    return Handler
-
-
-def main() -> int:
-    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s", stream=sys.stdout)
-    index = Index(INDEX_PATH)
-    creds = Credentials()
-    syncer = Syncer(index, creds, RESYNC_SECONDS)
-    syncer.start()
-    server = ThreadingHTTPServer(("", LISTEN_PORT), make_handler(index, syncer, creds))
-    LOG.info("listening on :%d, repo %s@%s, embed %s", LISTEN_PORT, NOTES_REPO, NOTES_BRANCH, EMBED_URL)
-    server.serve_forever()
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())

docs/superpowers/plans/2026-09-25-paperclip.md

@@ -0,0 +1,976 @@
+# Paperclip Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Deploy Paperclip (server-local agent runs, option B) into the homelab cluster behind an Authentik gate, with CNPG Postgres and OpenBao-sourced secrets.
+
+**Architecture:** One Flux `Kustomization` (`paperclip`) applies `kubernetes/apps/paperclip/`: a single-replica Deployment on a ceph-block PVC, CNPG `pclp-pg` (PG17) with barman backups to Hetzner, and a Traefik Ingress that backends only the `paperclip-gate` Authentik outpost. Namespace, baseline CNP, ClusterSecretStore allowlist and image automation live in the existing `infrastructure` tree.
+
+**Tech Stack:** Flux v2 (kustomize-controller, image-reflector v1.2.5), CNPG + barman-cloud plugin, External Secrets → OpenBao, Cilium CNP, Traefik, Authentik blueprints.
+
+**Spec:** `docs/superpowers/specs/2026-09-25-paperclip-design.md`
+
+## Global Constraints
+
+- Worktree: `/home/krenner/source/homelab-worktrees/paperclip`, branch `feat/paperclip`. Never edit the main checkout.
+- Commits as claude-bot: `git -c user.name='Claude Bot' -c user.email='claude-bot@renner.uno' commit …`, message ends with `Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>`.
+- Push: `git -c credential.helper='store --file=/home/krenner/.config/claude-bot/creds' push …`. PR via `tea pr create --login claude-bot`. Never self-merge.
+- One Bash command per call: no `&&`, `;`, `|` chaining, no `2>/dev/null`, no `python -c`.
+- Namespace `paperclip`; hostname `paperclip.renner.uno`; LAN-only (no `dns.renner.uno/public` label).
+- Image `ghcr.io/paperclipai/paperclip`, tag `nightly`, digest-pinned, `digestReflectionPolicy: Always`.
+- Postgres: CNPG `pclp-pg`, `ghcr.io/cloudnative-pg/postgresql:17`, 2 instances, `ceph-block`.
+- PVC `paperclip-data`, `ceph-block`, 20Gi, mounted at `/paperclip`.
+- OpenBao key `apps/paperclip` with fields `BETTER_AUTH_SECRET`, `PAPERCLIP_TOOL_ACTION_SIGNING_SECRET`, `PAPERCLIP_SECRETS_MASTER_KEY` (64-char hex), `CLAUDE_CODE_OAUTH_TOKEN`.
+- Env: `PAPERCLIP_DEPLOYMENT_MODE=authenticated`, `PAPERCLIP_DEPLOYMENT_EXPOSURE=private`, `PAPERCLIP_PUBLIC_URL=https://paperclip.renner.uno`, `HOST=0.0.0.0`, `PORT=3100`.
+- Manifest comments: one terse factual line; rationale goes in commit messages/PR.
+- No `kubectl apply`/`exec`/`rollout restart` against the live cluster; Flux applies after merge. Server-side `--dry-run=server` is allowed.
+
+## Review Focus
+
+1. **Pod replaced mid agent run** (image bump, secret rotation via Reloader, node drain) — `Recreate` kills in-flight runs; expected: Paperclip marks them failed/recoverable on restart, not wedged. Covered in Task 9 Step 6.
+2. **Shared OAuth token expired or missing** — expected: agent runs fail with a visible auth error in Paperclip, server stays healthy. Covered in Task 9 Step 7.
+3. **Direct access bypassing the gate** from another namespace pod — expected: connection refused/dropped by CNP. Covered in Task 9 Step 4.
+4. **Unauthenticated LAN request** to `paperclip.renner.uno` — expected: redirect to Authentik, never Paperclip's UI. Covered in Task 9 Step 3.
+5. **Nightly image with a breaking DB migration** — expected: a base backup exists before the first bump lands. Covered in Task 9 Step 5.
+
+---
+
+### Task 1: Upstream facts that decide two manifest values
+
+No repo changes; produces two decisions consumed by Task 6.
+
+**Interfaces:**
+- Produces: `SIGNUP_DISABLED` (`"true"` or `"false"`), `CONFIG_PARTIAL_OK` (yes/no), `ENTRYPOINT_NEEDS_ROOT` (yes/no).
+
+- [ ] **Step 1: Sign-up vs instance claim.** Read upstream `server/src/auth/better-auth.ts`, `server/src/routes/auth.ts`, and the claim flow for `authenticated`+`private` (search `claim` under `server/src`):
+
+Run: `gh api repos/paperclipai/paperclip/contents/server/src/routes/auth.ts -H "Accept: application/vnd.github.raw"`
+
+Decide: if the first-admin claim creates the account through a path that ignores `disableSignUp` (e.g. invite/claim token handler calling the adapter directly), `SIGNUP_DISABLED="true"`. Otherwise `SIGNUP_DISABLED="false"` — the Authentik gate is the access control, and a follow-up PR flips it to `"true"` after the admin is claimed.
+
+- [ ] **Step 2: Partial config.json.** Read `server/src/config.ts` and the config-file loader it calls (search `PAPERCLIP_CONFIG` / `readConfigFile`).
+
+Run: `gh api repos/paperclipai/paperclip/contents/server/src/config.ts -H "Accept: application/vnd.github.raw"`
+
+Decide `CONFIG_PARTIAL_OK=yes` if a file containing only `{"telemetry":{"enabled":false}}` parses (every other section optional/defaulted). If not, stop and report to the user — do not invent config fields.
+
+- [ ] **Step 3: Entrypoint privilege.**
+
+Run: `gh api repos/paperclipai/paperclip/contents/docker-entrypoint.sh -H "Accept: application/vnd.github.raw"`
+
+`ENTRYPOINT_NEEDS_ROOT=yes` if it calls `usermod`/`groupmod`/`chown`/`gosu`. Record the result for Task 6.
+
+- [ ] **Step 4: Anonymous health response.**
+
+Run: `gh api repos/paperclipai/paperclip/contents/server/src/routes/health.ts -H "Accept: application/vnd.github.raw"`
+
+Confirm `GET /api/health` returns HTTP 200 to an anonymous caller when healthy (probes use it). If it returns 401 anonymously, use a `tcpSocket` probe on 3100 in Task 6 instead.
+
+---
+
+### Task 2: Namespace, baseline CNP, ClusterSecretStore allowlist
+
+**Files:**
+- Modify: `kubernetes/infrastructure/namespaces/namespaces.yaml` (append)
+- Create: `kubernetes/infrastructure/network-policies/paperclip-ns.yaml`
+- Modify: `kubernetes/infrastructure/network-policies/kustomization.yaml`
+- Modify: `kubernetes/infrastructure/post-crds/openbao-clustersecretstore.yaml` (sorted list)
+
+**Interfaces:**
+- Produces: namespace `paperclip`; ESO may reference `openbao` from it.
+
+- [ ] **Step 1: Append the namespace** to the end of `namespaces.yaml`:
+
+```yaml
+---
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: paperclip
+```
+
+- [ ] **Step 2: Create `paperclip-ns.yaml`** (copy of `paperless-ns.yaml`'s `default-deny-ingress`, ns renamed, paperless-only CNPs dropped):
+
+```yaml
+# Traefik is not admitted: the only entry is the paperclip-gate outpost (apps/paperclip/network-policy.yaml).
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: default-deny-ingress
+  namespace: paperclip
+spec:
+  endpointSelector: {}
+  ingress:
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": paperclip
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": database
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": monitoring
+    - fromEntities:
+        - host
+        - remote-node
+    - fromEndpoints:
+        - matchExpressions:
+            - key: io.kubernetes.pod.namespace
+              operator: Exists
+      icmps:
+        - fields:
+            - { type: 0,   family: IPv4 }
+            - { type: 3,   family: IPv4 }
+            - { type: 8,   family: IPv4 }
+            - { type: 11,  family: IPv4 }
+            - { type: 12,  family: IPv4 }
+            - { type: 1,   family: IPv6 }
+            - { type: 2,   family: IPv6 }
+            - { type: 3,   family: IPv6 }
+            - { type: 4,   family: IPv6 }
+            - { type: 128, family: IPv6 }
+            - { type: 129, family: IPv6 }
+            - { type: 133, family: IPv6 }
+            - { type: 134, family: IPv6 }
+            - { type: 135, family: IPv6 }
+            - { type: 136, family: IPv6 }
+            - { type: 143, family: IPv6 }
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": paperclip
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": authentik
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ingress
+            app.kubernetes.io/name: traefik
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": rook-ceph
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": kube-system
+            k8s-app: kube-dns
+      toPorts:
+        - ports:
+            - port: "53"
+              protocol: UDP
+    - toEntities:
+        - host
+        - remote-node
+        - world
+    - toEndpoints:
+        - matchExpressions:
+            - key: io.kubernetes.pod.namespace
+              operator: Exists
+      icmps:
+        - fields:
+            - { type: 0,   family: IPv4 }
+            - { type: 3,   family: IPv4 }
+            - { type: 8,   family: IPv4 }
+            - { type: 11,  family: IPv4 }
+            - { type: 12,  family: IPv4 }
+            - { type: 1,   family: IPv6 }
+            - { type: 2,   family: IPv6 }
+            - { type: 3,   family: IPv6 }
+            - { type: 4,   family: IPv6 }
+            - { type: 128, family: IPv6 }
+            - { type: 129, family: IPv6 }
+            - { type: 133, family: IPv6 }
+            - { type: 134, family: IPv6 }
+            - { type: 135, family: IPv6 }
+            - { type: 136, family: IPv6 }
+            - { type: 143, family: IPv6 }
+```
+
+- [ ] **Step 3: Register it** — in `network-policies/kustomization.yaml` add `  - paperclip-ns.yaml` directly after `  - paperless-ns.yaml`.
+
+- [ ] **Step 4: Allowlist the namespace** — in `openbao-clustersecretstore.yaml` insert `        - paperclip` between `        - monitoring` and `        - paperless` (list is sorted).
+
+- [ ] **Step 5: Render check.**
+
+Run: `kubectl kustomize kubernetes/infrastructure/network-policies`
+Expected: output contains `namespace: paperclip`, exit 0.
+
+Run: `kubectl apply --dry-run=server -f kubernetes/infrastructure/post-crds/openbao-clustersecretstore.yaml`
+Expected: `clustersecretstore.external-secrets.io/openbao configured (server dry run)`.
+
+- [ ] **Step 6: Commit** `feat(paperclip): namespace, baseline CNP and OpenBao allowlist`.
+
+---
+
+### Task 3: OpenBao secret `apps/paperclip`
+
+Out-of-repo; must exist before merge or the ExternalSecret never becomes Ready.
+
+**Interfaces:**
+- Produces: OpenBao KV `apps/paperclip` with the four fields in Global Constraints.
+
+- [ ] **Step 1:** Ask wiki-synth for the claude-bot OpenBao write procedure (note `claude-bot-openbao-write-only-identity`): login method, KV mount, exact `bao kv put`/`patch` syntax for an `apps/<name>` key. Use exactly what it returns.
+
+- [ ] **Step 2: Write the generated fields** as claude-bot (write-only identity — never read values back). Values generated inline by the command:
+
+`BETTER_AUTH_SECRET` = `openssl rand -hex 32`, `PAPERCLIP_TOOL_ACTION_SIGNING_SECRET` = `openssl rand -hex 32`, `PAPERCLIP_SECRETS_MASTER_KEY` = `openssl rand -hex 32` (64 hex chars = 32 bytes, accepted format).
+
+Generate each with a separate `openssl rand -hex 32` call and pass it directly into the `bao kv put` from Step 1; do not echo or log them.
+
+- [ ] **Step 3: Ask the user** to add the shared token themselves (never handle it): "Run `claude setup-token`, then `bao kv patch <mount>/apps/paperclip CLAUDE_CODE_OAUTH_TOKEN=<token>`." Wait for confirmation.
+
+- [ ] **Step 4: Verify field names only** (no values): `bao kv metadata get <mount>/apps/paperclip` shows a current version ≥2. If claude-bot cannot read metadata, ask the user to confirm the four keys exist.
+
+---
+
+### Task 4: Image automation + initial digest
+
+**Files:**
+- Create: `kubernetes/infrastructure/image-automation/paperclip.yaml`
+- Modify: `kubernetes/infrastructure/image-automation/kustomization.yaml`
+
+**Interfaces:**
+- Produces: ImagePolicy `flux-system:paperclip`; `INITIAL_DIGEST` (`sha256:…`) for Task 6.
+
+- [ ] **Step 1: Create `paperclip.yaml`:**
+
+```yaml
+apiVersion: image.toolkit.fluxcd.io/v1
+kind: ImageRepository
+metadata:
+  name: paperclip
+  namespace: flux-system
+spec:
+  image: ghcr.io/paperclipai/paperclip
+  interval: 1h
+---
+# Upstream publishes no semver tags; track the nightly tag by digest.
+apiVersion: image.toolkit.fluxcd.io/v1
+kind: ImagePolicy
+metadata:
+  name: paperclip
+  namespace: flux-system
+spec:
+  imageRepositoryRef:
+    name: paperclip
+  digestReflectionPolicy: Always
+  filterTags:
+    pattern: '^nightly$'
+  policy:
+    alphabetical:
+      order: asc
+```
+
+- [ ] **Step 2: Register** — add `  - paperclip.yaml` after `  - paperless-ngx.yaml` in `image-automation/kustomization.yaml`.
+
+- [ ] **Step 3: Server dry-run.**
+
+Run: `kubectl apply --dry-run=server -f kubernetes/infrastructure/image-automation/paperclip.yaml`
+Expected: both objects `created (server dry run)`; no `unknown field "digestReflectionPolicy"` error.
+
+- [ ] **Step 4: Resolve the current nightly digest.**
+
+The workstation has `docker` only (no skopeo/crane/buildx, checked 2026-09-25); `docker manifest inspect` lists per-platform digests but not the index digest, so pull once:
+
+Run: `docker pull ghcr.io/paperclipai/paperclip:nightly`
+Run: `docker image inspect --format '{{index .RepoDigests 0}}' ghcr.io/paperclipai/paperclip:nightly`
+Expected: `ghcr.io/paperclipai/paperclip@sha256:<64 hex>`. Record the `sha256:…` part as `INITIAL_DIGEST`.
+Run: `docker image rm ghcr.io/paperclipai/paperclip:nightly`
+
+- [ ] **Step 5: Commit** `feat(paperclip): Flux image automation tracks the nightly digest`.
+
+---
+
+### Task 5: Postgres (CNPG) and secrets in the app directory
+
+**Files:**
+- Create: `kubernetes/apps/paperclip/cluster.yaml`, `objectstore.yaml`, `scheduled-backup.yaml`, `hetzner-creds-externalsecret.yaml`, `externalsecret.yaml`
+
+**Interfaces:**
+- Produces: Secret `pclp-pg-app` (CNPG, key `uri`), Service `pclp-pg-rw`, Secret `paperclip-secrets`.
+
+- [ ] **Step 1: `cluster.yaml`:**
+
+```yaml
+apiVersion: postgresql.cnpg.io/v1
+kind: Cluster
+metadata:
+  name: pclp-pg
+  namespace: paperclip
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled # data guard
+spec:
+  instances: 2
+  smartShutdownTimeout: 10
+  stopDelay: 60
+  # PG17: upstream's external Postgres target (docker/docker-compose.yml).
+  imageName: ghcr.io/cloudnative-pg/postgresql:17
+  storage:
+    storageClass: ceph-block
+    size: 10Gi
+  plugins:
+    - name: barman-cloud.cloudnative-pg.io
+      isWALArchiver: true
+      parameters:
+        barmanObjectName: pclp-pg-store
+  bootstrap:
+    initdb:
+      database: paperclip
+      owner: paperclip
+  postgresql:
+    parameters:
+      shared_buffers: "256MB"
+      max_connections: "200"
+      checkpoint_timeout: "15min"
+      max_wal_size: "2GB"
+      wal_compression: "zstd"
+      archive_timeout: "30min"
+    synchronous:
+      method: any
+      number: 1
+      dataDurability: preferred
+  resources:
+    requests:
+      cpu: 200m
+      memory: 512Mi
+    limits:
+      memory: 1Gi
+```
+
+- [ ] **Step 2: `objectstore.yaml`:**
+
+```yaml
+apiVersion: barmancloud.cnpg.io/v1
+kind: ObjectStore
+metadata:
+  name: pclp-pg-store
+  namespace: paperclip
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled # backup target
+spec:
+  configuration:
+    destinationPath: s3://renner-uno-k8s-cnpg/pclp-pg
+    endpointURL: https://fsn1.your-objectstorage.com
+    s3Credentials:
+      accessKeyId:
+        name: hetzner-s3-creds
+        key: ACCESS_KEY_ID
+      secretAccessKey:
+        name: hetzner-s3-creds
+        key: ACCESS_SECRET_KEY
+      region:
+        name: hetzner-s3-creds
+        key: REGION
+    wal:
+      compression: gzip
+    data:
+      compression: gzip
+  retentionPolicy: "30d"
+```
+
+- [ ] **Step 3: `scheduled-backup.yaml`:**
+
+```yaml
+apiVersion: postgresql.cnpg.io/v1
+kind: ScheduledBackup
+metadata:
+  name: pclp-pg-daily
+  namespace: paperclip
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled # backup schedule
+spec:
+  schedule: "0 45 2 * * *"
+  immediate: true
+  backupOwnerReference: self
+  cluster:
+    name: pclp-pg
+  method: plugin
+  pluginConfiguration:
+    name: barman-cloud.cloudnative-pg.io
+```
+
+- [ ] **Step 4: `hetzner-creds-externalsecret.yaml`:**
+
+```yaml
+apiVersion: external-secrets.io/v1
+kind: ExternalSecret
+metadata:
+  name: hetzner-s3-creds
+  namespace: paperclip
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled # barman S3 creds
+spec:
+  refreshInterval: 1h
+  secretStoreRef:
+    kind: ClusterSecretStore
+    name: openbao
+  target:
+    name: hetzner-s3-creds
+  dataFrom:
+    - extract:
+        key: infrastructure/hetzner-s3
+```
+
+- [ ] **Step 5: `externalsecret.yaml`:**
+
+```yaml
+apiVersion: external-secrets.io/v1
+kind: ExternalSecret
+metadata:
+  name: paperclip-secrets
+  namespace: paperclip
+spec:
+  refreshInterval: 10m
+  secretStoreRef:
+    kind: ClusterSecretStore
+    name: openbao
+  target:
+    name: paperclip-secrets
+  dataFrom:
+    - extract:
+        key: apps/paperclip
+```
+
+- [ ] **Step 6: Commit** `feat(paperclip): CNPG pclp-pg with barman backups and OpenBao secrets`. (Rendered together with Task 6's kustomization.)
+
+---
+
+### Task 6: Server workload, storage, ingress, network policy
+
+**Files:**
+- Create: `kubernetes/apps/paperclip/pvc.yaml`, `paperclip.yaml`, `gate-service.yaml`, `ingress.yaml`, `network-policy.yaml`, `kustomization.yaml`
+
+**Interfaces:**
+- Consumes: `pclp-pg-app` (`uri`), `paperclip-secrets`, ImagePolicy `flux-system:paperclip`, `INITIAL_DIGEST` (Task 4), `SIGNUP_DISABLED` / `ENTRYPOINT_NEEDS_ROOT` / probe choice (Task 1).
+- Produces: pods labelled `app.kubernetes.io/name: paperclip` listening on 3100; Service `paperclip:3100`; ExternalName `ak-outpost-paperclip-gate:9000`.
+
+- [ ] **Step 1: `pvc.yaml`:**
+
+```yaml
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: paperclip-data
+  namespace: paperclip
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled # data guard
+spec:
+  accessModes: [ReadWriteOnce]
+  storageClassName: ceph-block
+  resources:
+    requests:
+      storage: 20Gi
+```
+
+- [ ] **Step 2: `paperclip.yaml`** — replace `<INITIAL_DIGEST>` with the Task 4 value and `<SIGNUP_DISABLED>` with the Task 1 value. If Task 1 found `ENTRYPOINT_NEEDS_ROOT=no`, replace the container `securityContext` with `runAsNonRoot: true`, `runAsUser: 1000`, `runAsGroup: 1000`, `allowPrivilegeEscalation: false`, `capabilities: {drop: ["ALL"]}`. If Task 1 chose `tcpSocket`, replace each `httpGet` block with `tcpSocket: {port: http}`.
+
+```yaml
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: paperclip
+  namespace: paperclip
+  labels:
+    app.kubernetes.io/name: paperclip
+  annotations:
+    secret.reloader.stakater.com/reload: paperclip-secrets
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: paperclip
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: paperclip
+    spec:
+      enableServiceLinks: false
+      securityContext:
+        fsGroup: 1000
+        fsGroupChangePolicy: OnRootMismatch
+        seccompProfile:
+          type: RuntimeDefault
+      initContainers:
+        # Telemetry is only configurable in config.json.
+        - name: disable-telemetry
+          image: ghcr.io/paperclipai/paperclip:nightly@<INITIAL_DIGEST> # {"$imagepolicy": "flux-system:paperclip"}
+          command:
+            - sh
+            - -c
+            - |
+              set -eu
+              f=/paperclip/instances/default/config.json
+              mkdir -p "$(dirname "$f")"
+              if [ -s "$f" ]; then
+                jq '.telemetry.enabled = false' "$f" > "$f.tmp"
+              else
+                printf '%s\n' '{"telemetry":{"enabled":false}}' > "$f.tmp"
+              fi
+              mv "$f.tmp" "$f"
+          securityContext:
+            runAsNonRoot: true
+            runAsUser: 1000
+            runAsGroup: 1000
+            allowPrivilegeEscalation: false
+            capabilities:
+              drop: ["ALL"]
+          volumeMounts:
+            - name: data
+              mountPath: /paperclip
+      containers:
+        - name: paperclip
+          image: ghcr.io/paperclipai/paperclip:nightly@<INITIAL_DIGEST> # {"$imagepolicy": "flux-system:paperclip"}
+          ports:
+            - name: http
+              containerPort: 3100
+          env:
+            - name: DATABASE_URL
+              valueFrom:
+                secretKeyRef:
+                  name: pclp-pg-app
+                  key: uri
+            - name: HOST
+              value: "0.0.0.0"
+            - name: PORT
+              value: "3100"
+            - name: PAPERCLIP_DEPLOYMENT_MODE
+              value: authenticated
+            - name: PAPERCLIP_DEPLOYMENT_EXPOSURE
+              value: private
+            - name: PAPERCLIP_PUBLIC_URL
+              value: https://paperclip.renner.uno
+            - name: PAPERCLIP_AUTH_DISABLE_SIGN_UP
+              value: "<SIGNUP_DISABLED>"
+          envFrom:
+            - secretRef:
+                name: paperclip-secrets
+          volumeMounts:
+            - name: data
+              mountPath: /paperclip
+          startupProbe:
+            httpGet:
+              path: /api/health
+              port: http
+            periodSeconds: 10
+            timeoutSeconds: 5
+            failureThreshold: 60
+          livenessProbe:
+            httpGet:
+              path: /api/health
+              port: http
+            periodSeconds: 30
+            timeoutSeconds: 10
+          readinessProbe:
+            httpGet:
+              path: /api/health
+              port: http
+            periodSeconds: 10
+            timeoutSeconds: 5
+          resources:
+            requests:
+              cpu: 250m
+              memory: 1Gi
+            limits:
+              memory: 4Gi
+          securityContext:
+            runAsNonRoot: false
+            allowPrivilegeEscalation: true
+            capabilities:
+              drop: ["ALL"]
+              add: ["CHOWN", "DAC_OVERRIDE", "FOWNER", "SETGID", "SETUID"]
+      volumes:
+        - name: data
+          persistentVolumeClaim:
+            claimName: paperclip-data
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: paperclip
+  namespace: paperclip
+spec:
+  selector:
+    app.kubernetes.io/name: paperclip
+  ports:
+    - name: http
+      port: 3100
+      targetPort: http
+```
+
+- [ ] **Step 3: `gate-service.yaml`:**
+
+```yaml
+# ExternalName mirror so the Ingress can backend the paperclip-gate outpost in the authentik namespace.
+apiVersion: v1
+kind: Service
+metadata:
+  name: ak-outpost-paperclip-gate
+  namespace: paperclip
+spec:
+  type: ExternalName
+  externalName: ak-outpost-paperclip-gate.authentik.svc.cluster.local
+  ports:
+    - port: 9000
+      targetPort: 9000
+```
+
+- [ ] **Step 4: `ingress.yaml`:**
+
+```yaml
+# LAN-only (no dns.renner.uno/public label); backends the gate outpost, never the Service.
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: paperclip
+  namespace: paperclip
+  annotations:
+    cert-manager.io/cluster-issuer: letsencrypt-cluster
+spec:
+  ingressClassName: traefik
+  tls:
+    - hosts:
+        - paperclip.renner.uno
+      secretName: paperclip-tls
+  rules:
+    - host: paperclip.renner.uno
+      http:
+        paths:
+          - path: /
+            pathType: Prefix
+            backend:
+              service:
+                name: ak-outpost-paperclip-gate
+                port:
+                  number: 9000
+```
+
+- [ ] **Step 5: `network-policy.yaml`** (Traefik pod-side port is 8443, verified via `kubectl get pods -n ingress -l app.kubernetes.io/name=traefik`):
+
+```yaml
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: paperclip-gate-ingress
+  namespace: paperclip
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: paperclip
+  ingress:
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": authentik
+            goauthentik.io/outpost-name: paperclip-gate
+      toPorts:
+        - ports:
+            - port: "3100"
+              protocol: TCP
+---
+# Covers the server and every agent it spawns (same pod).
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: paperclip-egress
+  namespace: paperclip
+spec:
+  endpointSelector:
+    matchLabels:
+      app.kubernetes.io/name: paperclip
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": paperclip
+            cnpg.io/cluster: pclp-pg
+      toPorts:
+        - ports:
+            - port: "5432"
+              protocol: TCP
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": kube-system
+            k8s-app: kube-dns
+      toPorts:
+        - ports:
+            - port: "53"
+              protocol: UDP
+            - port: "53"
+              protocol: TCP
+    # Gitea, Woodpecker, HA, Loki via their *.renner.uno names.
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ingress
+            app.kubernetes.io/name: traefik
+      toPorts:
+        - ports:
+            - port: "8443"
+              protocol: TCP
+    - toEntities:
+        - kube-apiserver
+      toPorts:
+        - ports:
+            - port: "6443"
+              protocol: TCP
+    - toEntities:
+        - world
+      toPorts:
+        - ports:
+            - port: "443"
+              protocol: TCP
+```
+
+- [ ] **Step 6: `kustomization.yaml`:**
+
+```yaml
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+resources:
+  - externalsecret.yaml
+  - hetzner-creds-externalsecret.yaml
+  - cluster.yaml
+  - objectstore.yaml
+  - scheduled-backup.yaml
+  - pvc.yaml
+  - paperclip.yaml
+  - gate-service.yaml
+  - ingress.yaml
+  - network-policy.yaml
+```
+
+- [ ] **Step 7: Render check.**
+
+Run: `kubectl kustomize kubernetes/apps/paperclip`
+Expected: exit 0; output contains `kind: Cluster`, `kind: Deployment`, `claimName: paperclip-data`, and no literal `<INITIAL_DIGEST>` or `<SIGNUP_DISABLED>`.
+
+Run: `grep -rn "<INITIAL_DIGEST>\|<SIGNUP_DISABLED>" kubernetes/apps/paperclip`
+Expected: no output, exit 1.
+
+- [ ] **Step 8: Commit** `feat(paperclip): server Deployment, PVC, gated Ingress and CNPs`.
+
+---
+
+### Task 7: Authentik gate
+
+**Files:**
+- Modify: `kubernetes/apps/authentik/blueprints-configmap.yaml` (insert before the line `      # navidrome-gate: the Navidrome web UI.`)
+- Create: `kubernetes/apps/authentik/paperclip-gate-egress.yaml`
+- Modify: `kubernetes/apps/authentik/kustomization.yaml`
+
+**Interfaces:**
+- Produces: outpost pods labelled `goauthentik.io/outpost-name: paperclip-gate`, Service `ak-outpost-paperclip-gate.authentik:9000`.
+
+- [ ] **Step 1: Insert blueprint entries** (same indentation as the neighbouring `wiki-gate` block):
+
+```yaml
+      # paperclip-gate: Paperclip UI; the CNP admits only this outpost. Paperclip keeps its own login behind it.
+      - id: paperclip-users-group
+        model: authentik_core.group
+        identifiers:
+          name: Paperclip Users
+        state: present
+        attrs:
+          is_superuser: false
+          users:
+            - !Find [authentik_core.user, [username, k.renner]]
+
+      - id: paperclip-gate-provider
+        model: authentik_providers_proxy.proxyprovider
+        identifiers:
+          name: paperclip-gate
+        state: present
+        attrs:
+          authentication_flow: !Find [authentik_flows.flow, [slug, passkey-passwordless-authentication]]
+          authorization_flow: !Find [authentik_flows.flow, [slug, passkey-stepup-authorization]]
+          invalidation_flow: !Find [authentik_flows.flow, [slug, default-provider-invalidation-flow]]
+          mode: proxy
+          internal_host: http://paperclip.paperclip.svc.cluster.local:3100
+          internal_host_ssl_validation: false
+          external_host: https://paperclip.renner.uno
+          cookie_domain: renner.uno
+          access_token_validity: hours=24
+          refresh_token_validity: days=30
+          intercept_header_auth: true
+
+      - model: authentik_core.application
+        identifiers:
+          slug: paperclip-gate
+        state: present
+        attrs:
+          name: Paperclip
+          slug: paperclip-gate
+          provider: !KeyOf paperclip-gate-provider
+          meta_launch_url: https://paperclip.renner.uno
+          meta_description: "Paperclip agent orchestration"
+
+      - model: authentik_policies.policybinding
+        identifiers:
+          target: !Find [authentik_core.application, [slug, paperclip-gate]]
+          group: !KeyOf paperclip-users-group
+          order: 0
+        state: present
+        attrs:
+          enabled: true
+          negate: false
+
+      - model: authentik_outposts.outpost
+        identifiers:
+          name: paperclip-gate
+        state: present
+        attrs:
+          type: proxy
+          service_connection: !KeyOf k8s-conn
+          providers:
+            - !KeyOf paperclip-gate-provider
+          config:
+            log_level: info
+            authentik_host: "http://authentik-server.authentik.svc.cluster.local"
+            authentik_host_insecure: false
+            authentik_host_browser: "https://auth.renner.uno/"
+            refresh_interval: "minutes=5"
+            container_image: null
+            object_naming_template: "ak-outpost-%(name)s"
+            kubernetes_namespace: authentik
+            kubernetes_replicas: 1
+            kubernetes_service_type: ClusterIP
+            kubernetes_disabled_components: ["ingress"]
+            kubernetes_image_pull_secrets: []
+            kubernetes_ingress_annotations: {}
+            kubernetes_ingress_class_name: null
+            kubernetes_ingress_secret_name: "authentik-outpost-tls"
+            kubernetes_json_patches: null
+
+```
+
+- [ ] **Step 2: `paperclip-gate-egress.yaml`:**
+
+```yaml
+# Lets the paperclip-gate outpost reach the Paperclip server; scoped to this outpost's pods.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: paperclip-gate-egress
+  namespace: authentik
+spec:
+  endpointSelector:
+    matchLabels:
+      goauthentik.io/outpost-name: paperclip-gate
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": paperclip
+            app.kubernetes.io/name: paperclip
+      toPorts:
+        - ports:
+            - port: "3100"
+              protocol: TCP
+```
+
+- [ ] **Step 3: Register** — add `  - paperclip-gate-egress.yaml` after `  - paperless-gate-egress.yaml` in `apps/authentik/kustomization.yaml`.
+
+- [ ] **Step 4: Render + dry-run.**
+
+Run: `kubectl kustomize kubernetes/apps/authentik`
+Expected: exit 0 (valid YAML incl. the blueprint ConfigMap).
+
+Run: `kubectl apply --dry-run=server -f kubernetes/apps/authentik/paperclip-gate-egress.yaml`
+Expected: `ciliumnetworkpolicy.cilium.io/paperclip-gate-egress created (server dry run)`.
+
+- [ ] **Step 5: Commit** `feat(authentik): paperclip-gate outpost and Paperclip Users group`.
+
+---
+
+### Task 8: Flux Kustomization
+
+**Files:**
+- Create: `kubernetes/clusters/homelab/paperclip.yaml`
+
+- [ ] **Step 1:**
+
+```yaml
+apiVersion: kustomize.toolkit.fluxcd.io/v1
+kind: Kustomization
+metadata:
+  name: paperclip
+  namespace: flux-system
+spec:
+  interval: 10m
+  retryInterval: 1m
+  timeout: 20m
+  sourceRef:
+    kind: GitRepository
+    name: flux-system
+  path: ./kubernetes/apps/paperclip
+  prune: true
+  wait: true
+  dependsOn:
+    - name: infrastructure
+    - name: infrastructure-crds
+    - name: infrastructure-ceph-cluster
+    - name: infrastructure-cnpg-barman
+  healthCheckExprs:
+    - apiVersion: postgresql.cnpg.io/v1
+      kind: Cluster
+      failed: "has(status.phase) && status.phase.startsWith('Failed')"
+      current: "has(status.phase) && (status.phase == 'Cluster in healthy state' || status.readyInstances > 0)"
+```
+
+- [ ] **Step 2: Server dry-run.**
+
+Run: `kubectl apply --dry-run=server -f kubernetes/clusters/homelab/paperclip.yaml`
+Expected: `kustomization.kustomize.toolkit.fluxcd.io/paperclip created (server dry run)`.
+
+- [ ] **Step 3: Commit** `feat(paperclip): Flux Kustomization`.
+
+---
+
+### Task 9: PR, merge by user, live verification
+
+- [ ] **Step 1: Push and open the PR** (Task 3 must be complete).
+
+Run: `git -c credential.helper='store --file=/home/krenner/.config/claude-bot/creds' push -u origin feat/paperclip`
+Run: `tea pr create --login claude-bot --base master --head feat/paperclip --title "feat(paperclip): deploy Paperclip (server-local runs, trial)" --description "<summary of spec decisions + link to spec + bootstrap steps>\n\n🤖 Generated with [Claude Code](https://claude.com/claude-code)"`
+Then wait for the user to merge. Do not merge.
+
+- [ ] **Step 2: Reconciliation.** After merge:
+
+Run: `flux get kustomizations paperclip`
+Expected: `Ready True`. If False, read `kubectl describe kustomization paperclip -n flux-system` and the failing object's events; diagnose, don't poke.
+
+Run: `kubectl get cluster -n paperclip pclp-pg`
+Expected: `Cluster in healthy state`, 2 ready instances.
+
+Run: `kubectl get pods -n paperclip -l app.kubernetes.io/name=paperclip`
+Expected: `1/1 Running`.
+
+- [ ] **Step 3: Gate enforced (Review Focus 4).**
+
+Run: `curl -sS -o /dev/null -w '%{http_code} %{redirect_url}\n' https://paperclip.renner.uno/`
+Expected: `302` redirecting to `https://paperclip.renner.uno/outpost.goauthentik.io/…` or `auth.renner.uno`, never `200`.
+
+- [ ] **Step 4: Gate not bypassable (Review Focus 3).**
+
+`default` has no CNPs (verified 2026-09-25), so its egress is open and only paperclip's ingress policy can block the probe.
+
+Run: `kubectl get ciliumnetworkpolicies -n default`
+Expected: `No resources found` (else pick another CNP-free namespace).
+
+Run: `kubectl run cnp-probe -n default --rm -i --restart=Never --image=curlimages/curl:8.10.1 -- curl -sS -m 5 http://paperclip.paperclip.svc.cluster.local:3100/api/health`
+Expected: timeout (exit 28), no HTTP response.
+
+- [ ] **Step 5: First backup (Review Focus 5).**
+
+Run: `kubectl get backup -n paperclip`
+Expected: one `pclp-pg-daily-…` with phase `completed`.
+
+- [ ] **Step 6: Image policy.**
+
+Run: `kubectl get imagepolicy -n flux-system paperclip -o jsonpath='{.status.latestRef}'`
+Expected: `nightly` tag with a `sha256:` digest.
+
+- [ ] **Step 7: Hand-over to the user (browser steps, theirs):**
+  1. Open `https://paperclip.renner.uno`, pass Authentik, claim the instance as admin.
+  2. Create a test agent with the `claude-local` adapter and a trivial task ("reply with OK"); it must complete with the shared token (Review Focus 2: if it fails, report the run's error text — an expired token must show as an auth error, not a hang).
+  3. If `SIGNUP_DISABLED` was `"false"`, open the follow-up PR flipping `PAPERCLIP_AUTH_DISABLE_SIGN_UP` to `"true"` once the admin is claimed.
+  4. Review Focus 1 is observed on the first nightly bump: check the interrupted-run status in the UI.
+
+- [ ] **Step 8: Wiki.** Route the verified Paperclip facts (execution topology, env vars, image tags, digest-reflection usage) through wiki-ingest → wiki-distill once the deployment has proven out.

docs/superpowers/specs/2026-09-25-paperclip-design.md

@@ -0,0 +1,206 @@
+# Paperclip in the homelab cluster — design
+
+Date: 2026-09-25
+Status: draft, awaiting review
+
+## Goal
+
+Run [Paperclip](https://github.com/paperclipai/paperclip) (MIT, agent-orchestration
+server + UI) in the homelab cluster for real use: its agents do actual work against
+homelab systems. This spec covers the **platform** only. Bot identities and their
+credentials are out of scope (see "Deferred").
+
+## Decisions
+
+| Topic | Decision |
+|---|---|
+| Execution model | **B — server-local runs** (upstream default): agent CLIs are child processes of the server, in the server container. Trial only. |
+| Upgrade path | If Paperclip is kept: **C** — `@paperclipai/plugin-kubernetes` sandbox provider + kubernetes-sigs agent-sandbox controller (`agents.x-k8s.io`), per-run Pods and Secrets. Separate spec. |
+| Claude auth | **One shared Claude subscription AI connection**, created once via Paperclip's local sign-in (`claude auth login` into `ai-local-logins/<id>`); Paperclip stores and refreshes it. Server-env `CLAUDE_CODE_OAUTH_TOKEN` is not used: `stripAiAuthBindings` blanks it for managed runs, and upstream accepts no pasted setup-token (the device-code flow needs a Daytona sandbox). |
+| Bot credentials | Out of scope. Future "AIR-Manager" Paperclip role handles permission requests. |
+| Exposure | LAN-only, behind an Authentik forward-auth gate, then Paperclip's own login. |
+
+### Known trade-off of B (accepted for the trial)
+
+Upstream `runChildProcess()` spawns every local adapter with
+`{...sanitizeInheritedPaperclipEnv(process.env), ...opts.env}`; the sanitizer strips only
+`PAPERCLIP_*` keys. Every agent therefore sees `DATABASE_URL`, `BETTER_AUTH_SECRET`,
+the shared filesystem (including every other agent's workspace and the
+`ai-local-logins` credentials) and every agent-scoped secret it can read from disk. Under B, no bot may hold
+a credential we would not tolerate any other bot using. Per-bot isolation is C's job.
+
+## Architecture
+
+```
+LAN client ─▶ Traefik ─▶ ExternalName svc ─▶ ak-outpost-paperclip-gate (authentik ns)
+                                                   │ forward-auth OK
+                                                   ▼
+                                  paperclip Deployment (ns paperclip, 1 replica)
+                                   ├─ server :3100  (+ agent CLIs as child processes)
+                                   ├─ PVC /paperclip  (ceph-block)
+                                   └─▶ pclp-pg-rw (CNPG, PG17, 2 instances) ─▶ barman ▶ Hetzner S3
+```
+
+## Components
+
+All modeled on `kubernetes/apps/paperless/` unless stated.
+
+### Namespace and baseline policy
+- `Namespace paperclip` in `kubernetes/infrastructure/namespaces/namespaces.yaml`.
+- Baseline default-deny CNP `kubernetes/infrastructure/network-policies/paperclip-ns.yaml`
+  (copy of `paperless-ns.yaml`, renamed).
+- Add `paperclip` to the `openbao` ClusterSecretStore namespace allowlist.
+
+### Flux
+- `kubernetes/clusters/homelab/paperclip.yaml`: Flux `Kustomization`,
+  `path: ./kubernetes/apps/paperclip`, `prune: true`, `wait: true`, same `dependsOn` and
+  CNPG `healthCheckExprs` as `paperless.yaml`.
+- `kubernetes/apps/paperclip/kustomization.yaml`: plain resource list.
+
+### Server Deployment (`paperclip.yaml`)
+- Image `ghcr.io/paperclipai/paperclip:nightly@sha256:<digest>` with the Flux
+  image-automation marker for digest updates.
+- `replicas: 1`, `strategy: Recreate` (single writer on an RWO PVC).
+- Starts as root: upstream `docker-entrypoint.sh` remaps the `node` user, chowns
+  `/paperclip`, then drops to uid 1000 via `gosu`. Caps limited to CHOWN,
+  DAC_OVERRIDE, FOWNER, SETGID, SETUID; `fsGroup: 1000`.
+- `serviceAccountName: paperclip-prometheus-read` with
+  `automountServiceAccountToken: true`. All agents share the pod filesystem, so all
+  of them can read that token — it is mounted anyway because the Role behind it
+  grants exactly `get` on `services/proxy` for
+  `kube-prometheus-stack-prometheus:9090` and
+  `kube-prometheus-stack-alertmanager:9093` in `monitoring`, and nothing else
+  (`kubernetes/apps/paperclip/prometheus-read-rbac.yaml`, REN-36 then REN-95).
+  It was `false` until then.
+  Three things about that grant are worth knowing before you use or debug it:
+  - **It scopes the HTTP method, not the endpoint.** The apiserver maps GET→`get`,
+    POST→`create`, PUT→`update` on a service-proxy request, so `get` alone refuses
+    `POST /-/reload` and `POST /-/quit` in the authorizer. It does *not* narrow
+    which GET endpoints are reachable: the whole read surface of both services is,
+    including `/api/v1/status/config` (scrape credentials there render as
+    `<secret>`) and the web UI.
+  - **Force GET in the caller.** Prometheus accepts both GET and POST on
+    `/api/v1/query`; under this Role only the GET form works, so a client library
+    that switches to POST for long queries will 403 for a reason that has nothing
+    to do with the query.
+  - **If reads start 403ing, check the Service name against `resourceNames`
+    first.** `resourceNames` is matched verbatim, and
+    `kubernetes/apps/monitoring/helmrelease-prometheus.yaml` floats the chart at
+    `>=72.0.0` — a release rename breaks this path silently,
+    with everything still looking correct in git. The port-qualified form is the
+    only one granted: `svc`, `svc:9090` and `http:svc:9090` are three different
+    strings to the authorizer.
+- Port 3100; `Service paperclip` ClusterIP :3100.
+- Probes: startup and readiness on `GET /api/health` with `Host: localhost` (the
+  private-hostname guard rejects the pod IP); liveness is a TCP check, since
+  `/api/health` runs `SELECT 1` and a DB outage must not restart the server.
+- Env:
+  - `DATABASE_URL` assembled from CNPG `pclp-pg-app` Secret (`uri` key).
+  - `HOST=0.0.0.0`, `PORT=3100`
+  - `PAPERCLIP_DEPLOYMENT_MODE=authenticated`
+  - `PAPERCLIP_DEPLOYMENT_EXPOSURE=private`
+  - `PAPERCLIP_PUBLIC_URL=https://paperclip.renner.uno`
+  - `PAPERCLIP_AUTH_DISABLE_SIGN_UP=false` until the first admin is claimed: the
+    claim (`/bootstrap/claim`) requires an already signed-in Better Auth account.
+    A follow-up flips it to `true`.
+  - `PAPERCLIP_TELEMETRY_DISABLED=1` (read by `resolveTelemetryConfig`).
+  - `PAPERCLIP_API_URL=http://paperclip.paperclip.svc.cluster.local:3100` plus
+    `PAPERCLIP_ALLOWED_HOSTNAMES=paperclip.paperclip.svc.cluster.local`: upstream
+    otherwise hands agents the public URL, and the gate (`intercept_header_auth`)
+    rejects their bearer tokens. The Service name also works for sandbox pods under
+    option C; agent JWTs still authenticate each call.
+  - `envFrom: secretRef paperclip-secrets`
+- Reloader: `secret.reloader.stakater.com/reload: paperclip-secrets`.
+
+### Storage
+- `PVC paperclip-data`, `ceph-block`, 20Gi, mounted at `/paperclip` (`PAPERCLIP_HOME`
+  and `HOME`): instance config, workspaces, repo checkouts, per-agent `claude-runtime`
+  dirs.
+
+### Postgres
+- CNPG `Cluster pclp-pg`: 2 instances, `imageName: ghcr.io/cloudnative-pg/postgresql:17`,
+  `ceph-block` 10Gi, `bootstrap.initdb {database: paperclip, owner: paperclip}`,
+  barman-cloud plugin as WAL archiver.
+- `ObjectStore pclp-pg-store` → `s3://renner-uno-k8s-cnpg/pclp-pg`, creds via the
+  existing `hetzner-s3-creds` ExternalSecret pattern.
+- `ScheduledBackup` nightly, `method: plugin`, `backupOwnerReference: self`.
+- `kustomize.toolkit.fluxcd.io/prune: disabled` on Cluster, ObjectStore, ScheduledBackup.
+- PG17 deviates from the cluster-wide PG16 convention because upstream's external
+  Postgres target is `postgres:17` (`docker/docker-compose.yml`) and its embedded default
+  is PG18.
+
+### Secrets
+One ExternalSecret `paperclip-secrets` ← OpenBao `apps/paperclip` (whole-object
+extract):
+- `BETTER_AUTH_SECRET`
+- `PAPERCLIP_TOOL_ACTION_SIGNING_SECRET`
+- `PAPERCLIP_SECRETS_MASTER_KEY` (64-char hex). Held in OpenBao, not on the PVC, so
+  encrypted agent secrets in Postgres survive PVC loss.
+
+### Ingress and SSO
+- Hostname `paperclip.renner.uno`, Traefik `Ingress`, `cert-manager.io/cluster-issuer:
+  letsencrypt-cluster`, **no** `dns.renner.uno/public` label (LAN-only).
+- Authentik gate, per the paperless pattern:
+  - blueprint entries in `kubernetes/apps/authentik/blueprints-configmap.yaml`:
+    proxy provider `paperclip-gate` (`internal_host:
+    http://paperclip.paperclip.svc.cluster.local:3100`, `external_host:
+    https://paperclip.renner.uno`), application, group "Paperclip Users", policy binding;
+  - outpost `ak-outpost-paperclip-gate`;
+  - `kubernetes/apps/authentik/paperclip-gate-egress.yaml` (no `auth.renner.uno`
+    outpost Ingress: like wiki/shelly gates, the outpost serves its own callback on
+    `paperclip.renner.uno`);
+  - `ExternalName` Service in `paperclip` ns pointing at the outpost; the app Ingress
+    targets it on :9000.
+- Paperclip's own email/password login stays on (no header/OIDC login upstream).
+
+### Network policy (`kubernetes/apps/paperclip/network-policy.yaml`)
+Covers the server **and every agent** (same pod under B).
+- Ingress to the app pod: only from the `paperclip-gate` outpost (by label) on 3100.
+- Egress from the app pod:
+  - `pclp-pg` pods :5432
+  - kube-dns :53
+  - `ingress` namespace (Traefik) :8443 (pod-side websecure port) — reaches Gitea, Woodpecker, HA, Loki by their
+    `*.renner.uno` names
+  - `kube-apiserver` entity :6443
+  - `world` TCP :443 (Anthropic API, GitHub, npm)
+- No SSH egress; git over HTTPS only. The namespace baseline
+  (`paperclip-ns.yaml`) also restricts `world` to TCP 443 — Cilium unions allows
+  across CNPs, so an unrestricted baseline would override this list.
+- Known gap (house baseline): pods in `monitoring`/`database` and other pods in
+  `paperclip` can reach :3100 without the gate; Paperclip's own login still applies.
+
+### Image automation
+`kubernetes/infrastructure/image-automation/paperclip.yaml`: `ImageRepository`
+(`ghcr.io/paperclipai/paperclip`) + `ImagePolicy` filtering tag `^nightly$` with
+`digestReflectionPolicy: Always` (image-reflector-controller v1.2.5 in cluster; field
+available since v0.35.0). First use of digest reflection in this repo.
+
+## Guardrails
+- No bot receives cluster-admin or any binding in `flux-system`, `authentik`, `openbao`.
+- Git work goes through PRs; bots get no merge rights on protected branches.
+- Nothing is granted under B that we would not tolerate any agent misusing.
+
+## Bootstrap (one-time)
+1. Place `apps/paperclip` values in OpenBao (the three generated secrets).
+2. Merge; Flux deploys with sign-up enabled.
+3. Sign up through the gate and claim the instance as admin; then set
+   `PAPERCLIP_AUTH_DISABLE_SIGN_UP=true`.
+4. In onboarding "Connect a model", choose Claude subscription and run the shown login
+   **as the node user** (the entrypoint makes exec sessions root, and root-owned 0600
+   credentials are unreadable to the server):
+   `kubectl exec -it -n paperclip deploy/paperclip -- gosu node sh -c 'export CLAUDE_CONFIG_DIR=<shown path> && mkdir -p "$CLAUDE_CONFIG_DIR" && claude auth login'`
+   Repeat after loss of the PVC or database.
+
+## Verification
+- Flux Kustomization `paperclip` Ready; CNPG `pclp-pg` healthy; first ScheduledBackup
+  completes.
+- `curl https://paperclip.renner.uno/api/health` through the gate returns healthy after
+  Authentik login; unauthenticated request redirects to `auth.renner.uno`.
+- Direct pod access from another namespace is denied (CNP).
+- A test agent (claude-local) completes a trivial task using the shared Claude connection.
+- Flux ImagePolicy reports a `nightly@sha256:` digest.
+
+## Deferred
+- **Option C** (sandboxed runs) — required before any per-bot credential isolation.
+- **AIR-Manager** role and all bot identities/credentials.
+- Public exposure.

kubernetes/NETWORK-traffic-classes.md

@@ -1,117 +1,99 @@
 # Traffic classes on the k8s nodes
 
-Issue #576. Three of the four classes are separated; management is not.
+Issue #576. Five classes, each on its own VLAN.
 
 ## Where each class runs
 
 | class | link | mechanism |
 |---|---|---|
 | Ceph replication | `ceph-san`, VLAN 49 | the whole link is handed to the OSD pod (Multus `host-device`) |
-| world→pod | `ingress0`, VLAN 47 | BGP session sourced there, so the advertised next hop is that address |
-| pod→world | `ingress0`, VLAN 47 | routing rule on the pod source prefix → table 47 |
-| pod↔pod | `pod0`, VLAN 48 | routing rule on pod source **and** destination → table 48 |
-| management | `mgmt0`, VLAN 40 | not started |
+| pod↔pod | `pod0`, VLAN 48 | pod-sourced traffic is routed from table 47; bird puts the peers' pod CIDRs there via `pod0` |
+| world↔pod | `ingress0`, VLAN 47 | BGP session sourced there (next hop of every LoadBalancer VIP); pod egress takes table 47's default |
+| node↔node | `node0`, VLAN 46 | the node InternalIP and etcd live there; so do apiserver↔kubelet, cilium-health and host→remote-pod traffic |
+| world↔node | `mgmt0`, VLAN 40 | API VIP for kubectl, talosctl from outside, node egress (DNS, NTP, image pulls) |
 
 Hard constraint, set by the owner: inter-node and pod traffic never traverses
-the UDM. Everything above stays on the 10G switch.
+the UDM. Everything but world↔node stays on the 10G switch. VLAN 46 and 48 are
+statically addressed, without a gateway; VLAN 46 is isolated in UniFi.
 
-## Why policy routing rather than moving the node IP
+## Pod traffic: one table
 
-Cilium reaches a peer's pods at that peer's Kubernetes InternalIP, in native
-and tunnel mode alike. `--ipv4-node` is overwritten from the Node object, and
-`nodePort.directRoutingDevice` selects the BPF NodePort device, not a next hop.
-Both were established by test (#837, #838), not by reading.
+Two rules, one per family, send every packet with a pod source to table 47
+(`talos/patches/site-homelab.yaml`, priorities 1000/1001, below main). The
+table decides by longest prefix; no rule carves out exceptions.
 
-So the InternalIP decides where pod traffic goes — unless the lookup happens in
-a different table. The rules do that, and nothing about the node's identity
-moves. The routing decision runs before Cilium's masquerade hook, so a rule
-matching the pod source sees the pod address, and the masquerade then rewrites
-it to the egress link's.
+bird (`kubernetes/infrastructure/bird/`) owns table 47 and fills it with:
 
-## Rule order
+- the node's attached networks — `mgmt0`, `ingress0`, `pod0`, `node0` —
+  taken from the interfaces;
+- the peers' pod CIDRs, learned over an iBGP mesh on `pod0`, next hop the
+  peer's `pod0` address;
+- the node's own pod CIDR via `cilium_host`, originated rather than read back
+  from the kernel (reading it back made BIRD depend on the shape of Cilium's
+  route, and KRT rejects the IPv4 one);
+- the egress default via the UDM on `ingress0`, from the node annotations
+  `net.renner.uno/egress-gw-v4/v6`.
 
-Lower priority runs first.
+`node0` has to be in the table: a pod reaching a node address — the
+`kubernetes` ClusterIP, which socket LB rewrites to the apiserver's advertise
+address, or a hostNetwork pod — would otherwise take the default to the UDM.
 
-| prio | match | table |
-|---|---|---|
-| 980 | src pod, dst pod | 48 |
-| 981 | src pod, dst pod (v4) | 48 |
-| 990 | dst `fd00::/8` | main |
-| 991 | dst `10.0.0.0/8` | main |
-| 1000 | src pod | 47 |
-| 1001 | src pod (v4) | 47 |
+The routing decision runs before Cilium's masquerade hook, so the rule sees the
+pod address and the masquerade then rewrites it to the egress link's.
 
-980/981 match source **and** destination. Source alone would also catch the
-node's own traffic to a pod, which would then leave with a `pod0` source — an
-address `CiliumNode.spec.addresses` does not list, so Cilium classifies it as
-`world` rather than `host`.
-
-990/991 are Cilium's `ipv4/ipv6NativeRoutingCIDR`, and they are what keeps
-cluster-internal traffic out of table 47. Without them a pod's packet to a
-service matches the source rule, finds only a default route, and leaves for the
-UDM — measured, in-cluster DNS stopped resolving.
-
-Traffic to a pod on the same node falls through on its own: table 48 holds only
-the peers' CIDRs, so the lookup misses and 990/991 send it to main.
-
-## Where each table comes from
-
-- **47** — two default routes declared per node in `talos/patches/k8s-*.yaml`.
-  The IPv6 gateway is the UDM's configured address on the segment, not the
-  RA-learned link-local; `accept_ra_defrtr` stays `0` on `ingress0` so the RA
-  default does not land in main.
-- **48** — filled by the `bird` DaemonSet: an iBGP mesh over `pod0`, one route
-  per peer pod CIDR with the peer's `pod0` address as next hop.
-
-## The bird mesh
-
-`kubernetes/infrastructure/bird/`. Each node advertises only its own pod CIDR
-and installs what it learns into table 48. Cilium's routes in main are
-untouched, which is why the two coexist.
+Node-originated traffic does not match the rules and uses main, where Cilium
+routes each peer's pod CIDR via the peer's InternalIP — on `node0`.
 
 No address is written into this repo twice. Each node publishes its `pod0`
-addresses as `nodeAnnotations`, beside the `VLANConfig` that assigns them; the
-pod CIDRs come from the Node objects. An init container renders both from the
-API with `kubectl -o go-template-file`, excludes the node itself with a field
-selector, and ends with `bird -p` so BIRD's own parser gates startup.
+addresses as `nodeAnnotations` beside the `VLANConfig` that assigns them; the
+pod CIDRs come from the Node objects. bird's init container renders both from
+the API and ends with `bird -p`, so BIRD's own parser gates startup.
 
-The node's own CIDR is **originated**, not read back from the kernel. Reading
-it back made BIRD depend on the shape Cilium installs, and the IPv4 route is
-rejected outright: its next hop sits inside the prefix it describes, which KRT
-refuses with `strange next-hop`. The IPv6 route is a device route and passed,
-so the failure was silent on one family only.
+## Node↔node
+
+`kubelet.nodeIP.validSubnets` and `cluster.etcd.advertisedSubnets` point at
+VLAN 46. Talos picks one InternalIP per family, IPv4 first — the service CIDR
+is IPv4, and the apiserver's advertise address must match it.
+
+The API VIP stays on `mgmt0` for kubectl. No in-cluster client uses it:
+kubelet, scheduler, controller-manager and Cilium go through KubePrism
+(`localhost:7445`), which balances over every address of every control-plane
+member; pods use the ClusterIP.
+
+Talos never updates an existing etcd member's peer URLs
+(siderolabs/talos#12720). Changing `advertisedSubnets` means, per member:
+stage the config, `talosctl etcd forfeit-leadership`, `talosctl etcd leave`,
+reboot. `leave` removes only `/var/lib/etcd`; the node rejoins as a learner
+and is promoted. EPHEMERAL does not need a wipe.
+
+## Moving node IPs: the mixed state
+
+While some nodes have their InternalIP on the new network and some on the old,
+host→remote-pod traffic leaves with a node address that is not an InternalIP:
+the kernel picks the source from the route to the peer's InternalIP. Cilium
+classifies that as `world`, and every ingress policy that admits only
+`host`/`remote-node` (rook-ceph, monitoring, …) drops it. Kernel Ceph clients
+lose the OSDs on the other side; RBD mounts hang.
+
+Bridge it for the duration with a temporary `CiliumClusterwideNetworkPolicy`
+admitting the node subnets of both networks, both families, with
+`enableDefaultDeny: false` so it restricts nothing else, and delete it once
+every node has moved. Keep all OSD nodes schedulable while one is drained:
+draining two takes Ceph below its data availability.
 
 ## Failure behaviour
 
-Benign at every layer. An empty or absent table 48 makes 980/981 fall through
-to main, and pod traffic returns to VLAN 40 — unseparated, not broken. The same
-holds if the DaemonSet cannot start, which is the case during a cold rebuild
-until Gitea serves its image.
-
-## Management, the remaining class
-
-Not started. The levers differ per service and none of them is the node IP:
-
-| service | can leave the node IP? | lever |
-|---|---|---|
-| etcd | yes | `cluster.etcd.advertisedSubnets`; needs a reboot per node |
-| kube-apiserver endpoint | yes | it is a VIP — `Layer2VIPConfig.link` |
-| Talos API | n/a | `apid` answers on every node address; a firewall question |
-| kubelet | no | it *is* the node IP |
-
-Two approaches were rejected by the owner and are not to be proposed again:
-moving the node IP and etcd to `pod0` (#838), and relabelling VLAN 40 as the
-cluster VLAN while moving etcd and the API VIP to a new management VLAN.
-Replacing the CNI (#839, Antrea) was shelved.
-
-Note that the reasoning which produced that list also concluded pod↔pod was
-blocked, which policy routing disproved. Treat it as a record of decisions, not
-as an exhaustive analysis.
+If bird stops, table 47 empties, the rules find nothing and fall through to
+main: pod traffic returns to the node networks, unseparated rather than
+misrouted. The same holds during a cold rebuild until Gitea serves bird's
+image. The default route lives in table 47, not in the Talos config, so a
+default cannot outlive the specific routes and send pod traffic to the UDM.
 
 ## Sources
 
-Cilium v1.20.1 `pkg/datapath/linux/node.go`, `pkg/node/types/node.go`,
-`pkg/node/sync/local_node_sync.go`; BIRD 2.17 user manual (kernel protocol,
-`learn all`, `strict bind`, `direct`); Talos v1.13 `RoutingRuleConfig` and
-`machine.nodeAnnotations`. PRs #831, #833, #835, #836, #837, #842, #852, #854,
-#855; closed unmerged: #830, #838, #839.
+Cilium v1.20 `pkg/datapath/linux/node.go`; BIRD 2.17 user manual (kernel
+protocol, `strict bind`, `direct`); Talos v1.13 `RoutingRuleConfig`,
+`machine.nodeAnnotations`, configuration patching (lists are appended,
+`$patch: delete`), etcd `LeaveCluster`; KubePrism docs;
+siderolabs/talos#12720. PRs #831, #833, #835, #836, #837, #842, #852, #854,
+#855, #857, #858, #918, #919, #922; closed unmerged: #830, #838, #839.

kubernetes/PROPOSAL-wiki-web-api.md

@@ -0,0 +1,315 @@
+# Wiki: local web interface and agent API, as a modular .NET solution
+
+## Goal
+
+A web interface for the Notes wiki that runs entirely in the cluster, and an
+HTTP API through which agents ask the wiki typed questions and deposit facts.
+Built as one C# solution with a MudBlazor front end, and built so that every
+external dependency is replaceable: where the notes live, which model embeds
+them, what judges a typed question, what runs an agent, how the vector index
+is stored. The first adapters are what the homelab uses today. The point of
+the design is that the next one is a project and a config line, not a
+rewrite.
+
+## Where we are
+
+- `wiki-index` (Python, ns `ai`) syncs `wiki/` from `k.renner/notes` by blob
+  SHA, embeds through `wiki-embed` (bge-m3 on llama-server), answers
+  `POST /rank`. It keeps slug, title and vector, not the text. LAN-only behind
+  Traefik BasicAuth; the Gitea webhook hits the ClusterIP.
+- `~/Notes/bin/wiki_rank.py` and `wiki_ask.py` rank through the index, build a
+  document context from a local vault checkout, send it with typed questions
+  to TypeSafe, gate on `covered`. The wiki-synth agent composes answers from
+  those judgements.
+- New knowledge enters as a file in `raw/` (facts with provenance, never
+  hypotheses), distilled nightly in CI.
+- The house C# project, shelly-manager, sets the conventions: net10.0, an
+  `.slnx` with `src/` and `test/`, xUnit, MudBlazor 9, Blazor Web App with
+  interactive server rendering, minimal APIs, an Authentik proxy-header
+  authentication handler for humans and an API-key handler for machines, a
+  multi-stage Dockerfile, `dotnet test` on every push and a kaniko release
+  pipeline to the Gitea registry. Human-facing apps sit behind an Authentik
+  outpost (`<app>-gate`) with a CNP that admits only that outpost.
+
+## Architecture: ports in the core, adapters around it
+
+```
+                 ┌──────────────────────────────────────────┐
+                 │  Wiki.Host  (ASP.NET Core + Blazor/Mud)  │  composition root,
+                 │  /api/v1, /rank, /webhook, UI pages      │  config, auth
+                 └───────────────┬──────────────────────────┘
+                                 │ uses
+                 ┌───────────────▼──────────────────────────┐
+                 │  Wiki.Application  (use cases)           │  Sync, Rank, Ask,
+                 │                                          │  DepositFact, Queue-
+                 └───────────────┬──────────────────────────┘  Question, Evaluate
+                                 │ depends only on
+                 ┌───────────────▼──────────────────────────┐
+                 │  Wiki.Core  (domain + ports)             │  Note, RawSource,
+                 │  no package references                   │  Report, LinkGraph,
+                 └──────────────────────────────────────────┘  the interfaces
+   implemented by ▲            ▲            ▲            ▲            ▲
+   Adapters.Gitea  Adapters.LlamaCpp  Adapters.TypeSafe  Adapters.FileIndex  Adapters.Anthropic
+   (later: Database, Ollama/Onnx, local Jev, pgvector/Qdrant, OpenAI-compatible answer models)
+```
+
+Rules that keep it modular, enforced by project references and a test:
+
+- `Wiki.Core` references nothing. It holds the domain types and the port
+  interfaces. Adapters reference Core only; they never reference each other
+  or the host. `Wiki.Host` is the only project that knows which adapters
+  exist.
+- Every port has a **contract test suite** in `test/Wiki.Contracts` that
+  runs against every adapter of that port. A new adapter is done when the
+  suite is green. That is what makes "swap the storage" a safe operation.
+- Adapters are selected by configuration and registered by name (keyed
+  services), so several can be active at once. The UI's lab page and the
+  evaluator use that to compare, the API uses the one marked active.
+- Nothing in Core or Application knows about HTTP, Gitea, TypeSafe, files or
+  Kubernetes.
+
+### Ports
+
+| port | what it abstracts | first adapter | examples of the next ones |
+|---|---|---|---|
+| `INoteStore` | read notes, raw sources and reports at a head; create a new document (a fact becomes a raw note in the application layer, the store only moves text) | Gitea repository API (`k.renner/notes`, blob SHA) | git checkout on a volume; a database (CNPG Postgres) with the vault imported; an Obsidian sync folder |
+| `IEmbeddingModel` | text → vector, with a model id and dimension | llama-server `/v1/embeddings` (bge-m3 via wiki-embed, the model that won the German-document comparison on accuracy per cost) | the qwen3 embedding variants from that comparison; ONNX in-process; any OpenAI-compatible endpoint |
+| `IVectorIndex` | store vectors per (model id, document), query top-k | in-memory cosine with a file snapshot per model, as today | pgvector in the database; Qdrant |
+| `IChunker` | split a document into addressable units for typed questions | paragraph splitter (headings, fences), as in `wiki_ask.py` | sentence windows; semantic chunks |
+| `IJudge` | typed questions over a state: Choice, Noul, Score, with probabilities | TypeSafe System One (`jev-latest`, cloud; the wiki records no locally hosted Jev today) | a locally hosted Jev with the same request shape, should TypeSafe offer one; a constrained-output local LLM adapter, flagged as uncalibrated |
+| `IAnswerModel` | prose from a question plus evidence: the selected paragraphs with their notes and the judgements; must cite by slug; streaming | Anthropic Messages API (Claude, model from config, Sonnet by default) | any OpenAI-compatible endpoint, including the local llama-server; a different Claude model per question |
+| `IEvaluator` | measure a retrieval or judge configuration against the labelled set (questions and cited sources in `output/`) | recall@k, hit rate, MRR, as `wiki_rank_eval.py` | judge accuracy against hand-labelled paragraphs |
+| `IIdentity` | who is calling, for attribution of facts and saved answers | Authentik proxy headers (humans), API key (machines), as in shelly-manager | OIDC bearer tokens |
+
+Two ports deserve a sentence of scope each. The judge returns judgements
+with probabilities, never prose, and the UI never presents its output as an
+answer, only as ranked, quoted evidence. The answer model turns that
+evidence into words and is allowed nothing else: it sees the selected
+paragraphs and the judgements, not the whole context, and its output is
+shown next to the evidence it was given. Both are single model calls inside
+the request. The service runs no agent: no tool loop, no vault edits. The
+wiki-synth agent in a Claude session remains the path for artifacts and for
+multi-question synthesis, and it can call `/questions` itself.
+
+### Solution layout
+
+```
+Wiki.slnx
+src/
+  Wiki.Core/               domain types, ports, frontmatter and link parsing
+  Wiki.Application/        use cases; depends on Core only
+  Wiki.Adapters.Gitea/     INoteStore (read at a head, create)
+  Wiki.Adapters.LlamaCpp/  IEmbeddingModel (OpenAI-compatible embeddings)
+  Wiki.Adapters.TypeSafe/  IJudge (System One client)
+  Wiki.Adapters.FileIndex/ IVectorIndex (memory + snapshot per model)
+  Wiki.Adapters.Anthropic/ IAnswerModel (Messages API, streaming)
+  Wiki.Host/               ASP.NET Core: API, webhook, MudBlazor UI, auth,
+                           composition root, appsettings
+test/
+  Wiki.Core.Tests/         parsing, link graph, chunker
+  Wiki.Contracts/          one suite per port, parameterised over adapters
+  Wiki.Application.Tests/  use cases with fake ports
+  Wiki.Host.Tests/         API surface, auth handlers
+```
+
+Own repository `k.renner/wiki`, same shape as shelly-manager: `Dockerfile`,
+`.woodpecker/test.yml` on pull requests, `.woodpecker/release.yml` building
+the image with kaniko into the Gitea registry on push to `main`, tagged
+`build.<unix-ts>` plus the commit SHA so Flux image automation can follow
+it. The homelab repo holds only the manifests. `claude-bot` needs write
+access to the new repository.
+
+Two things the shelly-manager history settles in advance: the TypeSafe
+client is written from the documented field contract (endpoint, bearer key,
+`jev-latest`, Noul/Choice/Score, 255-option cap, 32k-token state budget),
+there is no prior C# client to copy; and the UI needs no login page, because
+identity comes from the outpost's headers, which avoids the Blazor-circuit
+sign-in constraint shelly-manager ran into.
+
+### Configuration
+
+Adapters are declared in `appsettings` and selected by name; secrets come
+through environment variables, filled by ExternalSecret, never from the file.
+
+```json
+"Wiki": {
+  "NoteStore": { "Active": "gitea", "gitea": { "Repository": "k.renner/notes", "Branch": "main" } },
+  "Embeddings": {
+    "Active": "bge-m3",
+    "Models": {
+      "bge-m3":   { "Adapter": "LlamaCpp", "Url": "http://wiki-embed.ai.svc:8080", "Model": "bge-m3" },
+      "e5-large": { "Adapter": "Ollama",   "Url": "http://ollama.ai.svc:11434",    "Model": "multilingual-e5-large" }
+    }
+  },
+  "Judges": {
+    "Active": "typesafe",
+    "Models": {
+      "typesafe":  { "Adapter": "TypeSafe", "Url": "https://api.typesafe.ai/v1/systemone", "Model": "jev-latest" },
+      "jev-local": { "Adapter": "TypeSafe", "Url": "http://jev.ai.svc:8080/v1/systemone",  "Model": "jev" }   // illustration: same adapter, other endpoint
+    }
+  },
+  "Index": { "Active": "file", "file": { "Path": "/data/index" } },
+  "AnswerModels": {
+    "Active": "claude-sonnet",
+    "Models": {
+      "claude-sonnet": { "Adapter": "Anthropic", "Model": "claude-sonnet-5" },
+      "qwen-local":    { "Adapter": "OpenAI",    "Url": "http://llama-server.ai.svc:8080/v1", "Model": "qwen3.6-35b-a3b" }
+    }
+  }
+}
+```
+
+Every embedding model gets its own index namespace. Switching the active
+model is a config change after the new index has been built in the
+background; until then both exist and can be compared. Nothing is
+re-embedded in place.
+
+## The web interface (MudBlazor)
+
+Blazor Web App, interactive server rendering, hosted in `Wiki.Host`. Static
+assets vendored into the image, no CDN. Pages:
+
+1. **Search**: a query box, results ranked by the active index with score and
+   tags; for a phrased question, optionally the paragraph-level hits from the
+   judge with their probabilities, shown as quoted evidence.
+2. **Note**: frontmatter, rendered body with `[[wikilinks]]` resolved,
+   backlinks, sources linking to their raw file.
+3. **Browse**: by tag, by `updated`, recently changed, orphans.
+4. **Answers**: a question box. The answer streams in within seconds: rank,
+   judge, compose, all in the request. Below it the evidence: the paragraphs
+   the judge selected with their notes and probabilities, and the `covered`
+   gate; a low gate is shown as "the wiki does not cover this" instead of an
+   answer. A save button writes the answer to `output/` with the wiki-synth
+   frontmatter (`question`, `sources`, `confidence`), so it joins the
+   labelled set. Underneath, the saved answers, newest first, including the
+   ones the wiki-synth agent wrote from a session.
+5. **Maintenance**: the raw queue with ages, the last consolidation and audit
+   reports, index and sync status.
+6. **Lab**: pick any registered embedding model and judge, run the same query
+   through them side by side, and run the evaluator on the labelled set to
+   get recall@k per configuration. This is the page for "play around with
+   different models".
+7. **Settings**: read-only view of the active adapters and their health.
+
+No editing of notes in the browser. The vault stays a git repository edited
+by the consolidate agent and by hand.
+
+## API v1
+
+Reads are authenticated by the hostname's gate (humans) or API key
+(machines); writes always need an API key so a fact carries its author.
+
+| method and path | purpose |
+|---|---|
+| `GET /api/v1/status` | index commit and age, counts, active adapters |
+| `GET /api/v1/notes?q=&tag=&limit=` | list; `q` ranks through the active index |
+| `GET /api/v1/notes/{slug}` | frontmatter, body, links, backlinks, sources |
+| `GET /api/v1/tags` | tags with counts |
+| `POST /api/v1/rank` | `{question, k, model?}` → ranked slugs; also served at `/rank` for the existing scripts |
+| `POST /api/v1/ask` | `{queries, questions, k, judge?, model?}` → context, typed judgements with the answering paragraphs verbatim, `covered`. For agents and code. |
+| `POST /api/v1/questions` | `{question, k?, judge?, model?, answerModel?, save?}` → the answer text with `[[slug]]` citations, the evidence (paragraphs, probabilities), `covered`, model and token usage; streamed as server-sent events when requested; with `save`, the `output/` path and commit |
+| `POST /api/v1/facts` | deposit a raw source: title, source, `kind: fact`, observed_at, content, tags → path and commit |
+| `GET /api/v1/raw` | the distill queue with ages |
+| `GET /api/v1/reports` | `output/` documents by type, newest first |
+| `POST /api/v1/lab/eval` | run the evaluator for a (model, judge) pair |
+| `POST /webhook` | Gitea push, HMAC as today |
+
+`POST /api/v1/facts` is rejected without `source`, with a `kind` other than
+`fact`, or with content under a minimum length. The raw rule is enforced by
+shape: there is no field for a hypothesis. The file gets the vault's
+provenance frontmatter with `captured_by` set from the API key's identity.
+
+`POST /api/v1/ask` is `wiki_ask.py` moved server-side, so the scripts become
+thin clients and agents no longer need a vault checkout. Its result is
+judgements and quotations, never prose.
+
+`POST /api/v1/questions` is `/ask` plus one answer-model call, in process:
+rank the question, build the context, run the default typed questions
+(answered, where, covered), hand the selected paragraphs and judgements to
+the answer model, stream the prose back. Below the `covered` threshold it
+returns the gap instead of an answer. Time is the sum of the parts, about
+0.2 s rank, 1 to 4 s judge, a few seconds of generation. Cost is one model
+call over a few thousand tokens, a few cents on Sonnet, against a quarter
+dollar for a full agent run.
+
+## Auth and network
+
+- Humans: `wiki.renner.uno`, LAN-only (no `dns.renner.uno/public` label),
+  Ingress backend `ak-outpost-wiki-gate:9000`, a `wiki-gate-egress` CNP in
+  `authentik`, the proxy-header handler trusts `X-Authentik-*` only from that
+  outpost.
+- Machines: `wiki-index.renner.uno` with the API-key handler; the workstation
+  key stays where it is. CI reaches the ClusterIP.
+- Secrets (Gitea token, webhook secret, API keys, TypeSafe key, Anthropic
+  key) reach the
+  pod as environment variables from Kubernetes Secrets filled by
+  ExternalSecret from the `openbao` ClusterSecretStore. The pod itself has
+  no OpenBao access. The Python service fetched its secrets from OpenBao
+  directly through its own Kubernetes-auth role; that role, its policy and
+  the `ai` → `openbao` CNP entry are retired with it.
+- Egress from the pod: Gitea, wiki-embed, `api.typesafe.ai:443`,
+  `api.anthropic.com:443`. Every
+  further adapter that talks to a new endpoint gets its own CNP line, which
+  is the one place adding an adapter touches the cluster.
+
+## Health endpoints
+
+Three endpoints with three different meanings, wired to the .NET health
+checks with tags so that the probes and the humans see the right thing:
+
+| endpoint | meaning | probe |
+|---|---|---|
+| `GET /health` | the process serves HTTP. Always 200 once Kestrel listens; no dependency checks. | liveness |
+| `GET /ready` | the active note store has synced at least once and the active index is loaded; 503 while the first sync or index build runs. A failed later sync keeps the last good state and stays ready. | readiness, plus a startup probe with enough budget for the first sync (the Python service was killed mid-build when its `/health` returned 503 during the initial sync) |
+| `GET /api/v1/status` | the full picture for humans and the Settings page: index commit and age, counts, queue lengths, and one entry per registered adapter with its last check result and latency. | none |
+
+Adapter checks (Gitea reachable, embeddings endpoint answering, judge
+endpoint answering, answer model answering) are reported in `/status` and
+on the Settings page and are never part of `/ready`: a judge outage must not
+take the reader down. Their failures show up as red entries, not as
+restarts.
+
+## Phases
+
+One PR each, verified before the next. The Python `wiki-index` keeps running
+until phase 3 proves parity.
+
+0. **Skeleton.** Repository, solution, `Wiki.Core` with the ports and domain
+   types, `Wiki.Contracts` with the suites, `Wiki.Host` serving `/health`,
+   `/ready` and `/api/v1/status`. CI: test and release pipelines. Verify:
+   image in the registry, pipeline green, the Deployment's three probes
+   wired and the pod Ready.
+1. **Parity with the Python service.** Gitea, LlamaCpp and FileIndex
+   adapters, the sync use case, `/rank` and `/webhook`. Verify: the 40-query
+   recall eval gives the same numbers as the Python service on the same
+   commit; `wiki_rank.py --probe-index` against the new service passes.
+2. **Read API and UI.** Notes, tags, raw, reports; Search, Note, Browse,
+   Maintenance pages; `wiki-gate`. Verify: a note renders with resolved
+   links; an unauthenticated request is redirected to Authentik.
+3. **Ask.** Chunker, TypeSafe adapter, `/api/v1/ask`, scripts as clients.
+   Verify: `test_wiki_ask.py` passes against the API; a wiki-synth run with
+   the definition unchanged. Then the Python service is retired and the
+   `wiki-index` hostname points at the new one.
+4. **Facts.** API keys as an ExternalSecret, the deposit-fact use case, `/api/v1/facts`. Verify:
+   a fact posted from the workstation shows as a `claude-bot` commit and in
+   the next distill.
+5. **Questions.** Anthropic adapter, `/api/v1/questions` with streaming,
+   the Answers page with save. Verify: a question answered in under ten
+   seconds with citations that resolve to the shown evidence; a question the
+   wiki does not cover returns the gap; a saved answer appears in `output/`
+   with the wiki-synth frontmatter and is picked up by the evaluator.
+6. **Lab.** Evaluator, a second embedding model and a second answer model
+   registered, the Lab page. Verify: two models compared on the labelled set
+   from the UI.
+
+## Not in scope
+
+- Editing notes in the browser.
+- Replacing Obsidian; the vault stays a git repository and `notes-push.path`
+  keeps pushing.
+- Public exposure; LAN and VPN only.
+- Running or triggering agents. The service makes single model calls inside
+  a request; it never runs a tool loop or edits the vault. Distill,
+  consolidate and multi-question synthesis stay Claude processes in CI or in
+  a session; the service shows their inputs and outputs and serves their API
+  calls.

kubernetes/RUNBOOK-cold-rebuild.md

@@ -107,7 +107,7 @@ config, and the nodes are addressed by name.
 ## Step 1b — boot all three machines from install media
 
 ```
-https://factory.talos.dev/image/0f6fb8bd8bb53013e9743b7c7b099d96fb9c32fd2464ced32693c2dc0ea5feb9/v1.13.9/metal-amd64.iso
+https://factory.talos.dev/image/0f6fb8bd8bb53013e9743b7c7b099d96fb9c32fd2464ced32693c2dc0ea5feb9/v1.13.10/metal-amd64.iso
 ```
 
 Write it to a USB stick and boot each MS-01 from it. Each comes up in

kubernetes/RUNBOOK-home-assistant-image-rollback.md

@@ -0,0 +1,71 @@
+# Runbook — roll back a Home Assistant image bump
+
+Use when an automatic Home Assistant update (a
+`chore(image-automation): bump ghcr.io/home-assistant/home-assistant:<new>`
+commit on `master`) breaks Home Assistant. The owner approves the rollback
+and its timing. Atlas writes the PR; Argus merges it.
+
+## Why a plain `git revert` is not enough
+
+Image automation (`kubernetes/infrastructure/image-automation/`) re-scans
+the registry every hour and rewrites the tag every 30 minutes to the
+highest release matching the `home-assistant` ImagePolicy. A plain revert
+of the bump commit would be bumped forward again within about an hour.
+
+So the rollback commit does two things at once: it reverts the tag **and**
+drops the `# {"$imagepolicy": ...}` marker from that image line. Without
+the marker the Setters strategy has nothing to rewrite, so the old tag
+stays pinned. It is still one commit, and reverting that commit restores
+automation.
+
+## Before you roll back
+
+1. Find the bump commit:
+   ```sh
+   git log --oneline -5 origin/master -- kubernetes/apps/home-assistant/deployment.yaml
+   ```
+2. Check the release notes of the broken version for a **database schema**
+   change. Home Assistant migrates the recorder schema (CNPG cluster
+   `ha-pg`) on start and cannot downgrade it. An older version started on a
+   newer schema runs without recorder history until the schema matches
+   again. If the bump crossed a schema change, tell the owner before going
+   on: the choice is roll back without history, or restore the recorder
+   database from backup as a separate step.
+
+## Roll back (one commit, one PR)
+
+```sh
+git worktree add -b atlas/ha-rollback-<old> "$WT" origin/master && cd "$WT"
+git revert --no-commit <bump-sha>
+# In kubernetes/apps/home-assistant/deployment.yaml, delete the trailing
+#   # {"$imagepolicy": "flux-system:home-assistant"}
+# from the home-assistant image line. Leave every other line alone.
+git diff --cached --stat   # expect: 1 file, 1 line changed
+git commit -m "fix(home-assistant): pin <old>, pause image automation (bump to <new> broke HA)"
+git push -u origin HEAD
+```
+
+The PR description states: broken version, pinned version, symptom, schema
+check result, and "revert this commit to resume automatic updates".
+HA-impact review and merge follow [README.md](../README.md#2-ha-impact-review-before-merge).
+
+On merge Flux applies the Deployment (strategy `Recreate`): Home Assistant
+restarts once, about one to two minutes of downtime.
+
+## Verify
+
+```sh
+kubectl -n home-assistant get deploy home-assistant \
+  -o jsonpath='{.spec.template.spec.containers[?(@.name=="home-assistant")].image}'
+kubectl -n home-assistant rollout status deploy/home-assistant
+```
+
+Read-only commands only (see REN-240 protection rule 1). One hour after
+merge, confirm no new `chore(image-automation)` commit touched
+`deployment.yaml`.
+
+## Resume automatic updates
+
+When a fixed Home Assistant release is out, revert the rollback commit. The
+marker comes back, and image automation bumps to the newest release on its
+next pass. That bump is itself a Home Assistant restart.

kubernetes/apps/audiobookshelf/deployment.yaml

@@ -122,7 +122,7 @@ spec:
             claimName: audiobookshelf-metadata
         - name: books
           persistentVolumeClaim:
-            claimName: audiobookshelf-books
+            claimName: audiobookshelf-books-bulk
         - name: le-ca
           emptyDir:
             medium: Memory

kubernetes/apps/audiobookshelf/pvcs.yaml

@@ -35,14 +35,14 @@ spec:
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:
-  name: audiobookshelf-books
+  name: audiobookshelf-books-bulk
   namespace: audiobookshelf
   annotations:
     kustomize.toolkit.fluxcd.io/prune: disabled
 spec:
   accessModes:
     - ReadWriteMany
-  storageClassName: ceph-filesystem
+  storageClassName: ceph-filesystem-bulk
   resources:
     requests:
       storage: 300Gi

kubernetes/apps/authentik/blueprints-configmap.yaml

@@ -976,6 +976,280 @@ data:
             kubernetes_ingress_secret_name: "authentik-outpost-tls"
             kubernetes_json_patches: null
 
+      # wiki-gate: the wiki service (ns ai) behind the same passkey flows as
+      # llama-gate. The service trusts X-Authentik-Username only from pod
+      # addresses, and its CNP admits only this outpost, so the gate is the
+      # whole access control for wiki.renner.uno.
+      - id: wiki-gate-provider
+        model: authentik_providers_proxy.proxyprovider
+        identifiers:
+          name: wiki-gate
+        state: present
+        attrs:
+          authentication_flow: !Find [authentik_flows.flow, [slug, passkey-passwordless-authentication]]
+          authorization_flow: !Find [authentik_flows.flow, [slug, passkey-stepup-authorization]]
+          invalidation_flow: !Find [authentik_flows.flow, [slug, default-provider-invalidation-flow]]
+          mode: proxy
+          internal_host: http://wiki.ai.svc.cluster.local:8080
+          internal_host_ssl_validation: false
+          external_host: https://wiki.renner.uno
+          cookie_domain: renner.uno
+          access_token_validity: hours=24
+          refresh_token_validity: days=30
+          intercept_header_auth: true
+
+      - model: authentik_core.application
+        identifiers:
+          slug: wiki-gate
+        state: present
+        attrs:
+          name: Wiki
+          slug: wiki-gate
+          provider: !KeyOf wiki-gate-provider
+          meta_launch_url: https://wiki.renner.uno
+          meta_description: "Notes wiki: search, notes, answers, maintenance"
+
+      - model: authentik_outposts.outpost
+        identifiers:
+          name: wiki-gate
+        state: present
+        attrs:
+          type: proxy
+          service_connection: !KeyOf k8s-conn
+          providers:
+            - !KeyOf wiki-gate-provider
+          config:
+            log_level: info
+            authentik_host: "http://authentik-server.authentik.svc.cluster.local"
+            authentik_host_insecure: false
+            authentik_host_browser: "https://auth.renner.uno/"
+            refresh_interval: "minutes=5"
+            container_image: null
+            object_naming_template: "ak-outpost-%(name)s"
+            kubernetes_namespace: authentik
+            kubernetes_replicas: 1
+            kubernetes_service_type: ClusterIP
+            kubernetes_disabled_components: ["ingress"]
+            kubernetes_image_pull_secrets: []
+            kubernetes_ingress_annotations: {}
+            kubernetes_ingress_class_name: null
+            kubernetes_ingress_secret_name: "authentik-outpost-tls"
+            kubernetes_json_patches: null
+
+      # paperclip-gate: Paperclip UI; the CNP admits only this outpost. Paperclip keeps its own login behind it.
+      - id: paperclip-users-group
+        model: authentik_core.group
+        identifiers:
+          name: Paperclip Users
+        state: present
+        attrs:
+          is_superuser: false
+          users:
+            - !Find [authentik_core.user, [username, k.renner]]
+
+      - id: paperclip-gate-provider
+        model: authentik_providers_proxy.proxyprovider
+        identifiers:
+          name: paperclip-gate
+        state: present
+        attrs:
+          authentication_flow: !Find [authentik_flows.flow, [slug, passkey-passwordless-authentication]]
+          authorization_flow: !Find [authentik_flows.flow, [slug, passkey-stepup-authorization]]
+          invalidation_flow: !Find [authentik_flows.flow, [slug, default-provider-invalidation-flow]]
+          mode: proxy
+          internal_host: http://paperclip.paperclip.svc.cluster.local:3100
+          internal_host_ssl_validation: false
+          external_host: https://paperclip.renner.uno
+          cookie_domain: renner.uno
+          access_token_validity: hours=24
+          refresh_token_validity: days=30
+          intercept_header_auth: true
+
+      - model: authentik_core.application
+        identifiers:
+          slug: paperclip-gate
+        state: present
+        attrs:
+          name: Paperclip
+          slug: paperclip-gate
+          provider: !KeyOf paperclip-gate-provider
+          meta_launch_url: https://paperclip.renner.uno
+          meta_description: "Paperclip agent orchestration"
+
+      - model: authentik_policies.policybinding
+        identifiers:
+          target: !Find [authentik_core.application, [slug, paperclip-gate]]
+          group: !KeyOf paperclip-users-group
+          order: 0
+        state: present
+        attrs:
+          enabled: true
+          negate: false
+
+      - model: authentik_outposts.outpost
+        identifiers:
+          name: paperclip-gate
+        state: present
+        attrs:
+          type: proxy
+          service_connection: !KeyOf k8s-conn
+          providers:
+            - !KeyOf paperclip-gate-provider
+          config:
+            log_level: info
+            authentik_host: "http://authentik-server.authentik.svc.cluster.local"
+            authentik_host_insecure: false
+            authentik_host_browser: "https://auth.renner.uno/"
+            refresh_interval: "minutes=5"
+            container_image: null
+            object_naming_template: "ak-outpost-%(name)s"
+            kubernetes_namespace: authentik
+            kubernetes_replicas: 1
+            kubernetes_service_type: ClusterIP
+            kubernetes_disabled_components: ["ingress"]
+            kubernetes_image_pull_secrets: []
+            kubernetes_ingress_annotations: {}
+            kubernetes_ingress_class_name: null
+            kubernetes_ingress_secret_name: "authentik-outpost-tls"
+            kubernetes_json_patches: null
+
+      # navidrome-gate: the Navidrome web UI. Navidrome auto-creates the
+      # X-Authentik-Username user on first login; the first user is admin.
+      # /rest and /share bypass the gate at the Traefik router.
+      - id: navidrome-gate-provider
+        model: authentik_providers_proxy.proxyprovider
+        identifiers:
+          name: navidrome-gate
+        state: present
+        attrs:
+          authorization_flow: !Find [authentik_flows.flow, [slug, default-provider-authorization-implicit-consent]]
+          invalidation_flow: !Find [authentik_flows.flow, [slug, default-provider-invalidation-flow]]
+          mode: proxy
+          internal_host: http://navidrome.navidrome.svc.cluster.local
+          internal_host_ssl_validation: false
+          external_host: https://navidrome.renner.uno
+          cookie_domain: renner.uno
+          access_token_validity: hours=24
+          refresh_token_validity: days=30
+          intercept_header_auth: true
+
+      - model: authentik_core.application
+        identifiers:
+          slug: navidrome-gate
+        state: present
+        attrs:
+          name: Navidrome
+          slug: navidrome-gate
+          provider: !KeyOf navidrome-gate-provider
+          meta_launch_url: https://navidrome.renner.uno
+          meta_description: "Music streaming (Subsonic)"
+
+      - model: authentik_outposts.outpost
+        identifiers:
+          name: navidrome-gate
+        state: present
+        attrs:
+          type: proxy
+          service_connection: !KeyOf k8s-conn
+          providers:
+            - !KeyOf navidrome-gate-provider
+          config:
+            log_level: info
+            authentik_host: "http://authentik-server.authentik.svc.cluster.local"
+            authentik_host_insecure: false
+            authentik_host_browser: "https://auth.renner.uno/"
+            refresh_interval: "minutes=5"
+            container_image: null
+            object_naming_template: "ak-outpost-%(name)s"
+            kubernetes_namespace: authentik
+            kubernetes_replicas: 1
+            kubernetes_service_type: ClusterIP
+            kubernetes_disabled_components: ["ingress"]
+            kubernetes_image_pull_secrets: []
+            kubernetes_ingress_annotations: {}
+            kubernetes_ingress_class_name: null
+            kubernetes_ingress_secret_name: "authentik-outpost-tls"
+            kubernetes_json_patches: null
+
+      # lidarr-gate: Lidarr runs with AuthenticationMethod=External, so the gate
+      # is its only access control.
+      - id: music-librarians-group
+        model: authentik_core.group
+        identifiers:
+          name: Music Librarians
+        state: present
+        attrs:
+          is_superuser: false
+          # Membership is owned here; blueprint apply reverts UI edits.
+          users:
+            - !Find [authentik_core.user, [username, k.renner]]
+
+      - id: lidarr-gate-provider
+        model: authentik_providers_proxy.proxyprovider
+        identifiers:
+          name: lidarr-gate
+        state: present
+        attrs:
+          authorization_flow: !Find [authentik_flows.flow, [slug, default-provider-authorization-implicit-consent]]
+          invalidation_flow: !Find [authentik_flows.flow, [slug, default-provider-invalidation-flow]]
+          mode: proxy
+          internal_host: http://lidarr.navidrome.svc.cluster.local
+          internal_host_ssl_validation: false
+          external_host: https://lidarr.renner.uno
+          cookie_domain: renner.uno
+          access_token_validity: hours=24
+          refresh_token_validity: days=30
+          intercept_header_auth: true
+
+      - model: authentik_core.application
+        identifiers:
+          slug: lidarr-gate
+        state: present
+        attrs:
+          name: Lidarr
+          slug: lidarr-gate
+          provider: !KeyOf lidarr-gate-provider
+          meta_launch_url: https://lidarr.renner.uno
+          meta_description: "Music library gaps: discographies, Spotify artists"
+
+      - model: authentik_policies.policybinding
+        identifiers:
+          target: !Find [authentik_core.application, [slug, lidarr-gate]]
+          group: !KeyOf music-librarians-group
+          order: 0
+        state: present
+        attrs:
+          enabled: true
+          negate: false
+          failure_result: false
+
+      - model: authentik_outposts.outpost
+        identifiers:
+          name: lidarr-gate
+        state: present
+        attrs:
+          type: proxy
+          service_connection: !KeyOf k8s-conn
+          providers:
+            - !KeyOf lidarr-gate-provider
+          config:
+            log_level: info
+            authentik_host: "http://authentik-server.authentik.svc.cluster.local"
+            authentik_host_insecure: false
+            authentik_host_browser: "https://auth.renner.uno/"
+            refresh_interval: "minutes=5"
+            container_image: null
+            object_naming_template: "ak-outpost-%(name)s"
+            kubernetes_namespace: authentik
+            kubernetes_replicas: 1
+            kubernetes_service_type: ClusterIP
+            kubernetes_disabled_components: ["ingress"]
+            kubernetes_image_pull_secrets: []
+            kubernetes_ingress_annotations: {}
+            kubernetes_ingress_class_name: null
+            kubernetes_ingress_secret_name: "authentik-outpost-tls"
+            kubernetes_json_patches: null
+
       - model: authentik_outposts.outpost
         identifiers:
           name: authentik Embedded Outpost

kubernetes/apps/authentik/kustomization.yaml

@@ -7,6 +7,7 @@ resources:
   - helmrelease.yaml
   - paperless-gate-ingress.yaml
   - paperless-gate-egress.yaml
+  - paperclip-gate-egress.yaml
   - vscode-gate-egress.yaml
   - shelly-manager-gate-egress.yaml
   - esphome-gate-egress.yaml
@@ -16,6 +17,9 @@ resources:
   - prometheus-gate-egress.yaml
   - wordpress-gate-egress.yaml
   - llama-gate-egress.yaml
+  - wiki-gate-egress.yaml
+  - navidrome-gate-egress.yaml
+  - lidarr-gate-egress.yaml
   - ldap-outpost-cert.yaml
   - ldap-outpost-lb.yaml
   - ldap-outpost-cnp.yaml

kubernetes/apps/authentik/lidarr-gate-egress.yaml

@@ -0,0 +1,20 @@
+# Lets the lidarr-gate outpost reach Lidarr; the matching ingress rule is in
+# network-policies/navidrome-ns.yaml.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: lidarr-gate-egress
+  namespace: authentik
+spec:
+  endpointSelector:
+    matchLabels:
+      goauthentik.io/outpost-name: lidarr-gate
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": navidrome
+            app.kubernetes.io/name: lidarr
+      toPorts:
+        - ports:
+            - port: "8686"
+              protocol: TCP

kubernetes/apps/authentik/navidrome-gate-egress.yaml

@@ -0,0 +1,20 @@
+# Lets the navidrome-gate outpost reach Navidrome; the matching ingress rule
+# is in network-policies/navidrome-ns.yaml.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: navidrome-gate-egress
+  namespace: authentik
+spec:
+  endpointSelector:
+    matchLabels:
+      goauthentik.io/outpost-name: navidrome-gate
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": navidrome
+            app.kubernetes.io/name: navidrome
+      toPorts:
+        - ports:
+            - port: "4533"
+              protocol: TCP

kubernetes/apps/authentik/paperclip-gate-egress.yaml

@@ -0,0 +1,19 @@
+# Lets the paperclip-gate outpost reach the Paperclip server; scoped to this outpost's pods.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: paperclip-gate-egress
+  namespace: authentik
+spec:
+  endpointSelector:
+    matchLabels:
+      goauthentik.io/outpost-name: paperclip-gate
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": paperclip
+            app.kubernetes.io/name: paperclip
+      toPorts:
+        - ports:
+            - port: "3100"
+              protocol: TCP

kubernetes/apps/authentik/wiki-gate-egress.yaml

@@ -0,0 +1,21 @@
+# Lets the wiki-gate outpost reach the wiki service. The authentik namespace
+# baseline CNP otherwise stops any authentik pod from entering ai. Scoped to
+# this outpost's pod label; the matching ingress rule is in ai-ns.yaml.
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: wiki-gate-egress
+  namespace: authentik
+spec:
+  endpointSelector:
+    matchLabels:
+      goauthentik.io/outpost-name: wiki-gate
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ai
+            app.kubernetes.io/name: wiki
+      toPorts:
+        - ports:
+            - port: "8080"
+              protocol: TCP

kubernetes/apps/gatus/configmap.yaml

@@ -30,7 +30,38 @@ metadata:
 #     Web Service 9 = Music Assistant
 #     Web Service 10 = TRMNL
 #     Web Service 11 = Paperless (gate-level only, see its entry)
+#
+# CORRECTION (REN-21, 2026-09-25) to the Web Service 11 comment further down,
+# which claims that an in-cluster probe of the paperless ClusterIP "needs a
+# CiliumNetworkPolicy first — the monitoring probes get away without one only
+# because their targets share a namespace with the prober".
+#
+# That is wrong, and it was wrong when written. It is stated here rather than
+# fixed in place because the sentence lives inside the `config.yaml: |` block
+# scalar below: any edit to it changes ConfigMap DATA, and
+# `configmap.reloader.stakater.com/reload: gatus-config`
+# (deployment.yaml#reloader) would then roll the gatus Deployment — a
+# production restart for a comment. This header is outside `data:`, so it is
+# inert; the served config is byte-identical.
+#
+#   * ../../infrastructure/network-policies/paperless-ns.yaml#monitoring-admit
+#     ALREADY admits the whole monitoring namespace on every port, and
+#     monitoring is where
+#     blackbox-exporter runs. Cilium unions the allow rules of every CNP
+#     selecting an endpoint, so no extra policy is required.
+#   * The in-cluster paperless probe therefore exists, added with no change
+#     to the paperless namespace:
+#     monitoring-rules/probes-blackbox.yaml, target
+#     http://paperless.paperless.svc.cluster.local:80/accounts/login/
+#   * The claim does NOT generalise to gatus itself. Gatus runs in `apps`,
+#     which is a different source namespace; a gatus-sourced in-cluster check
+#     would need its own allow, exactly as
+#     cnp-cross-namespace-allow-apps-gatus-to-monitored-services describes.
+#
+# The Web Service 11 entry below stays a reachability-only check on purpose.
+# The app-health half of the pair is the blackbox probe named above.
 data:
+  # anchor: gatus-endpoints
   config.yaml: |
     storage:
       type: sqlite

kubernetes/apps/gatus/deployment.yaml

@@ -6,6 +6,7 @@ metadata:
   labels:
     app: gatus
   annotations:
+    # anchor: reloader
     # Gatus reads its config once at startup and never re-reads the file, so
     # without this a ConfigMap change reconciles cleanly and then does nothing:
     # git and the running checks drift apart silently, with no failed resource
@@ -46,7 +47,7 @@ spec:
           type: RuntimeDefault
       containers:
         - name: gatus
-          image: twinproduction/gatus:v5.36.0 # {"$imagepolicy": "flux-system:gatus"}
+          image: twinproduction/gatus:v5.37.0 # {"$imagepolicy": "flux-system:gatus"}
           securityContext:
             allowPrivilegeEscalation: false
             capabilities:

kubernetes/apps/gitea/ingress.yaml

@@ -2,7 +2,7 @@ apiVersion: networking.k8s.io/v1
 kind: Ingress
 metadata:
   name: gitea
-  namespace: apps
+  namespace: gitea
   labels:
     app.kubernetes.io/name: gitea
   annotations:

kubernetes/apps/gitea/ingressroute-ssh.yaml

@@ -2,7 +2,7 @@ apiVersion: traefik.io/v1alpha1
 kind: IngressRouteTCP
 metadata:
   name: gitea-ssh
-  namespace: apps
+  namespace: gitea
   labels:
     app.kubernetes.io/name: gitea
 spec:

kubernetes/apps/gitea/kustomization.yaml

@@ -6,7 +6,6 @@ kind: Kustomization
 # and verified. See MIGRATION-gitea-woodpecker.md.
 resources:
   - statefulset.yaml
-  - pdb.yaml
   - service.yaml
   - ingressroute-ssh.yaml
   - ingress.yaml

kubernetes/apps/gitea/networkpolicy.yaml

@@ -1,28 +1,8 @@
-apiVersion: cilium.io/v2
-kind: CiliumNetworkPolicy
-metadata:
-  name: gitea-allow-metrics-scrape
-  namespace: apps
-spec:
-  # Prometheus (monitoring ns) scrapes gitea's /metrics on :3000. The apps
-  # default-deny-ingress only allows same-ns / home-assistant / host, so the
-  # scrape is dropped (context deadline exceeded). This became necessary only
-  # after the LXC->in-cluster cutover — the old static target was external, so
-  # no in-cluster policy applied. Scoped to the gitea pod + :3000.
-  endpointSelector:
-    matchLabels:
-      app.kubernetes.io/name: gitea
-  ingress:
-    - fromEndpoints:
-        - matchLabels:
-            "k8s:io.kubernetes.pod.namespace": monitoring
-      toPorts:
-        - ports:
-            - port: "3000"
-              protocol: TCP
----
+# Prometheus scrape and Traefik are admitted by the namespace baseline in
+# infrastructure/network-policies/gitea-ns.yaml.
+#
 # Secret rotation: the flux-git rotator (openbao ns) reaches gitea to mint a new
-# flux PAT. It uses the in-cluster Service (gitea.apps.svc:3000) — the public
+# flux PAT. It uses the in-cluster Service (gitea.gitea.svc:3000) — the public
 # gitea.renner.uno name hairpins off the ingress LB from a pod (same failure as
 # pve.renner.uno). Scoped to the gitea pod :3000 and the rotator label; the
 # egress half is rotator-gitea-flux-egress in network-policies/openbao-ns.yaml.
@@ -30,7 +10,7 @@ apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
 metadata:
   name: gitea-allow-flux-rotator
-  namespace: apps
+  namespace: gitea
 spec:
   endpointSelector:
     matchLabels:
@@ -54,7 +34,7 @@ apiVersion: cilium.io/v2
 kind: CiliumNetworkPolicy
 metadata:
   name: gitea-allow-review-bot-rotator
-  namespace: apps
+  namespace: gitea
 spec:
   endpointSelector:
     matchLabels:

kubernetes/apps/gitea/pdb.yaml

@@ -1,23 +0,0 @@
----
-# Eviction protection for the single-replica Gitea StatefulSet.
-#
-# Gitea runs on SQLite (single writer) and its repos + DB live on a single
-# ReadWriteOnce PVC, so it is deliberately replicas:1 and cannot fail over.
-# With no PDB, a node drain / maintenance evicts the pod abruptly, mid-write,
-# with zero coordination. minAvailable:1 on a 1-replica workload makes the
-# eviction API refuse *voluntary* evictions: `kubectl drain` blocks and
-# surfaces this pod, forcing a conscious decision (drain with
-# --disable-eviction, or scale down first) instead of yanking the live DB.
-# It does NOT block involuntary disruptions (node crash, OOM).
-apiVersion: policy/v1
-kind: PodDisruptionBudget
-metadata:
-  name: gitea
-  namespace: apps
-  labels:
-    app.kubernetes.io/name: gitea
-spec:
-  minAvailable: 1
-  selector:
-    matchLabels:
-      app.kubernetes.io/name: gitea

kubernetes/apps/gitea/service.yaml

@@ -2,7 +2,7 @@ apiVersion: v1
 kind: Service
 metadata:
   name: gitea
-  namespace: apps
+  namespace: gitea
   labels:
     app.kubernetes.io/name: gitea
 spec:

kubernetes/apps/gitea/statefulset.yaml

@@ -2,7 +2,7 @@ apiVersion: apps/v1
 kind: StatefulSet
 metadata:
   name: gitea
-  namespace: apps
+  namespace: gitea
   labels:
     app.kubernetes.io/name: gitea
 spec:
@@ -152,7 +152,7 @@ spec:
           # NOTE: volumeClaimTemplates is immutable on an existing StatefulSet,
           # so bumping this value cannot be `kubectl apply`-ed in place — Flux
           # will reject the whole object. Landing this size change requires a
-          # one-time `kubectl delete statefulset gitea -n apps --cascade=orphan`
+          # one-time `kubectl delete statefulset gitea -n gitea --cascade=orphan`
           # (pod + PVC keep running) so Flux re-creates the STS from this
           # template and re-adopts them. After the cleanup_packages GC reclaims
           # the ~10G of orphaned blobs, real usage is ~9G, so 40Gi is headroom.

kubernetes/apps/grafana-db/postgres.yaml

@@ -4,7 +4,9 @@ metadata:
   name: gf-pg
   namespace: monitoring
 spec:
-  instances: 1
+  instances: 2
+  smartShutdownTimeout: 10
+  stopDelay: 60
   imageName: ghcr.io/cloudnative-pg/postgresql:16
   storage:
     storageClass: ceph-block

kubernetes/apps/home-assistant/deployment.yaml

@@ -95,7 +95,7 @@ spec:
               memory: 32Mi
       containers:
         - name: home-assistant
-          image: ghcr.io/home-assistant/home-assistant:2026.9.3 # {"$imagepolicy": "flux-system:home-assistant"}
+          image: ghcr.io/home-assistant/home-assistant:2026.9.4 # {"$imagepolicy": "flux-system:home-assistant"}
           env:
             - name: TZ
               value: Europe/Berlin

kubernetes/apps/home-assistant/ha-pg-cluster.yaml

@@ -30,16 +30,71 @@ metadata:
   name: ha-pg
   namespace: home-assistant
 spec:
-  instances: 1
+  instances: 2
+  # switchover, not CNPG's default `restart`: any change that rolls the pods
+  # (walStorage below did) then promotes the already-rolled replica instead of
+  # restarting the primary in place, so the recorder loses seconds of writes
+  # rather than stopDelay plus a pg_wal move plus startup (REN-176).
+  primaryUpdateMethod: switchover
+  smartShutdownTimeout: 10
+  stopDelay: 60
   imageName: ghcr.io/cloudnative-pg/postgresql:16
   storage:
     storageClass: ceph-block
-    # Sized for the migrated statistics plus 7 days of states at the current
-    # rate: statistics 10.8M rows + statistics_short_term 1.9M (~1.8 GB with
-    # indexes in Postgres) and ~18M states rows for the purge_keep_days:7 window
-    # (~4 GB). 20Gi leaves room for autovacuum bloat and index churn.
-    # ceph-block has allowVolumeExpansion: true, so this can grow in place.
-    size: 20Gi
+    # Sized for WAL headroom during an archiving stall, not for table growth.
+    # The data itself needs far less: migrated statistics 10.8M rows +
+    # statistics_short_term 1.9M (~1.8 GB with indexes in Postgres) and ~18M
+    # states rows for the purge_keep_days:7 window (~4 GB).
+    #
+    # 20Gi left only 6.63 GiB free, and that is the number that matters here.
+    # WAL archiving stalled for 10 minutes and the un-archived backlog reached
+    # 1.84 GiB — 28% of that free space (REN-129).
+    #
+    # That stall was first blamed on Velero repository maintenance. It is not
+    # the cause: REN-136 measured seven nights and the archiver's sustained
+    # drain (~14 segments/min, its serial ceiling) is flat whether Kopia is
+    # running or not, and Kopia's nearest batch drifts across 02:03-02:24Z
+    # while the burst stays pinned to 02:13Z. What is real is a nightly
+    # producer burst — 56-119 segments queued at 02:13-02:15 UTC on all seven
+    # nights, against a steady-state queue of 0. Postgres cannot recycle a WAL
+    # segment until the archiver has shipped it, so a burst that arrives while
+    # archiving is stalled walks the volume toward full, and a full PGDATA
+    # stops the recorder outright.
+    #
+    # 32Gi takes free space to ~18.6 GiB, which puts the same 1.84 GiB backlog
+    # at 10% of headroom rather than 28%. In REN-136's terms: at 6.36 GiB free
+    # one stalled night survives and two fill the volume; ~18.6 GiB carries
+    # about four. Approved by the owner on REN-135.
+    # ceph-block has allowVolumeExpansion: true, so this grows in place.
+    #
+    # Since REN-176 WAL lives on walStorage below, so the stall headroom above
+    # now belongs to that volume. 32Gi stays because a PVC cannot shrink.
+    size: 32Gi
+  # WAL on its own volume (ha-pg-N-wal), approved by the owner on REN-176.
+  #
+  # PERMANENT. The CNPG webhook rejects removing walStorage once set ("walStorage
+  # cannot be disabled once configured"), so reverting this block fails
+  # admission and leaves the home-assistant Kustomization stuck. Undoing it
+  # means a new Cluster recovered from ha-pg-store. Adding it on 1.30.1 needs
+  # no re-clone: the operator creates the PVC, rolls each pod, and the instance
+  # manager moves pg_wal onto it before Postgres starts.
+  #
+  # This does NOT keep the recorder up when WAL fills: a full WAL disk still
+  # stops Postgres. What it does is keep WAL writable if PGDATA fills (a purge
+  # that stops running), and make WAL-volume usage a direct archiver-backlog
+  # signal.
+  #
+  # Sized so stall survival is no worse than the shared 32Gi gives today.
+  # Today: ~18.6 GiB free on PGDATA / 4.66 GiB/day of WAL = ~4.0 days.
+  #   1.86 GiB  normal pg_wal footprint (max_wal_size 2GB)
+  # + 18.64 GiB four days of full archiver outage at 4.66 GiB/day
+  #             (the nightly 56-119 segment burst, REN-136, is inside that rate)
+  # = 20.5 GiB -> 21Gi: (21 - 1.86) / 4.66 = ~4.1 days. 18Gi would give ~3.5,
+  # worse than today. RBD is thin, so the ceiling costs nothing until a stall
+  # fills it. The webhook requires min_wal_size < max_wal_size < this size.
+  walStorage:
+    storageClass: ceph-block
+    size: 21Gi
   plugins:
     - name: barman-cloud.cloudnative-pg.io
       isWALArchiver: true
@@ -124,8 +179,7 @@ spec:
       # padding rather than data.
       #
       # The cost is RPO: up to 30min of WAL is unrecoverable if this cluster and
-      # its volume are lost outright. This is a single-instance cluster, so the
-      # archive is the only off-node copy.
+      # its volumes are lost outright.
       archive_timeout: "30min"
       # Fewer checkpoints => fewer full-page writes, since a page is only
       # fully-logged on its first modification after a checkpoint.

kubernetes/apps/home-assistant/kustomization.yaml

@@ -16,6 +16,9 @@ resources:
   # health broker credential: OpenBao-backed + Class-A rotated (first of the
   # eight users to migrate off mosquitto-bootstrap.yaml).
   - mosquitto-health-credentials-externalsecret.yaml
+  # open3e-int broker credential: open3e-esp32 gateway on the Vitocal internal
+  # CAN; OpenBao-backed + Class-A rotated.
+  - mosquitto-open3e-int-credentials-externalsecret.yaml
   - mosquitto-config.yaml
   - mosquitto-certificate.yaml
   - mosquitto-service.yaml

kubernetes/apps/home-assistant/mosquitto-deployment.yaml

@@ -12,7 +12,7 @@ metadata:
     #
     # Listed per-Secret rather than with the blanket auto annotation, so that
     # enrolling another broker user is a visible edit to this line.
-    secret.reloader.stakater.com/reload: mosquitto-health-credentials
+    secret.reloader.stakater.com/reload: mosquitto-health-credentials,mosquitto-open3e-int-credentials
 spec:
   replicas: 1
   strategy:
@@ -78,6 +78,7 @@ spec:
               mosquitto_passwd -b /mosquitto/auth/passwd admin "$ADMIN_PASSWORD"
               mosquitto_passwd -b /mosquitto/auth/passwd sml2mqtt "$SML2MQTT_PASSWORD"
               mosquitto_passwd -b /mosquitto/auth/passwd open3e "$OPEN3E_PASSWORD"
+              mosquitto_passwd -b /mosquitto/auth/passwd open3e-int "$OPEN3E_INT_PASSWORD"
               mosquitto_passwd -b /mosquitto/auth/passwd evcc "$EVCC_PASSWORD"
               mosquitto_passwd -b /mosquitto/auth/passwd health "$HEALTH_PASSWORD"
               mosquitto_passwd -b /mosquitto/auth/passwd zigbee2mqtt "$ZIGBEE2MQTT_PASSWORD"
@@ -124,6 +125,12 @@ spec:
               user open3e
               topic readwrite open3e/#
 
+              # open3e-esp32 gateway on the Vitocal internal CAN: own subtree
+              # under open3e/ so edge-1's listener on open3e/cmnd never sees
+              # its commands.
+              user open3e-int
+              topic readwrite open3e/int/#
+
               user evcc
               topic read open3e/#
               topic readwrite evcc/#
@@ -187,6 +194,13 @@ spec:
                 secretKeyRef:
                   name: mosquitto-open3e-credentials
                   key: password
+            # open3e-int: OpenBao-backed + Class-A rotated (see
+            # mosquitto-open3e-int-credentials-externalsecret.yaml).
+            - name: OPEN3E_INT_PASSWORD
+              valueFrom:
+                secretKeyRef:
+                  name: mosquitto-open3e-int-credentials
+                  key: password
             - name: EVCC_PASSWORD
               valueFrom:
                 secretKeyRef:

kubernetes/apps/home-assistant/mosquitto-open3e-int-credentials-externalsecret.yaml

@@ -0,0 +1,21 @@
+# Projects the mosquitto `open3e-int` broker credential (open3e-esp32 gateway
+# on the Vitocal internal CAN) from OpenBao into the Secret the broker
+# consumes. Rotator: kubernetes/infrastructure/openbao/rotation/mosquitto-open3e-int.yaml.
+#
+# creationPolicy is implicitly Owner: deleting this ExternalSecret deletes the
+# live Secret. Set creationPolicy: Orphan first to detach without destroying.
+apiVersion: external-secrets.io/v1
+kind: ExternalSecret
+metadata:
+  name: mosquitto-open3e-int-credentials
+  namespace: home-assistant
+spec:
+  refreshInterval: 10m
+  secretStoreRef:
+    kind: ClusterSecretStore
+    name: openbao
+  target:
+    name: mosquitto-open3e-int-credentials
+  dataFrom:
+    - extract:
+        key: apps/home-assistant/mosquitto-open3e-int-credentials

kubernetes/apps/ical-proxy/deployment.yaml

@@ -24,7 +24,7 @@ spec:
           # unchanged, so Kubernetes has no reason to replace anything, and
           # this image is deliberately NOT enrolled in Flux image automation
           # — see kubernetes/infrastructure/image-automation/README.md, which
-          # lists ghcr.io/konairius/ical-proxy as having "only main, latest,
+          # lists gitea.renner.uno/konairius/ical-proxy as having "only main, latest,
           # and main-<sha> commit builds; no release tags", so there is no
           # ImagePolicy watching it. The pods therefore sat on a 21-day-old
           # image while :latest moved underneath them, and the only way to
@@ -35,15 +35,23 @@ spec:
           # what is RUNNING recoverable from this repo rather than "whatever
           # :latest meant the last time a pod happened to restart".
           #
-          # main-99a912b is the merge of konairius/ical-proxy#12, which adds
-          # the summary_match/summary_replace parameters this deployment
-          # needs to strip "| Abfuhrkalender - Landkreis Amberg-Sulzbach"
-          # from every waste collection.
+          # The image comes from the Gitea package registry, built by
+          # Woodpecker (.woodpecker.yml in konairius/ical-proxy) on every
+          # push to main. It is amd64-only and pulls without credentials, so
+          # no imagePullSecret. main-4783abe adds two hardening changes on
+          # top of main-db90a59: an SSRF guard (upstreams and redirect hops
+          # resolving to private/loopback/link-local addresses are refused;
+          # public redirects are still followed, up to 5) and a cap of 20
+          # concurrent /proxy upstream fetches (over the cap: 429 +
+          # Retry-After). main-73b5d97 then raises the server WriteTimeout
+          # from 10s to 40s (the 30s upstream fetch timeout + 10s headroom),
+          # so a slow-but-legitimate upstream is no longer cut off before
+          # /proxy can answer. The summary_match/summary_replace parameters
+          # this deployment needs are unchanged.
           #
           # Bumping this is a manual, reviewed edit. The tag format is
-          # main-<short-sha> (docker/metadata-action type=sha with the branch
-          # as prefix).
-          image: ghcr.io/konairius/ical-proxy:main-99a912b
+          # main-<short-sha>, the first 7 characters of the main commit.
+          image: gitea.renner.uno/konairius/ical-proxy:main-73b5d97
           ports:
             - containerPort: 8080
               name: http

kubernetes/apps/immich/cluster.yaml

@@ -59,8 +59,7 @@ spec:
       # padding rather than data.
       #
       # The cost is RPO: up to 30min of WAL is unrecoverable if this cluster and
-      # its volume are lost outright. This is a single-instance cluster, so the
-      # archive is the only off-node copy.
+      # its volumes are lost outright.
       archive_timeout: "30min"
       # Write-reduction tuning. These databases are largely idle, and the
       # Postgres defaults favour throughput over storage that wears out.

kubernetes/apps/immich/pvcs.yaml

@@ -28,6 +28,38 @@ spec:
     requests:
       storage: 300Gi
 ---
+# Media root on the EC 2+1 bulk pool; thumbs mount separately below.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: immich-library-bulk
+  namespace: immich
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled
+spec:
+  accessModes:
+    - ReadWriteMany
+  storageClassName: ceph-filesystem-bulk
+  resources:
+    requests:
+      storage: 300Gi
+---
+# Thumbnails stay replicated: read affinity serves them from the local OSD.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: immich-thumbs
+  namespace: immich
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled
+spec:
+  accessModes:
+    - ReadWriteMany
+  storageClassName: ceph-filesystem
+  resources:
+    requests:
+      storage: 30Gi
+---
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:

kubernetes/apps/immich/server.yaml

@@ -110,6 +110,8 @@ spec:
           volumeMounts:
             - name: library
               mountPath: /data
+            - name: thumbs
+              mountPath: /data/thumbs
             - name: immich-config-file
               mountPath: /config
               readOnly: true
@@ -145,7 +147,10 @@ spec:
       volumes:
         - name: library
           persistentVolumeClaim:
-            claimName: immich-library
+            claimName: immich-library-bulk
+        - name: thumbs
+          persistentVolumeClaim:
+            claimName: immich-thumbs
         - name: immich-config-file
           secret:
             secretName: immich-secrets

kubernetes/apps/influxdb/helmrelease.yaml

@@ -61,6 +61,9 @@ spec:
       enabled: true
       storageClass: ceph-block
       size: 50Gi
+    # A minAvailable:1 PDB on a single replica blocks every node drain.
+    pdb:
+      create: false
     # Bootstrap admin/org/bucket on the fresh PVC. Admin password + token come
     # from the influxdb-bootstrap Secret (OpenBao via ESO), deterministic across
     # rebuilds so the bootstrap token is the live admin token. `influx setup`

kubernetes/apps/kustomization.yaml

@@ -3,8 +3,8 @@ kind: Kustomization
 resources:
   - gatus/
   - external-dns/
-  - gitea/
   - ical-proxy/
+  # gitea/ has its own `gitea` Kustomization (clusters/homelab/gitea.yaml, #878).
   # woodpecker/ moved to its own dedicated `woodpecker` Kustomization
   # (clusters/homelab/woodpecker.yaml) so its wp-pg CNPG cluster's barman
   # dependency + Cluster health-gating don't contaminate this shared apps KS.

kubernetes/apps/libation/cronjob.yaml

@@ -4,7 +4,7 @@
 # the entrypoint exit after one pass instead of looping internally —
 # the CronJob handles scheduling.
 #
-# Books land in /audiobooks (the audiobookshelf-books PVC, shared
+# Books land in /audiobooks (the audiobookshelf-books-bulk PVC, shared
 # with the audiobookshelf Deployment). Config + DB live on
 # libation-config (RWO ceph-block, seeded once from
 # libation-bootstrap secret).
@@ -107,6 +107,9 @@ spec:
                 - |
                   trap 'cp /config-internal/AccountsSettings.json /libation-state/AccountsSettings.json' EXIT
                   set -e
+                  # Strip the Audible intro/outro; re-liberated books replace their file instead of gaining a " (1)" copy.
+                  jq '.StripAudibleBrandAudio = true | .OverwriteExisting = true' /config-internal/Settings.json > /config-internal/Settings.json.tmp
+                  mv /config-internal/Settings.json.tmp /config-internal/Settings.json
                   /libation/LibationCli scan
                   /libation/LibationCli liberate
               env:
@@ -135,7 +138,7 @@ spec:
                 claimName: libation-config
             - name: books
               persistentVolumeClaim:
-                claimName: audiobookshelf-books
+                claimName: audiobookshelf-books-bulk
             - name: bootstrap
               secret:
                 secretName: libation-bootstrap

kubernetes/apps/libation/pvc.yaml

@@ -5,7 +5,7 @@
 # SOPS-encrypted Secret on first boot (sentinel file gate); after that
 # Libation reads and writes it directly so token rotation persists.
 #
-# Books are written to the audiobookshelf-books PVC (already provisioned
+# Books are written to the audiobookshelf-books-bulk PVC (already provisioned
 # by the audiobookshelf seed Kustomization, RWX ceph-filesystem). We
 # share the same namespace so the PVC reference is direct — k8s PVCs
 # are namespace-scoped.

kubernetes/apps/monitoring-rules/cilium-servicemonitor.yaml

@@ -1,8 +1,75 @@
+---
+# The agent metrics Service, declared here because the chart does not render
+# one under our values (REN-147).
+#
+# cilium's templates/cilium-agent/service.yaml is gated twice:
+#
+#   {{- if and .Values.agent (not .Values.preflight.enabled) .Values.prometheus.enabled }}
+#   {{- if (or .Values.prometheus.serviceMonitor.enabled .Values.prometheus.metricsService) }}
+#
+# infrastructure/cilium/helmrelease.yaml sets `prometheus.enabled: true` and
+# `prometheus.serviceMonitor.enabled: false`, and leaves `metricsService` at its
+# default of false. First gate passes, second does not. The chart's fallback
+# branch is envoy-only and needs envoy in the agent pod, but envoy runs as its
+# own DaemonSet here — so no `cilium-agent` Service was rendered at all.
+#
+# The consequence was silent: agents exported metrics on :9962 all along, the
+# ServiceMonitor below selected a Service that did not exist, and the scrape
+# pool sat at 0 active / 46 dropped targets. `cilium_*` had zero series, so
+# CiliumBPFMapPressureHigh and CiliumBPFMapFull in prometheusrule-cilium.yaml
+# evaluated an empty vector forever and could never fire.
+#
+# Why a Service here and not `prometheus.metricsService: true` in the
+# HelmRelease: that flag is the chart-native fix and its rendered diff is
+# exactly this one object (it appears in no other template — not the
+# ConfigMap, not the DaemonSet). But setting it changes the HelmRelease values,
+# which makes helm-controller run `helm upgrade` on the CNI release, and that
+# three-way-merges against live objects. PROPOSAL-cilium-single-owner.md
+# records that `cilium-config` has carried a `kubectl-patch` field manager; if
+# any of that drift is still live, the upgrade reconciles it, and the
+# `configmap.reloader.stakater.com/reload: cilium-config` annotation on the
+# HelmRelease then rolls every agent node by node. A Service applied by the
+# apps Kustomization cannot do that. If the drift is ever proven gone, delete
+# this object and set `prometheus.metricsService: true` instead.
+#
+# Name and selector follow authentik-outpost-servicemonitors.yaml: a distinct
+# name plus `app.kubernetes.io/managed-by: homelab`, so flipping either chart
+# flag later renders the chart's own `cilium-agent` Service without an
+# ownership conflict and without double-scraping the same endpoints.
+apiVersion: v1
+kind: Service
+metadata:
+  name: cilium-agent-metrics
+  namespace: kube-system
+  labels:
+    app.kubernetes.io/managed-by: homelab
+    app.kubernetes.io/name: cilium-agent
+    app.kubernetes.io/part-of: cilium
+    # Copied onto every series by the ServiceMonitor's targetLabels below.
+    # The dashboard the chart ships queries
+    # cilium_bpf_map_pressure{k8s_app="cilium"}, so this label is load-bearing.
+    k8s-app: cilium
+spec:
+  type: ClusterIP
+  # Headless on purpose: agent metrics are per-node and every endpoint has to
+  # be scraped. A ClusterIP would hand Prometheus one random agent per scrape.
+  clusterIP: None
+  # Agent pods only. cilium-envoy carries k8s-app: cilium-envoy and exposes
+  # 9964, so it cannot match.
+  selector:
+    k8s-app: cilium
+  ports:
+    # Named and numbered exactly as the chart's own Service, so the
+    # ServiceMonitor below stays a field-for-field copy of the chart's.
+    - name: metrics
+      port: 9962
+      targetPort: prometheus
+      protocol: TCP
+---
 # Cilium's agent ServiceMonitor, declared here rather than by the chart.
 #
-# Reproduces what cilium 1.20.1 renders with prometheus.serviceMonitor.enabled,
-# taken field for field off the live object so scraping is unchanged: same name,
-# same namespace, same selector, same 10s interval and node relabeling.
+# Reproduces what cilium renders with prometheus.serviceMonitor.enabled: same
+# name, same namespace, same 10s interval and node relabeling.
 #
 # It lives here because the chart's template is gated on monitoring.coreos.com/v1
 # existing at install time, and on an empty cluster it does not — those CRDs
@@ -22,13 +89,14 @@ metadata:
     app.kubernetes.io/part-of: cilium
 spec:
   selector:
+    # Pinned to the Service above. Without managed-by this would also match a
+    # chart-rendered `cilium-agent` Service and scrape the same endpoints twice.
     matchLabels:
+      app.kubernetes.io/managed-by: homelab
       app.kubernetes.io/name: cilium-agent
   namespaceSelector:
     matchNames:
       - kube-system
-  # The chart's Service keeps rendering (prometheus.enabled), so the target and
-  # its `metrics` port are there regardless of who declares the ServiceMonitor.
   targetLabels:
     - k8s-app
   endpoints:

kubernetes/apps/monitoring-rules/external-secrets-podmonitor.yaml

@@ -0,0 +1,71 @@
+# Scrape the external-secrets (ESO) controller's own metrics. Until this
+# existed, nothing in this cluster could see a failed ExternalSecret sync:
+# `externalsecret_status_condition` held 0 series, so no rule could be written
+# about a key that silently stops syncing (REN-100).
+#
+# This is the scrape only. The alert rule is deliberately NOT in this change —
+# its `for:` window has to be tuned against real data, and that data does not
+# exist until this has been live for a while. See REN-103.
+#
+# A PodMonitor, NOT the chart's `serviceMonitor.enabled`, and NOT a hand-written
+# ServiceMonitor. Both alternatives were checked against chart 2.11.0:
+#
+#   * A hand-written ServiceMonitor has nothing to select. `metrics.service.enabled`
+#     defaults to false and infrastructure/external-secrets/helmrelease.yaml
+#     overrides nothing, so templates/service.yaml renders no Service for the
+#     controller. The container port exists (`name: metrics`, containerPort 8080)
+#     but nothing fronts it — the same trap that forced a PodMonitor for the Flux
+#     controllers.
+#   * Setting `serviceMonitor.enabled: true` would work, but it triggers a Helm
+#     UPGRADE of ESO, which re-renders the Deployment and re-applies its CRDs
+#     (`crds: CreateReplace` in the HelmRelease). ESO is what all 54 ExternalSecret
+#     manifests in this repo depend on. A PodMonitor gets the same series while
+#     touching the release not at all, which is the trade this repo already made
+#     for Velero's node-agent (velero-podmonitor.yaml) and for Flux.
+#
+# THIS NEEDS A NETWORK POLICY CHANGE, unlike the Flux PodMonitor: the cluster is
+# default-deny and external-secrets-ns.yaml admitted only intra-namespace traffic,
+# host/remote-node and ICMP — no `monitoring` grant, so this scrape would have
+# been dropped and the metric would have stayed absent. The matching ingress rule
+# is added to infrastructure/network-policies/external-secrets-ns.yaml in the
+# same pull request. Without it this file is decoration.
+apiVersion: monitoring.coreos.com/v1
+kind: PodMonitor
+metadata:
+  name: external-secrets
+  namespace: monitoring
+  labels:
+    app.kubernetes.io/name: external-secrets
+    # Not required for discovery — podMonitorSelector is empty, see
+    # apps/monitoring/helmrelease-prometheus.yaml — but set for consistency with
+    # the other monitors in this directory.
+    release: kube-prometheus-stack
+spec:
+  # Lives in `monitoring` rather than in `external-secrets`, matching
+  # external-dns-servicemonitor.yaml and the Flux PodMonitor. Every object in
+  # the external-secrets namespace is owned by the ESO HelmRelease; keeping this
+  # out of that namespace means this change cannot interact with the release.
+  namespaceSelector:
+    matchNames:
+      - external-secrets
+  selector:
+    matchLabels:
+      # The CONTROLLER only. The webhook and cert-controller carry
+      # `app.kubernetes.io/name: external-secrets-webhook` and
+      # `-cert-controller` respectively (see the chart's
+      # external-secrets-webhook.selectorLabels / -cert-controller.selectorLabels
+      # helpers), so an exact match on the unsuffixed name excludes both. That is
+      # intended: the `externalsecret_*` series are emitted by the controller, and
+      # the other two would only add idle targets.
+      app.kubernetes.io/name: external-secrets
+      app.kubernetes.io/instance: external-secrets
+  podMetricsEndpoints:
+    # Named container port, not a hardcoded 8080, so this follows
+    # `metrics.listen.port` if it is ever overridden in the HelmRelease.
+    - port: metrics
+      # 60s, not the chart's 30s default: `externalsecret_status_condition` is a
+      # slow-moving gauge over ~54 ExternalSecrets, and the alert built on it in
+      # REN-103 will use a multi-minute `for:` window. 30s would double the
+      # samples for no gain in detection time.
+      interval: 60s
+      scrapeTimeout: 10s

kubernetes/apps/monitoring-rules/flux-podmonitor.yaml

@@ -0,0 +1,91 @@
+# Scrape the Flux controllers' own metrics. Until this existed Prometheus held
+# no `gotk_*` series at all, so nothing about Flux's own behaviour was
+# queryable and no rule could be written about it (REN-97).
+#
+# THIS IS THE SCRAPE ONLY, and the readiness alert is not merely deferred —
+# it cannot be written from this endpoint. `gotk_reconcile_condition` and
+# `gotk_suspend_status` are still DECLARED by fluxcd/pkg/runtime (v0.110.3,
+# runtime/metrics/recorder.go), but at the tags pinned in gotk-components.yaml
+# no controller calls the recorder that populates them: grepping the six
+# tagged source trees for RecordReadiness / RecordCondition / RecordSuspend
+# returns zero call sites, and an unpopulated GaugeVec exports no series. So
+# the readiness metrics would stay empty however this file is written. Flux
+# moved per-object readiness to kube-state-metrics custom-resource-state
+# (`gotk_resource_info{ready=...}`); that is REN-102, and it is a prerequisite
+# for the not-Ready alert rather than a follow-up to it.
+#
+# What this scrape does yield, and why it is still worth merging on its own:
+#   * `gotk_reconcile_duration_seconds` per kind/name/namespace — reconcile
+#     latency and, by its absence, an object that has stopped being reconciled
+#     at all. RecordDuration IS called, but by FIVE of the six controllers, not
+#     all six: measured post-merge, 170 series from kustomize (33), source (43),
+#     image-reflector (68), image-automation (1) and helm (25), and ZERO from
+#     notification-controller. That is not a gap waiting to fill in — its three
+#     reconcilers are registered but never walked (controller_runtime_reconcile_
+#     total for controller=alert/provider/receiver is 0 over 56 h of uptime with
+#     2 Alerts and 1 Provider live), so the family is absent from that pod's
+#     /metrics entirely and five is the permanent value. DO NOT write a rule of
+#     the form `count(... gotk_reconcile_duration_seconds ...) < 6`, or an
+#     absent() per pod: it fires forever on notification-controller. Liveness for
+#     all six pods is `up{job="monitoring/flux-controllers"}` (below); that pod's
+#     reconcile health is `controller_runtime_reconcile_errors_total`. REN-127.
+#   * controller-runtime's own series: reconcile totals and errors per
+#     controller, workqueue depth and retry counts, leader election.
+#   * the target itself, so `up{job="monitoring/flux-controllers"}` and the
+#     chart's default TargetDown rule start covering the Flux control plane.
+#   * the scrape REN-102 needs in place anyway.
+#
+# A PodMonitor, NOT a ServiceMonitor, and that is forced rather than preferred:
+# gotk-components.yaml creates Services for only two of the six controllers
+# (source-controller, notification-controller), and both publish port 80 ->
+# targetPort `http`, which is the artifact/webhook server on :9090 — not the
+# metrics port. kustomize-, helm-, image-reflector- and image-automation-
+# controller have no Service at all, so there is nothing for a ServiceMonitor to
+# select. All six pods do carry the same two labels and the same `http-prom`
+# (:8080) container port, so one PodMonitor covers the set.
+#
+# NO NETWORK POLICY CHANGE IS NEEDED, verified rather than assumed, and unlike
+# external-secrets-podmonitor.yaml which did need one: flux-system has no
+# CiliumNetworkPolicy of its own (infrastructure/network-policies/ has no
+# flux-system file), and gotk-components.yaml ships an `allow-scraping`
+# NetworkPolicy admitting every namespace to :8080 (`from: [namespaceSelector:
+# {}]`, port 8080). So `monitoring` is already admitted, as it was for REN-21.
+# Nothing in the Flux install is touched.
+apiVersion: monitoring.coreos.com/v1
+kind: PodMonitor
+metadata:
+  name: flux-controllers
+  namespace: monitoring
+  labels:
+    app.kubernetes.io/part-of: flux
+    # Not required for discovery — podMonitorSelector is empty, see
+    # apps/monitoring/helmrelease-prometheus.yaml — but set for consistency
+    # with the other monitors in this directory.
+    release: kube-prometheus-stack
+spec:
+  # Lives in `monitoring` with a namespaceSelector rather than in flux-system,
+  # matching external-dns-servicemonitor.yaml and external-secrets-podmonitor.yaml.
+  # Flux owns every object in flux-system via gotk-components.yaml; keeping this
+  # out of that namespace means this change cannot interact with the Flux install
+  # itself.
+  namespaceSelector:
+    matchNames:
+      - flux-system
+  selector:
+    matchLabels:
+      app.kubernetes.io/part-of: flux
+      app.kubernetes.io/instance: flux-system
+  podMetricsEndpoints:
+    # Matches the controllers' `http-prom` port (containerPort 8080). Present on
+    # all six Deployments in gotk-components.yaml; a future controller without
+    # it simply yields no target rather than a broken scrape config.
+    - port: http-prom
+      path: /metrics
+      # 60s, not the 30s used elsewhere here, matching the ESO PodMonitor: the
+      # useful series are slow-moving per-object gauges and histograms over a
+      # few hundred Flux objects, and any rule built on them (REN-102) will use
+      # a multi-minute `for:`. 30s would double the samples for no gain in
+      # detection time — and Prometheus is already near its retention cap
+      # (REN-107).
+      interval: 60s
+      scrapeTimeout: 10s

kubernetes/apps/monitoring-rules/kustomization.yaml

@@ -21,6 +21,9 @@ resources:
   # alerted if that stopped — HetznerBackupEndpointDown only probes the S3
   # endpoint, which proves reachability, not that anything was written.
   - prometheusrule-cnpg-backup.yaml
+  # Home Assistant, its recorder DB, Mosquitto and Zigbee2MQTT (REN-244). The
+  # only signal for them before was the generic 15m KubePodNotReady warning.
+  - prometheusrule-home-assistant.yaml
   # DNS publication. Nothing scraped external-dns, so it deleted six LAN
   # records on 2026-09-12 and stayed silent for five days.
   - prometheusrule-external-dns.yaml
@@ -30,6 +33,9 @@ resources:
   - prometheusrule-dns-resolution.yaml
   - probes-dns.yaml
   - alloy-servicemonitor.yaml
+  # Agent metrics. Ships the metrics Service as well as the ServiceMonitor:
+  # the chart renders neither under our values, so this scraped nothing and
+  # the two CiliumBPFMapPressure alerts could never fire. See the file.
   - cilium-servicemonitor.yaml
   - cnpg-podmonitors.yaml
   - crowdsec-servicemonitor.yaml
@@ -48,4 +54,26 @@ resources:
   - loki-servicemonitor.yaml
   - smartctl-exporter-servicemonitor.yaml
   - intel-gpu-exporter-servicemonitor.yaml
+  # ESO's controller metrics. Needs the `monitoring` ingress grant added to
+  # infrastructure/network-policies/external-secrets-ns.yaml in the same change,
+  # or it scrapes nothing. See the file.
+  - external-secrets-podmonitor.yaml
+  # The alert built on what that PodMonitor collects. Split from it on purpose:
+  # the `for:` window had to be chosen against real series, and Prometheus held
+  # none for this metric until the scrape landed.
+  - prometheusrule-external-secrets.yaml
+  # Flux's own control plane. The controllers were never scraped, so nothing
+  # about Flux's behaviour was queryable. Scrape only: per-object readiness is
+  # not on this endpoint at the pinned controller versions and needs
+  # kube-state-metrics instead. See the file.
+  - flux-podmonitor.yaml
+  # The readiness half the PodMonitor above cannot provide. Reads
+  # `gotk_resource_info`, which comes from the kube-state-metrics
+  # customResourceState block in apps/monitoring/helmrelease-prometheus.yaml —
+  # these two files must land and be reverted together.
+  - prometheusrule-flux.yaml
   - probes-blackbox.yaml
+  # TEMPORARY, REN-140. Negative-test probes for the REN-28 §7 bar, taking the
+  # "reachable before" reading that stops being takeable once the paperless PR
+  # merges. Delete this line and the file once the "after" reading is recorded.
+  - probes-ren28-negative.yaml

kubernetes/apps/monitoring-rules/probes-blackbox.yaml

@@ -88,6 +88,179 @@ spec:
         # /ready answers 200 unauthenticated and is in this namespace, so the
         # same intra-namespace allowance the two above rely on covers it.
         - http://loki.monitoring.svc.cluster.local:3100/ready
+        # APP-HEALTH probes for the two gated navidrome-namespace apps. Same
+        # rule as the block above (monitoring-vantage-point-rule), but these
+        # are the first targets here that leave the monitoring namespace.
+        #
+        # Both sit entirely behind an Authentik outpost — navidrome.renner.uno
+        # on ak-outpost-navidrome-gate
+        # (../navidrome/ingress.yaml#navidrome-gate), lidarr.renner.uno on
+        # ak-outpost-lidarr-gate (../navidrome/lidarr.yaml#lidarr-gate) — so an
+        # EXTERNAL probe of either name measures the gate, not the app, and
+        # stays green while the app behind
+        # a live gate is dead. That external signal is still wanted, and REN-2
+        # adds it from UptimeRobot; these are the other half of the pair.
+        #
+        # NO CiliumNetworkPolicy change is needed — but NOT because
+        # cnp-cross-namespace-allow-apps-gatus-to-monitored-services fails to
+        # apply here. It applies: its Generalization section names
+        # blackbox-exporter by hand. The reason we are safe is that the allow
+        # that note demands is PRESENT AND DEDICATED — the navidrome baseline
+        # CNP (infrastructure/network-policies/navidrome-ns.yaml) admits
+        # blackbox-exporter by pod label on exactly the two listeners these
+        # two probes reach, :4533 and :8686. It used to admit the whole
+        # monitoring namespace on every port; REN-28 narrowed it, and the
+        # ports in that rule are the ones these probes depend on, so a probe
+        # added here against a THIRD listener in that namespace would need
+        # that rule widened first. Run that check for the next namespace you
+        # add; do not skip it on the grounds that the note "is about gatus".
+        #
+        # EGRESS — the half the wiki note is silent on, and the half that is
+        # actually load-bearing here. Cilium enforces at BOTH ends, so the
+        # destination ingress allow above is worthless if the prober itself
+        # is in egress enforcement without a matching rule. It is not: an
+        # endpoint only enters egress enforcement once some policy selecting
+        # it declares an `egress:` block, and none does for monitoring. The
+        # namespace has exactly two Cilium policies and both are ingress-only:
+        # infrastructure/network-policies/monitoring-ns.yaml
+        # (default-deny-ingress) and monitoring-alloy-syslog.yaml
+        # (allow-syslog-world-ingress). blackbox-exporter runs in that same
+        # namespace (../monitoring/helmrelease-blackbox-exporter.yaml#namespace).
+        # Re-verified at 2b71dda1 across all 635 yaml files in the repo — 73
+        # Cilium objects, 38 carrying an egress or egressDeny block, none of
+        # them in the monitoring namespace, and the single
+        # CiliumClusterwideNetworkPolicy is scoped to `reserved.health` so it
+        # cannot select a pod. 17 other namespaces here already have an egress
+        # block. If a routine hardening pass ever adds one to monitoring, it
+        # MUST list the navidrome and paperless namespaces or these three
+        # cross-namespace targets go red while the *.monitoring.svc ones above
+        # keep working — a failure mode nobody will connect back to the
+        # hardening commit. The cnp-cross-namespace-* notes document only the
+        # ingress half; that silence is what makes this the trap.
+        #
+        # BLAST RADIUS if that egress dependency is ever broken. Nothing
+        # user-facing breaks: these are Prometheus-side Probe targets, not
+        # gatus checks, so status.renner.uno is unaffected and no workload is
+        # restarted. What breaks is the signal, and it breaks misleadingly.
+        # All three targets start failing BlackboxProbeFailing
+        # (prometheusrule-blackbox.yaml#blackbox-probe-failing — severity
+        # warning, for: 10m); none of the three is in that rule's exclusion
+        # list. BlackboxProbesAbsent does NOT fire, because the other 17
+        # targets keep producing
+        # probe_success — so it reads as three apps going down, not as a
+        # policy change. WHO NOTICES: warning routes to the default `email`
+        # receiver, admin@renner.uno (helmrelease-prometheus.yaml alertmanager
+        # route, group_wait 30s, grouped by alertname), so the owner gets one
+        # email roughly 10-11 minutes after the policy lands. HOW FAST UNDONE:
+        # the REVERT recipe below is a one-line deletion; from merge it is
+        # live within ~11 minutes (GitRepository interval 1m0s,
+        # monitoring-rules Kustomization interval 10m), or immediately with
+        # `flux reconcile kustomization monitoring-rules --with-source`.
+        #
+        # /ping on both: unauthenticated, and already the pods' own kubelet
+        # probe path (../navidrome/deployment.yaml#kubelet-probes,
+        # ../navidrome/lidarr.yaml#kubelet-probes). For lidarr that it is
+        # ungated is proven independently of the gate — its reconcile script
+        # (../navidrome/lidarr.yaml#reconcile-script) curls /ping in a
+        # loop BEFORE it has read the API key out of config.xml.
+        #
+        # STATUS CODE, unproven for all three of the new targets: module
+        # http_2xx follows redirects and passes only on a FINAL 2xx, while
+        # every piece of repo evidence behind these paths establishes only
+        # `< 400` — a kubelet httpGet accepts 200-399, and that reconcile
+        # script's `curl -f` without -L exits 0 on a 302. That evidence proves the
+        # paths are UNGATED, not that they answer 200. Check
+        # probe_http_status_code on the first scrapes; a target that settles
+        # on a 3xx gets REVERTED, not widened to accept 3xx.
+        #
+        # REVERT — an executed recipe, not a claim. Every command below was
+        # actually run against master 2b71dda1 on 2026-09-25; the results are
+        # what is recorded here.
+        #
+        # NO `git revert` works. All four candidates conflict in THIS file at
+        # 2b71dda1, because the comment blocks around these targets were
+        # rewritten after the probes landed (#979):
+        #
+        #   git revert -m 1 0a82d61f   merge of #977, paperless     CONFLICT
+        #   git revert -m 1 16238d39   merge of #976, nav + lidarr  CONFLICT
+        #   git revert cb9e73bd        the paperless commit         CONFLICT
+        #   git revert 2219c928        the nav + lidarr commit      CONFLICT
+        #
+        # Hand-resolving any of them is the trap: the natural --theirs
+        # resolution of the 16238d39 conflict also drops the paperless target
+        # below, leaving NO app-health probes and no signal that it happened.
+        #
+        # Delete the target lines forward instead. Only the target lines are
+        # part of the object; the comments are inert. Both of these were run
+        # and verified:
+        #
+        #   # all three app-health targets — blackbox-http 20 -> 17
+        #   sed -i '/svc\.cluster\.local:80\/ping$/d;/paperless\.paperless\.svc\.cluster\.local:80\/accounts\/login\/$/d' \
+        #     kubernetes/apps/monitoring-rules/probes-blackbox.yaml
+        #
+        #   # the paperless target alone — blackbox-http 20 -> 19
+        #   sed -i '/paperless\.paperless\.svc\.cluster\.local:80\/accounts\/login\/$/d' \
+        #     kubernetes/apps/monitoring-rules/probes-blackbox.yaml
+        #
+        # Verified after each run: the file still parses, blackbox-http drops
+        # to 17 and to 19 respectively, blackbox-edge-wan (3) and
+        # blackbox-tcp-relays (2) are untouched, and `git diff` contains
+        # nothing but the deleted target lines. Then drop whichever comment
+        # block is left orphaned, open a PR, and merge — timing as in the
+        # BLAST RADIUS paragraph above.
+        - http://navidrome.navidrome.svc.cluster.local:80/ping
+        - http://lidarr.navidrome.svc.cluster.local:80/ping
+        # anchor: paperless-app-health
+        # APP-HEALTH probe for paperless. Same rule as the block above
+        # (monitoring-vantage-point-rule); the blind spot it closes is the one
+        # gatus already documents in its Web Service 11 comment
+        # (../gatus/configmap.yaml#gatus-endpoints) — through the
+        # paperless-gate outpost EVERY path, /api/ included,
+        # answers 302 to an unauthenticated caller, so no external probe of
+        # paperless.renner.uno can tell a live app from a live gate. That
+        # external signal is still wanted, and REN-2 adds it from UptimeRobot;
+        # this is the other half of the pair.
+        #
+        # That gatus comment says the in-cluster probe "needs a
+        # CiliumNetworkPolicy first". That sentence is simply WRONG, not
+        # merely inapplicable:
+        # ../../infrastructure/network-policies/paperless-ns.yaml#monitoring-admit
+        # already admits the whole monitoring namespace, which is where
+        # blackbox-exporter runs,
+        # and Cilium unions the allow rules of every CNP selecting an
+        # endpoint (the mechanism ../paperless/network-policy.yaml#cnp-union
+        # relies on). The EGRESS paragraph in the block above applies here
+        # identically. Nothing in the paperless namespace is edited by this
+        # change.
+        #
+        # The stale sentence is deliberately left in place in
+        # apps/gatus/configmap.yaml: it sits INSIDE the `config.yaml: |`
+        # block scalar, so editing it changes ConfigMap data and
+        # `configmap.reloader.stakater.com/reload`
+        # (../gatus/deployment.yaml#reloader) rolls the gatus Deployment. The
+        # correction is written at the top of
+        # that file instead, outside `data:`, where it costs no restart.
+        #
+        # /accounts/login/ because paperless-ngx has no unauthenticated health
+        # endpoint at all. It is the app's own established "the webserver
+        # answers" signal — the same path its startup, liveness and readiness
+        # probes use (../paperless/paperless.yaml#http-probes) — and it
+        # renders a Django view
+        # rather than a static file. It does NOT cover the consumer/celery
+        # workers: see process-liveness-vs-protocol-liveness.
+        #
+        # The cluster-local Host header is accepted only because
+        # PAPERLESS_ALLOWED_HOSTS is "*"
+        # (../paperless/paperless.yaml#allowed-hosts). Narrow that and
+        # Django answers 400 here and this probe goes permanently red.
+        #
+        # This is the target most likely to fail the STATUS CODE check in the
+        # block above — a Django login view can answer 302 — so check it
+        # first, and revert rather than widen if it does. Use the REVERT
+        # recipe in that block: `git revert -m 1 0a82d61f` conflicts here too
+        # at 2b71dda1, and the tested one-line deletion for this target alone
+        # is the second command listed there (20 -> 19 targets).
+        - http://paperless.paperless.svc.cluster.local:80/accounts/login/
 # WAN-path probes of the public edge BY PROXY IP. The name-based
 # home.renner.uno probe above resolves in-cluster (CoreDNS/k8s_gateway)
 # and never leaves the LAN — the vantage-point blind spot that let the
@@ -150,3 +323,80 @@ spec:
       static:
         - proxy-1.renner.uno:33080
         - proxy-2.renner.uno:33080
+---
+# NEGATIVE CONTROLS for the paperless ingress scoping (REN-28 PR 2a).
+#
+# These four targets are here TO BE REFUSED. They are not availability
+# monitoring: every one of them is a workload in the `paperless` namespace that
+# blackbox-exporter has no business reaching, and each must read
+# `probe_success == 1` today and `0` once REN-28's PR 2b lands. That flip is the
+# test. Without them, "the whole-namespace admit is gone" is an assertion about
+# a CNP diff rather than a measured property of the cluster, and a later
+# re-widening of the grant would be silent.
+#
+# Why this can only be the ingress policy under test: blackbox-exporter runs in
+# `monitoring`, whose baseline CNP (infrastructure/network-policies/
+# monitoring-ns.yaml) is ingress-only — it declares no `egress:` at all — and
+# the cluster's default-deny.yaml is scoped `namespace: apps`, which covers
+# neither `monitoring` nor `paperless`. So the only thing that can turn these
+# red is paperless's own ingress policy, which is exactly the variable. Nothing
+# else can make the test pass or fail.
+#
+# WHY A SEPARATE jobName, and why these must NEVER be folded into the
+# blackbox-http Probe above: after PR 2b these targets fail permanently and BY
+# DESIGN. BlackboxProbeFailing and BlackboxProbesAbsent
+# (prometheusrule-blackbox.yaml#blackbox-probe-failing and
+# prometheusrule-blackbox.yaml#blackbox-probes-absent) are both scoped to
+# job="blackbox-http", so a distinct job is outside both and stays silent — and
+# the hand-maintained instance!~ exclusion regex in the first of them, whose
+# fragility that file warns about, does not have to grow a fifth entry. Adding
+# these to blackbox-http instead would
+# page the on-call forever.
+#
+# The inverted alert that gives these targets teeth — CNPNegativeControlReachable,
+# `probe_success{job="blackbox-negative-control"} == 1` — ships in PR 2b, NOT
+# here. Between 2a and 2b these controls are SUPPOSED to read 1, so the alert
+# would fire on the correct state. Do not add it before the policy change.
+#
+# A bare TCP connect is the right probe: it completes or it times out, with no
+# application-layer ambiguity about whether a 403 meant the policy or the app.
+# Known, accepted side effect: the pprl-pg-rw connect opens and closes without a
+# startup packet, so Postgres logs "incomplete startup packet" at LOG level once
+# a minute until PR 2b lands and the connect stops arriving.
+apiVersion: monitoring.coreos.com/v1
+kind: Probe
+metadata:
+  name: blackbox-negative-control-paperless
+  namespace: monitoring
+  labels:
+    # Load-bearing, as noted at the top of this file: probeSelector matches on
+    # it. Without it this Probe is never discovered and the control reads
+    # `absent`, not `1` — which would look like a passing negative test.
+    release: kube-prometheus-stack
+spec:
+  jobName: blackbox-negative-control
+  prober:
+    url: blackbox-exporter-prometheus-blackbox-exporter.monitoring.svc.cluster.local:9115
+  module: tcp_connect
+  interval: 60s
+  scrapeTimeout: 10s
+  targets:
+    staticConfig:
+      static:
+        # Proves BOTH halves of the pprl-pg scoping PR 2b will introduce: the
+        # source (blackbox is not Prometheus, and the pprl-pg-allow-prometheus
+        # rule 2b adds admits only Prometheus) and the port (5432 is not the
+        # exporter's 9187). Neither that rule nor paperless-allow-blackbox
+        # exists yet at this commit — that is the point, and why this target
+        # must read 1 now. Service is CNPG-generated from Cluster `pprl-pg`
+        # (../paperless/cluster.yaml#pprl-pg) and is the name paperless itself
+        # dials (../paperless/paperless.yaml#db-host).
+        - pprl-pg-rw.paperless.svc.cluster.local:5432
+        # Three witnesses that the whole-namespace `monitoring` admit is gone
+        # for workloads no new rule selects. Ports from the Services themselves:
+        # ../paperless/redis.yaml#service-port,
+        # ../paperless/tika.yaml#service-port,
+        # ../paperless/gotenberg.yaml#service-port.
+        - redis.paperless.svc.cluster.local:6379
+        - tika.paperless.svc.cluster.local:9998
+        - gotenberg.paperless.svc.cluster.local:3000

kubernetes/apps/monitoring-rules/probes-ren28-negative.yaml

@@ -0,0 +1,100 @@
+# TEMPORARY — negative-test probes for REN-28 §7. DELETE AFTER THE "AFTER"
+# READING IS TAKEN. This file exists to make two of the four §7 negative tests
+# measurable without granting `pods/exec` in `monitoring`, which the owner
+# declined on REN-140 as a wider grant than the change it would verify.
+#
+# WHAT IT IS. blackbox-exporter is already an HTTP/TCP prober running in the
+# `monitoring` namespace. Pointing it at targets that the REN-28 paperless PR
+# is supposed to CLOSE turns "is this path still open?" into a number in
+# Prometheus, readable through the existing read-only REN-36 grant. Every
+# target below answers today and MUST stop answering after the PR lands. A
+# `probe_success` that stays 1 after the merge means the rule does not do what
+# §5 claims.
+#
+# WHY IT DOES NOT PAGE ANYONE. jobName is `blackbox-ren28-negative`, not
+# `blackbox-http`. Every rule in prometheusrule-blackbox.yaml is scoped to
+# job="blackbox-http" (BlackboxProbeFailing) or absent(...{job="blackbox-http"})
+# (BlackboxProbesAbsent), so neither can match these series — before or after.
+# TargetDown (general.rules) cannot match either: the scrape target of a Probe
+# is blackbox-exporter itself, which answers 200 with probe_success 0 on a
+# failed probe, so `up` stays 1 regardless of the result. Verified against the
+# live rule set on 2026-09-26. This is deliberate: a monitor that pages for an
+# expected failure is worse than no monitor.
+#
+# NO POLICY CHANGE IS NEEDED TO MAKE THESE WORK TODAY. Ingress: the paperless
+# baseline still admits the whole `monitoring` namespace on every port
+# (../../infrastructure/network-policies/paperless-ns.yaml#monitoring-admit) —
+# that admission is
+# exactly what REN-28 removes. Egress: the `monitoring` namespace is in no
+# egress enforcement at all; its only two Cilium policies, monitoring-ns.yaml
+# and monitoring-alloy-syslog.yaml, are ingress-only. Same dependency the three
+# cross-namespace targets in probes-blackbox.yaml already rest on.
+#
+# TWO POD-IP TARGETS. pprl-pg's :9187 metrics port is on the POD, not on any
+# Service — that is why it is scraped by a PodMonitor and not a ServiceMonitor
+# (cnpg-podmonitors.yaml). There is no stable name for it, so the literal pod
+# IPs are used. They were read live on 2026-09-26; if either pod is rescheduled
+# the probe goes red for the wrong reason. BEFORE TRUSTING AN "AFTER" READING,
+# re-confirm the IP still belongs to the same pod:
+#
+#   kube_pod_info{namespace="paperless", pod=~"pprl-pg-.*"}   -> pod_ip
+#
+# The pprl-pg-r:5432 target below is the stable-name version of the same
+# statement and survives a reschedule; it is on the database port rather than
+# the metrics port, so the two are kept together on purpose.
+#
+# REVERT. Delete this file and its line in kustomization.yaml. Nothing else
+# references it; no workload is touched; Prometheus drops the job on the next
+# operator reconcile.
+apiVersion: monitoring.coreos.com/v1
+kind: Probe
+metadata:
+  name: blackbox-ren28-negative
+  namespace: monitoring
+  labels:
+    # probeSelector is NOT loosened on this cluster (helmrelease-prometheus.yaml
+    # keeps ruleSelector/probeSelector/scrapeConfigSelector strict), so this
+    # label is load-bearing, not decorative. Without it the Probe is created and
+    # never discovered — and a test that silently does not run reads exactly
+    # like a test that passed.
+    release: kube-prometheus-stack
+spec:
+  jobName: blackbox-ren28-negative
+  prober:
+    url: blackbox-exporter-prometheus-blackbox-exporter.monitoring.svc.cluster.local:9115
+  module: tcp_connect
+  interval: 60s
+  # 10s matches blackbox-tcp-relays. It is also the discriminator: a Cilium
+  # deny DROPS the SYN, so a refused path completes in ~1ms and a denied path
+  # burns the full timeout. probe_duration_seconds distinguishes "nothing is
+  # listening" from "policy stopped it" — probe_success alone does not.
+  scrapeTimeout: 10s
+  targets:
+    staticConfig:
+      static:
+        # REN-28 §7 test 4 — "from a monitoring pod that is not Prometheus, a
+        # connection to pprl-pg :9187 must time out". blackbox-exporter is
+        # precisely that pod. Both CNPG instances, because the PR's
+        # pprl-pg-allow-prometheus rule selects on a label both carry: if the
+        # selector were wrong in a way that left one instance open, one target
+        # would tell us and the other would not.
+        - 10.244.4.166:9187 # pprl-pg-2, node k8s-2
+        - 10.244.2.220:9187 # pprl-pg-3, node k8s-3
+        # Same statement on a stable name and the port that actually matters:
+        # after the PR the only monitoring→pprl-pg allow is Prometheus on
+        # :9187, so a non-Prometheus monitoring pod must not reach Postgres.
+        - pprl-pg-r.paperless.svc.cluster.local:5432
+        # REN-28 §7 test 2, workload dimension — "from blackbox-exporter, a
+        # paperless target other than the one it is scoped to must time out".
+        # The test as written in §7 says "a paperless POD port other than
+        # 8000"; there is no such port. The paperless pod's only other listener
+        # is the bridge-tls-proxy socat, bound to 127.0.0.1:11143
+        # (../paperless/paperless.yaml#bridge-tls-proxy), so every other port
+        # on that pod
+        # is closed today and answers RST — there is no "reachable before" to
+        # record and the port half of paperless-allow-blackbox (:8000) cannot
+        # be tested at all. gotenberg is the half that CAN be: it is one of the
+        # other eight workloads the endpointSelector:{} baseline currently
+        # exposes to all of `monitoring`, and the PR's per-workload selector is
+        # what closes it.
+        - gotenberg.paperless.svc.cluster.local:3000

kubernetes/apps/monitoring-rules/prometheusrule-blackbox.yaml

@@ -31,6 +31,7 @@ spec:
         # target means. A target whose loss deserves a critical page gets its
         # own rule — and lands in the list above. That is what keeps the
         # exclusion list from being arbitrary.
+        # anchor: blackbox-probe-failing
         - alert: BlackboxProbeFailing
           expr: probe_success{job="blackbox-http", instance!~"https://fsn1\\.your-objectstorage\\.com|https://netbird\\.renner\\.uno|http://proxy-[12]\\.renner\\.uno:9000/health|https://ci\\.renner\\.uno|https://openbao\\.renner\\.uno/v1/sys/health.*"} == 0
           for: 10m
@@ -46,6 +47,7 @@ spec:
         # The check that the checks are running. Every rule above reads
         # probe_success; if the series stop being produced at all, none of them
         # can fire and the silence looks exactly like health.
+        # anchor: blackbox-probes-absent
         - alert: BlackboxProbesAbsent
           expr: absent(probe_success{job="blackbox-http"})
           for: 15m

kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml

@@ -52,19 +52,123 @@ spec:
         - alert: CephPoolLostRedundancy
           # Names the pool, which POOL_NO_REDUNDANCY does not.
           #
-          # There is no pool-size metric, but bytes_used/stored IS the
-          # replication factor. The >1e6 floor drops pools too small to measure
-          # meaningfully, empty ones included.
+          # There is no pool-size metric, but for a REPLICATED pool
+          # bytes_used/stored IS the replication factor. The >1e6 floor drops
+          # pools too small to measure meaningfully, empty ones included.
           #
-          # No pool is exempt.
+          # ERASURE-CODED POOLS ARE EXCLUDED, AND MUST STAY EXCLUDED (REN-88).
+          #
+          # That ratio is not a replication factor for an EC pool, it is the
+          # coding overhead: k+m shards over k data shards. For the two ec:2+1
+          # pools here that is 3/2 = exactly 1.5 — sitting precisely on this
+          # rule's own threshold, on the wrong side of it half the time.
+          # Measured on `cephfs-bulk` (pool 15): used 480042086400 / stored
+          # 320028049408 = 1.5000000384, a margin of 12 KiB in 480 GB, so which
+          # way the comparison lands is decided by the last bits of a float
+          # division. It paged critical for 199 minutes on 2026-09-25 (12:37 to
+          # 15:56 UTC, two emails) with the cluster HEALTH_OK, zero degraded
+          # PGs and every OSD up, then cleared on its own. 50 of the 210
+          # samples in the preceding 30h were below 1.5; only that run held the
+          # 30m `for`. `migration.rgw.buckets.data` (pool 12) is the same ec:2+1
+          # shape and was one drift away from doing the same thing.
+          #
+          # The premise fails for EC too, not just the arithmetic: the old
+          # description claimed a single OSD loss destroys data, but 2+1 carries
+          # a parity shard and survives one OSD loss, same as replica-2.
+          #
+          # Do NOT "fix" this by nudging the threshold to 1.4 or 1.49. That
+          # keeps a knife-edge, just a quieter one, and it also narrows the real
+          # replica-2 -> replica-1 signal this rule exists to catch (1.0).
+          #
+          # EC pools keep full redundancy coverage without this rule: a real
+          # loss raises POOL_NO_REDUNDANCY, which is a live series in
+          # `ceph_health_detail` and is deliberately in scope for
+          # CephHealthWarn above. That check reads Ceph's own verdict instead of
+          # inferring one from a byte ratio, which is the right instrument here
+          # — an EC pool that loses a shard does not change its stored/used
+          # ratio at all, because Ceph does not delete the surviving shards.
+          # So there is no gap to reopen, and no second EC-specific threshold
+          # worth inventing.
+          #
+          # `on(pool_id)` is deliberately the whole matcher. ceph_pool_metadata
+          # carries labels ceph_pool_bytes_used does not (name, type,
+          # description, application, compression_mode), so default all-label
+          # matching would never join. Adding `cluster` would be defensive
+          # against a second Ceph cluster that does not exist, at the cost of
+          # the rule silently matching nothing if that label is ever dropped —
+          # an alert that fails open invisibly is the worse trade.
           expr: |
-            ceph_pool_bytes_used / ceph_pool_stored < 1.5 and ceph_pool_stored > 1e6
+            ceph_pool_bytes_used / ceph_pool_stored < 1.5
+              and on(pool_id) ceph_pool_metadata{type="replicated"} == 1
+              and ceph_pool_stored > 1e6
           for: 30m
           labels:
             severity: critical
           annotations:
-            summary: "Ceph pool {{ $labels.pool_id }} is running unreplicated"
-            description: "Pool {{ $labels.pool_id }} has a replication factor below 1.5, meaning a single OSD loss destroys data. No pool is meant to be unreplicated — an unreplicated pool also blocks ALL OSD maintenance cluster-wide, since `ceph osd ok-to-stop` refuses while its PGs would go offline. Check with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail`."
+            summary: "Ceph replicated pool {{ $labels.pool_id }} is running unreplicated"
+            description: "Replicated pool {{ $labels.pool_id }} has a replication factor below 1.5, meaning it is down to a single copy and one OSD loss destroys data. No replicated pool is meant to be unreplicated — it also blocks OSD maintenance for as long as it lasts, since `ceph osd ok-to-stop` refuses while its PGs would go offline. Erasure-coded pools are out of scope for this rule and are covered by POOL_NO_REDUNDANCY via CephHealthWarn (REN-88). Check with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail`."
+
+        - alert: CephPoolLostRedundancyBlind
+          # Canary for the rule directly above. Keep the two adjacent.
+          #
+          # CephPoolLostRedundancy is `critical` and it fails CLOSED and
+          # SILENTLY. Its expression joins three metrics on one label and
+          # filters on one specific label VALUE, and if any of those stops
+          # matching the expression returns nothing, which is indistinguishable
+          # from "every pool is fine". Health stays ok, no email is sent, and
+          # nothing anywhere reports that a critical storage alert has gone
+          # dark. Measured at REN-88 review time: `type` carried exactly two
+          # values, `replicated` and `erasure`, across 15 series — so the
+          # dependency is real, and a Rook or Ceph upgrade that renames the
+          # label, renames the value, or drops it in a relabel rule is enough.
+          #
+          # This expression is CephPoolLostRedundancy's with both thresholds
+          # removed, wrapped in absent(): "can that rule produce any series at
+          # all". It goes to 1 if either metric disappears, if `pool_id` is
+          # dropped from either side of the join, if the `type` label is
+          # renamed, or if `replicated` stops being its value.
+          #
+          # The one deliberate difference is that `ceph_pool_stored > 1e6` is
+          # NOT mirrored here. That floor is about data volume, not structure —
+          # replicated pools that are all legitimately tiny is a quiet state,
+          # not a broken dependency, and copying it in would make this fire for
+          # a benign reason.
+          #
+          # ONE CANARY PER FRAGILE JOIN, NOT ONE PER RULE GROUP. Judged on
+          # drift: a group-level canary has to enumerate every metric the group
+          # reads, and adding a rule to the group does not update it. Six months
+          # on it silently under-covers while LOOKING comprehensive, which is the
+          # same failure shape as a discovery step that passes having checked
+          # nothing. This one's scope is exactly one rule sitting six lines
+          # above it, so "is it still true" is answerable by reading the two
+          # blocks together. It can still drift — someone edits the expr above
+          # and not this — but that drift shows up in the same diff hunk, and
+          # when it does bite it FIRES rather than going quiet. For a guard,
+          # loud and wrong beats silent and wrong.
+          #
+          # 1h, not the 30m above. The exporter moved through 17 distinct mgr
+          # instances in 30d, and every one of those failovers makes all ceph_*
+          # metrics absent for as long as it takes the new mgr to be scraped;
+          # anything under ~15m would page on routine failover, and noise is
+          # what REN-88 was about. A long `for` costs nothing in detection value
+          # because the failure this catches is a schema change — permanent, not
+          # transient — and a blindness that self-heals inside an hour is not
+          # worth an email. Transient exporter loss is already covered by the
+          # chart's own TargetDown.
+          #
+          # `warning`, not `critical`: an alert stopped being able to fire.
+          # Storage is not broken.
+          expr: |
+            absent(
+              ceph_pool_bytes_used / ceph_pool_stored
+                and on(pool_id) ceph_pool_metadata{type="replicated"} == 1
+            )
+          for: 1h
+          labels:
+            severity: warning
+          annotations:
+            summary: "CephPoolLostRedundancy cannot fire — its metrics no longer join"
+            description: "The CephPoolLostRedundancy alert has been unable to produce any series for an hour, so a `critical` storage alert is silently dark. It joins ceph_pool_bytes_used / ceph_pool_stored against ceph_pool_metadata{type=\"replicated\"} on `pool_id`; one of those metrics, the `pool_id` label, the `type` label, or the `replicated` value has stopped matching — most likely a Rook or Ceph upgrade, or a relabel rule. Check what the exporter emits now with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail` and by querying `ceph_pool_metadata` directly, then fix the join in kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml — do not delete this canary (REN-93)."
 
         - alert: CephHealthError
           # HEALTH_ERR means data availability or redundancy is at risk.
@@ -95,3 +199,116 @@ spec:
           annotations:
             summary: "{{ $labels.instance }} /var below 30% free — Ceph mon will warn"
             description: "/var on {{ $labels.instance }} is under 30% free, which is Ceph's mon_data_avail_warn threshold, so the mon on this node will report MON_DISK_LOW. Find the actual consumer with `talosctl -n <node> usage /var/lib --depth 1` and then `/var/lib/containerd --depth 1` — note that overlayfs snapshots belong to running containers and are not reclaimable by image GC."
+
+        - alert: CephOSDNearfullApproaching
+          # Warns BEFORE Ceph's own nearfull check, which is the incident and
+          # not a warning. mon_osd_nearfull_ratio is 0.85 and is not
+          # overridden here.
+          #
+          # There is no rebalance to hide behind: failureDomain is host with
+          # size 3 on exactly three OSDs, so every OSD holds a copy of every
+          # replicated object and all three cross 0.85 within seconds of each
+          # other. Measured 2026-09-26: 81.86 / 81.87 / 81.83%.
+          #
+          # 0.75 is 0.85 minus the largest bulk-tier burst on record. Measured
+          # at 30s resolution over the 14d to 2026-09-26: the fullest OSD went
+          # 71.63% -> 81.87% between 05:18 and 05:43Z (+10.24 pp, ~275 GiB raw)
+          # and reclaimed within the hour. At a SUSTAINED 75%, one more burst of
+          # that size reaches nearfull, which is the point at which the level
+          # itself is the finding.
+          #
+          # `for: 30m` is the shortest wait that outlives a self-reclaiming
+          # burst. That same burst held >= 75% for 19.0 min, so 30m clears it by
+          # 11 min; a permanent step holds it indefinitely. Do NOT shorten this
+          # to 15m — the 09-26 burst then pages, with the cluster reclaiming on
+          # its own while someone reads the mail.
+          #
+          # Nor should it be lengthened to 1h. The fastest observed SUSTAINED
+          # climb was +11.44 pp in one hour (2026-09-25 08:02-09:02Z, a real
+          # ingest that never reclaimed). From 75% that reaches 0.85 in ~53 min,
+          # so a 1h pending window would fire after nearfull, not before it.
+          # 30m fires at ~80.7% in that worst case — about 22 min of lead — and
+          # at 2.24 pp/h (the worst 6h-sustained rate) it fires at ~76.1%, about
+          # 4h of lead. The 30m/1h choice is the whole trade: a burst reclaims
+          # in ~20 min and a step does not, so ~30m is the earliest honest
+          # verdict. REN-162 is measuring the burst envelope; if a burst is ever
+          # seen holding >= 75% for more than ~25 min, this `for` moves out, not
+          # the threshold.
+          #
+          # Replay over the 14d to 2026-09-26 08:00Z: the worst 30m-sustained
+          # floor on any OSD was 71.632% (09-26 04:31Z), so this rule fires ZERO
+          # times on the observed normal shape, including the 09-26 excursion to
+          # 81.87% and the 09-25 step from 59% to 71%. Margin 3.37 pp.
+          #
+          # `max by (ceph_daemon)` is not cosmetic. rook-ceph-exporter pods are
+          # replaced often — 32 distinct series for three OSDs in 14 days — and
+          # the raw expression carries `pod` and `instance`, so every restart
+          # would change the alert identity and reset a pending window.
+          expr: |
+            max by (ceph_daemon) (
+              ceph_osd_stat_bytes_used / ceph_osd_stat_bytes
+            ) >= 0.75
+          for: 30m
+          labels:
+            severity: warning
+          annotations:
+            summary: "Ceph {{ $labels.ceph_daemon }} sustained above 75% — one bulk burst from nearfull"
+            description: "{{ $labels.ceph_daemon }} has been at or above 75% utilisation for 30 minutes ({{ $value | humanizePercentage }} now). That is 0.85 nearfull minus the largest bulk-tier burst measured here (+10.24 pp in 25 minutes), so a single repeat of that burst trips mon_osd_nearfull_ratio — and with failureDomain host, size 3 and three OSDs, it trips on all three at once with no rebalance available. This is sustained, not a spike: a burst clears in about 20 minutes and never holds this for 30. Find what grew with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph df` and `ceph osd df tree`; the bulk tier (cephfs-bulk, pool 15) is the usual mover."
+
+        - alert: CephReplicapoolHeadroomLow
+          # The `replicapool` counterpart to the OSD rule above, in the units a
+          # reader can act on: GiB that still fit.
+          #
+          # These two are NOT independent signals and must not be read as
+          # corroboration. `ceph_pool_max_avail` for a size-3 pool on three
+          # equal OSDs is exactly (full_ratio - fullest_OSD_utilisation) x
+          # osd_size; verified against live data, the identity holds at 0.95 to
+          # five decimals. So this rule is the OSD series rescaled — a second
+          # TIER, deliberately set to trip after the OSD rule, not a second
+          # opinion. Ceph's own nearfull line of 0.85 sits at 89.4 GiB here.
+          #
+          # 160 GiB is what must still fit, measured:
+          #   89 GiB  max_avail at the 0.85 nearfull line, i.e. the part of
+          #           max_avail that is not ours to spend
+          #   38 GiB  ha-pg's unconsumed allocation after the approved 20Gi ->
+          #           32Gi expansion: 2 instances x (31.33 GiB provisioned -
+          #           12.4 GiB used) as of 2026-09-26
+          #   35 GiB  replicapool churn allowance — the worst observed single
+          #           day (+14.3 GiB, 09-23) plus eleven days at the 14d mean
+          #           of 1.9 GiB/day
+          #   = 162 GiB, rounded DOWN to 160 so the rounding costs quiet, not
+          #     headroom.
+          #
+          # Selecting on the pool NAME through ceph_pool_metadata, not on a
+          # hardcoded pool_id: ids are assignable and replicapool is only id 1
+          # by history. `on(pool_id)` is the whole matcher for the same reason
+          # given on CephPoolLostRedundancy above — ceph_pool_metadata carries
+          # labels ceph_pool_max_avail does not, so default matching never
+          # joins.
+          #
+          # `max by (pool_id)` twice over: it collapses the churning mgr `pod`
+          # and `instance` labels (8 series in 14 days), and during an mgr
+          # failover it takes the live value rather than a stale low one.
+          #
+          # `for: 30m` also covers a real exporter artefact: ceph_pool_max_avail
+          # reads exactly 0 for a single 30s sample after an mgr restart, twice
+          # in the 14d window (e.g. 09-23 12:27Z), on every mgr series at once.
+          # A `for` shorter than a few minutes would page on that alone. Note
+          # the glitch is one-directional, so it cannot cancel a genuine breach
+          # — 0 is still below the floor.
+          #
+          # Replay over the 14d to 2026-09-26 08:00Z: the lowest 30m-sustained
+          # ceiling was 208.97 GiB (09-26 04:31Z), so ZERO fires on the observed
+          # normal shape. The 09-26 burst took max_avail to 117.4 GiB for about
+          # 10 minutes and did not come close to holding it.
+          expr: |
+            max by (pool_id) (
+              ceph_pool_max_avail
+                and on(pool_id) ceph_pool_metadata{name="replicapool"}
+            ) < 160 * 1024^3
+          for: 30m
+          labels:
+            severity: warning
+          annotations:
+            summary: "replicapool usable headroom below 160 GiB"
+            description: "replicapool has had under 160 GiB usable for 30 minutes ({{ $value | humanize1024 }}B now). 160 GiB is the 89 GiB that max_avail still holds above Ceph's 0.85 nearfull line, plus the 38 GiB ha-pg has not yet consumed of its 32Gi expansion, plus 35 GiB of observed replicapool churn — below this, what is already committed no longer fits before nearfull. This is the same series as CephOSDNearfullApproaching rescaled (max_avail = (0.95 - fullest OSD) x OSD size), so it is a second tier and not independent confirmation. Check `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph df`; growth here is usually the bulk tier on the shared OSDs, not replicapool itself."

kubernetes/apps/monitoring-rules/prometheusrule-ci-openbao.yaml

@@ -34,9 +34,20 @@ spec:
           # query parameter, and an exact match would break silently if the
           # instance label ever encodes it differently.
           expr: probe_success{job="blackbox-http", instance=~"https://openbao\\.renner\\.uno/v1/sys/health.*"} == 0
-          # Two consecutive failed probes: enough to skip a single lost probe
-          # through Traefik, not enough to sit on a sealed store.
-          for: 1m
+          # Five minutes of continuously failing probes — about ten consecutive
+          # samples, since the probe and the rule evaluation both run at 30s
+          # (probes-blackbox.yaml, helmrelease-prometheus.yaml). Matches
+          # WoodpeckerDown above.
+          #
+          # This was 1m, the tightest probe window in the cluster by 5x, and it
+          # went pending on a single 5s HTTPS timeout with OpenBao healthy
+          # (REN-131). Isolated blips are ambient on this probe fleet at roughly
+          # one per two hours, so a window that clears on three samples pages on
+          # jitter. The widening costs nothing real: per the description below
+          # nothing breaks the moment OpenBao goes away, and OpenBaoQuorumLost
+          # (prometheusrule-openbao.yaml, for: 2m) still covers the fast
+          # failure, which is quorum loss rather than a lost probe.
+          for: 5m
           labels:
             severity: critical
           annotations:

kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml

@@ -13,7 +13,7 @@ spec:
       rules:
         # The Postgres clusters back up to Hetzner. HetznerBackupEndpointDown
         # probes that S3 endpoint, which proves the bucket is reachable and
-        # nothing more — these two rules check that something is actually being
+        # nothing more — these rules check that something is actually being
         # written to it.
 
         - alert: CNPGWalArchiveBacklog
@@ -23,22 +23,79 @@ spec:
           # that moment, and pg_wal keeps growing until it fills the PVC and
           # takes Postgres down.
           #
-          # This is the load-bearing alert of the two. It reads the
+          # This is the load-bearing alert of the group. It reads the
           # instance-level collector directly, so unlike the backup-age metric
           # discussed below it does not depend on Cluster status being
           # populated.
           #
-          # Threshold: >5 segments sustained for 30m. Brief non-zero values are
-          # normal under write bursts — a segment is 16 MiB, so 5 is roughly
-          # 80 MiB of un-archived WAL, well below anything that threatens the
-          # volume but far above idle noise.
-          expr: cnpg_collector_pg_wal_archive_status{value="ready"} > 5
-          for: 30m
+          # Threshold: >5 segments sustained for 20m. A segment is 16 MiB, so 5
+          # is roughly 80 MiB of un-archived WAL — well below anything that
+          # threatens the volume, but far above idle noise.
+          #
+          # WHAT THIS RULE KEYS ON: persistence, not depth. The failure that
+          # costs something is an archiver that has *stopped*, and a stopped
+          # archiver can sit at 60 segments indefinitely without ever reaching
+          # some larger number. A depth threshold tuned above the nightly
+          # envelope below would miss exactly that case, and would need raising
+          # again every time the recorder grows. So there is one rule here, and
+          # the tuning knob is `for` (REN-152).
+          #
+          # THE NIGHTLY BURST, which this rule must stay silent through.
+          # ha-pg-1 queues WAL every night on the same clock:
+          #
+          #   flat 0 until 02:12Z; fill from 02:13Z; peak 02:14-02:15Z;
+          #   drained by 02:26Z at the worst. 02:12Z is 04:12 local, which is
+          #   when Home Assistant's recorder purge runs.
+          #
+          # It is a write burst, not IO contention. Drain is a flat
+          # 13-15 segments/min every night — including nights when Kopia's
+          # repository-maintenance batch was saturating the RBD devices, and
+          # nights when it did not start until after the burst had already
+          # drained. That batch drifts across 02:03-02:24Z while the peak stays
+          # pinned to 02:15Z, and a drifting cause cannot produce an effect
+          # fixed to the clock. What varies is the fill (19-40 segments/min),
+          # so the height of the backlog is set by how much WAL is produced
+          # against an archiver with a flat ~14 segments/min ceiling.
+          #
+          # WHY 20m. Measured against Prometheus' full 27-night retention
+          # (2026-08-30..2026-09-26), at the real 30s scrape resolution, across
+          # all 12 CNPG instances. Only ha-pg-1 ever exceeds 5 at all; every
+          # other instance peaks at 2. Longest contiguous excursion above 5 on
+          # any night: 13.0 minutes, on 2026-09-11, at a peak of 141 segments.
+          # Firing evaluations over those 27 nights: 0 at 20m, and also 0 at
+          # 15m — but 15m leaves only 2 minutes of margin against the worst
+          # night, and the nightly peak climbed from 116 to 141 inside the
+          # measured month as the recorder grew. 20m leaves 7 minutes and does
+          # not have to be re-derived every time the database gets bigger.
+          #
+          # The 5 minutes 20m costs in detection latency buy nothing back: a
+          # real stall lasts hours, and at 32Gi (REN-141) the volume holds
+          # roughly four nights of stalled WAL before free space is in danger.
+          #
+          # `for` was 30m before REN-129. Outside the nightly windows the
+          # metric is a hard 0 on all 12 instances, so there was never a noise
+          # floor that needed 30m to wait out — the extra 10m only delayed the
+          # page.
+          #
+          # WHEN IT FIRES, the fix is drain rate or headroom, never a higher
+          # threshold here. ObjectStore `wal.maxParallel`
+          # (kubernetes/apps/home-assistant/ha-pg-objectstore.yaml) is unset,
+          # which the barman-cloud plugin reads as one WAL segment at a time;
+          # that is where the ~14 segments/min ceiling comes from. Headroom is
+          # `spec.storage.size` in ha-pg-cluster.yaml. Both are REN-135.
+          #
+          # `max by (namespace, pod)` is not cosmetic. The raw series carries
+          # `role`, which flips on failover; a changed label is a new series,
+          # and a new series restarts the `for` timer. Aggregating `role` away
+          # means a backlog that persists across a failover still pages 20m
+          # after the backlog started rather than 20m after the failover.
+          expr: max by (namespace, pod) (cnpg_collector_pg_wal_archive_status{value="ready"}) > 5
+          for: 20m
           labels:
             severity: warning
           annotations:
             summary: "CNPG WAL archiving is falling behind on {{ $labels.namespace }}/{{ $labels.pod }}"
-            description: "{{ $value }} WAL segments have been waiting to archive for over 30 minutes. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n {{ $labels.namespace }} logs {{ $labels.pod }} -c plugin-barman-cloud`."
+            description: "{{ $value }} WAL segments have been waiting to archive for over 20 minutes. NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local) is normal and drains itself within about 13 minutes, and this rule is tuned to stay silent through it — so if this fired, the backlog did NOT clear and the archiver is stuck. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n {{ $labels.namespace }} logs {{ $labels.pod }} -c plugin-barman-cloud`."
 
         - alert: CNPGBackupFailed
           # cnpg_collector_last_failed_backup_timestamp is a unix timestamp, 0

kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml

@@ -0,0 +1,227 @@
+# Alert when an ExternalSecret stops syncing.
+#
+# This is the second half of REN-103. The first half (#994) added the PodMonitor
+# and the `monitoring` ingress grant that make the series exist at all; before
+# that, `externalsecret_status_condition` held 0 series and a key that silently
+# stopped refreshing was invisible to Prometheus, to Alertmanager and to anyone
+# looking at a dashboard. The alert is split out because its `for:` window had to
+# be chosen against real data rather than guessed.
+#
+# `exported_namespace`, NOT `namespace`. The metric carries its own `namespace`
+# label (the ExternalSecret's), and the PodMonitor attaches a target label of the
+# same name (the controller pod's, always `external-secrets`). Prometheus resolves
+# that collision by renaming the metric's label to `exported_namespace`, so on
+# every one of the ~105 live series `namespace="external-secrets"`. Grouping by
+# `namespace` would therefore still fire correctly and still be useless: every
+# failing ExternalSecret in the cluster would be reported as living in
+# `external-secrets`. ESO's own shipped Grafana dashboard does group by
+# `namespace` — that is right for the chart's ServiceMonitor, which scrapes ESO
+# through its own Service and injects no conflicting target label, and wrong here.
+# This is a general trap for the PodMonitor-instead-of-ServiceMonitor trade this
+# repo keeps making: any metric carrying its own `namespace`, `pod` or `container`
+# label gets it shadowed.
+#
+# `honorLabels: true` on the PodMonitor endpoint would keep the clean label name
+# instead, and is deliberately not used: it drops the target labels, so nothing
+# records which controller replica reported, and it would make this the only
+# monitor in the directory doing so.
+#
+# A consequence worth knowing rather than fixing: aggregating by
+# `exported_namespace` leaves these alerts with no `namespace` label at all, and
+# Alertmanager's route groups on ['alertname', 'cluster', 'namespace']. So every
+# failing ExternalSecret arrives in ONE email listing all of them, each named by
+# its own annotation, instead of one email per namespace. For a fleet this size
+# that is the better shape.
+#
+# The obvious worry about that shape — that a SECOND ExternalSecret failing while
+# the first alert is already firing gets swallowed by the open group — does not
+# hold, and the setting that decides it is `group_interval: 5m` on the root route
+# in apps/monitoring/helmrelease-prometheus.yaml. `group_interval` governs
+# notification of a group whose alert SET has changed, which is a different thing
+# from `repeat_interval: 12h`, which governs re-sending an unchanged group. A
+# second failure therefore produces a fresh email naming both within 5 minutes.
+# This rule routes to the root `email` receiver: it is `severity: warning` and
+# carries no `source = "flux"`, so it matches none of the four child routes.
+#
+# The inhibit rules are also checked and do not touch these alerts. All three key
+# on `equal: ['namespace', 'alertname']` (or `['namespace']`), and no critical or
+# warning alert in the cluster shares this `alertname`. In particular
+# OpenBaoQuorumLost does NOT inhibit this rule — different alertname — so an
+# OpenBao outage can surface both. That is the intended order rather than a
+# problem: OpenBaoQuorumLost is critical at 2m and arrives first, well before
+# this rule's 15m.
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: external-secrets
+  namespace: monitoring
+  labels:
+    # kube-prometheus-stack's ruleSelector matches on this label; without it the
+    # rule is created but never loaded into Prometheus.
+    release: kube-prometheus-stack
+spec:
+  groups:
+    - name: external-secrets.rules
+      rules:
+        # Measured against the live Prometheus before this rule was proposed for
+        # merge, through the apiserver services/proxy read path:
+        #
+        #   count(externalsecret_status_condition)                       = 105
+        #   count(count by (exported_namespace,name) (...))              =  53
+        #   count(count by (exported_namespace,name) (...{condition="Ready"})) = 52
+        #   count(count by (exported_namespace) (...))                   =  20
+        #   count(...{condition="Ready",status="False"})                 =  52
+        #   the ExternalSecretSyncFailing expression below                = EMPTY
+        #
+        # So 105 = 52 ExternalSecrets x 2 Ready polarities + 1 Deleted tombstone,
+        # across 20 namespaces, and THIS RULE ARRIVES SILENT: it matches nothing
+        # in the cluster as it stands, which is what makes it safe to load. The
+        # 52 on the line below is a count, not an estimate.
+        #
+        # `== 1` IS LOAD-BEARING, not defensive. ESO writes BOTH polarities of the
+        # Ready condition on every reconcile — the true one at 1 and the opposite
+        # at 0 (esmetrics.UpdateExternalSecretCondition) — so
+        # `{condition="Ready",status="False"}` is present with value 0 for every
+        # healthy ExternalSecret. A bare selector would match all 52 of them.
+        #
+        # `max`, not the `sum` the vendor dashboard uses. Both give 1 today,
+        # because ESO runs a single controller replica. During a Deployment
+        # rollout Prometheus can briefly scrape the old and the new pod at once,
+        # and `sum(...) == 1` would then evaluate to 2 and silently STOP matching
+        # — the alert would resolve itself in the middle of a failure. `max` is 0
+        # or 1 whatever the replica count.
+        #
+        # The gauge, not `externalsecret_sync_calls_error`. The counter was the
+        # other candidate, on the theory that a counter cannot flap the way a
+        # condition gauge might. Reading the controller says the opposite on both
+        # halves: the counter increments on strictly MORE paths than the condition
+        # does — a failed Get of the ExternalSecret itself, a failed Get or label
+        # Patch of the target Secret, an unresolvable SecretStore — and none of
+        # those call markAsFailed(), so none of them set Ready=False. A cache blip
+        # or an apiserver hiccup moves the counter and nothing else. It is the
+        # noisier signal, not the quieter one. And the gauge does
+        # not flap, because it is set to 1 only by markAsFailed() and cleared by
+        # the next successful reconcile. There is no unset state in between.
+        #
+        # The live numbers say the same thing, measured after #994 landed:
+        # `externalsecret_sync_calls_error` has exactly one series in the whole
+        # cluster, `ai/wiki-anthropic`, at 8 errors over 47 calls — and that
+        # ExternalSecret is Ready=True and has never been otherwise in the
+        # observation window. So on today's data the counter carries 8 events that
+        # correspond to no sync failure at all, and the gauge carries none.
+        #
+        # That observation window is now long enough to mean something, and it is
+        # the answer to the only open question this rule was held back for — how
+        # much the gauge flaps in ordinary operation. Measured 2026-09-26T08:02Z,
+        # over the 464 consecutive scrapes since #994 started collecting at
+        # 00:18Z:
+        #
+        #   min_over_time(up{job="monitoring/external-secrets"}[8h])     = 1
+        #   sum(increase(externalsecret_sync_calls_total[8h]))           = 959
+        #   count(increase(externalsecret_sync_calls_total[8h]) > 0)     = 52
+        #   count(max_over_time(
+        #     externalsecret_status_condition{
+        #       condition="Ready",status="False"}[8h]) == 1)             = EMPTY
+        #   sum(increase(externalsecret_sync_calls_error[8h]))           = 0
+        #
+        # Every one of the 52 ExternalSecrets reconciled at least once in that
+        # window — 959 syncs, so the 1h and 10m refreshInterval cycles both turned
+        # over repeatedly — and across all of them the Ready gauge never reached 1
+        # in a single scrape. Zero flap amplitude, against real reconcile traffic
+        # rather than against silence. The counter was flat over the same window
+        # too, which is why 8 stays 8 above. `for: 15m` is therefore not damping
+        # anything measurable; it is derived from the failure modes below, and the
+        # data says no shorter window would have been noisy either.
+        #
+        # The 2026-09-20 `ai/wiki-embed-basic-auth` incident is evidence FOR this
+        # rule, not against it, though it reads the other way at first. That key
+        # sat in ESO retry backoff through ~10 failures and NEEDED A MANUAL
+        # force-sync (the incident is written up in
+        # clusters/homelab/uptime-monitor.yaml, as the in-repo proof for REN-100).
+        # Because it never self-healed, Ready=False was continuous from the first
+        # failure until a human intervened — hours, not minutes. `for: 15m` catches
+        # that with a wide margin. It was a stuck condition, not a flap; the flap
+        # reading would only hold if the retries had eventually succeeded on their
+        # own, and they did not.
+        - alert: ExternalSecretSyncFailing
+          expr: >-
+            max by (exported_namespace, name) (
+              externalsecret_status_condition{condition="Ready",status="False"}
+            ) == 1
+          # 15m, and the reason it can be this relaxed is that a failed sync is
+          # not an outage: consumers keep using the last Secret ESO wrote, so what
+          # degrades is freshness, not availability. Nothing here needs catching
+          # inside a minute.
+          #
+          # It needs to be at least this long because the provider-error path
+          # returns the error to controller-runtime, which retries with
+          # exponential backoff rather than waiting out `refreshInterval`. A
+          # momentary OpenBao blip is therefore cleared within seconds and never
+          # reaches 15m, while a rolling restart of the 3-replica OpenBao
+          # StatefulSet is ridden out entirely — and that outage has its own
+          # alert (OpenBaoQuorumLost, critical at 2m), so nothing is hidden by
+          # waiting. 15m of continuous Ready=False means repeated failure.
+          for: 15m
+          labels:
+            severity: warning
+          annotations:
+            summary: "ExternalSecret {{ $labels.exported_namespace }}/{{ $labels.name }} is not syncing"
+            description: >-
+              ESO has been unable to sync this ExternalSecret for 15 minutes. The
+              Secret it manages still holds its last successfully synced value, so
+              consumers keep working — what breaks is that a rotation at the
+              provider no longer propagates, which surfaces much later as a
+              credential that stopped being accepted. Get the reason with
+              `kubectl -n {{ $labels.exported_namespace }} get externalsecret
+              {{ $labels.name }} -o jsonpath='{.status.conditions}'`; the common
+              causes are a key that was renamed or removed in OpenBao, a
+              ClusterSecretStore that cannot authenticate, and a target Secret
+              owned by something else.
+
+        # Dead-man's-switch, mirroring ExternalDNSMetricsMissing and
+        # OpenBaoMetricsMissing. The alert above can only fire on series that
+        # exist, so deleting the PodMonitor, reverting the `monitoring` ingress
+        # grant in infrastructure/network-policies/external-secrets-ns.yaml, or
+        # bumping the ESO chart onto a metrics port that no longer matches would
+        # all disarm it silently — which is the exact failure mode REN-103 exists
+        # to end. TargetDown does not cover this: it reports a target that exists
+        # and is failing, not the absence of a target.
+        - alert: ExternalSecretMetricsMissing
+          expr: absent(externalsecret_status_condition)
+          for: 15m
+          labels:
+            severity: warning
+          annotations:
+            summary: "No ExternalSecret metrics are reaching Prometheus"
+            description: >-
+              `externalsecret_status_condition` has no series. Either the ESO
+              controller is gone or its scrape stopped being collected, and
+              ExternalSecretSyncFailing is blind for as long as this is true.
+              Check the PodMonitor `monitoring/external-secrets` still selects the
+              controller pods, and that external-secrets-ns.yaml still admits
+              `monitoring` on TCP/8080.
+
+        # NOT covered here, deliberately, so nobody assumes it is:
+        #
+        #   * `condition="Deleted"`. In chart 2.11.0 the only writer of that
+        #     condition is the NotFound branch of Reconcile, which fabricates
+        #     `Deleted=False` purely to make UpdateExternalSecretCondition drop
+        #     the object's Ready series. It is a tombstone for an ExternalSecret
+        #     that no longer exists, not a health state, and `Deleted=True` is
+        #     never emitted at all — a rule on it would be dead code. The one such
+        #     series in the cluster today (`ai/wiki-index-basic-auth`) is leftover
+        #     process state for an ExternalSecret that moved namespace; it clears
+        #     when the controller next restarts.
+        #   * An ExternalSecret that exists in the cluster but was never
+        #     reconciled emits no series at all, so nothing above can see it. That
+        #     needs object-existence data — kube-state-metrics custom resource
+        #     state — rather than ESO's own metrics, and is not in this change.
+        #   * A sync that fails and then succeeds on its own inside 15m. This is a
+        #     choice, not an oversight: the retry left the Secret holding a freshly
+        #     synced value, so nothing degraded and there is nothing for a human to
+        #     do. Alerting on it would mean paging for a self-healed event. The case
+        #     that would justify revisiting is the same failure recurring for days
+        #     without any single occurrence lasting 15m, which would show up as a
+        #     climbing `externalsecret_sync_calls_error` on an ExternalSecret that
+        #     is Ready=True. Nothing in the cluster shows that shape today, and the
+        #     rule for it needs a measured recurrence rate rather than a guess.

kubernetes/apps/monitoring-rules/prometheusrule-flux.yaml

@@ -0,0 +1,274 @@
+# Per-object Flux readiness, and the groundwork for the Ready-but-frozen
+# ImagePolicy signal (REN-102, closing REN-97).
+#
+# EVERYTHING HERE READS `gotk_resource_info`, WHICH IS A kube-state-metrics
+# SERIES, NOT A FLUX CONTROLLER SERIES. apps/monitoring-rules/flux-podmonitor.yaml
+# scrapes the controllers and yields reconcile durations and controller-runtime
+# counters, but readiness is not on that endpoint at the pinned controller
+# versions: `gotk_reconcile_condition` and `gotk_suspend_status` are declared by
+# fluxcd/pkg/runtime and populated by nothing. The series these rules read are
+# created by the `customResourceState` block in
+# apps/monitoring/helmrelease-prometheus.yaml. If that block is removed, every
+# rule below goes blind — which is what FluxImagePolicyMetricsMissing exists to
+# say out loud.
+#
+# THE STALENESS HALF DELIBERATELY SHIPS WITHOUT AN ALERT. See the second group.
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: flux
+  namespace: monitoring
+  labels:
+    # ruleSelector matches on this label; without it the rule is created but
+    # never loaded into Prometheus.
+    release: kube-prometheus-stack
+spec:
+  groups:
+    - name: flux.rules
+      rules:
+        # Why `max by (...)` and not the bare selector: it collapses the KSM
+        # target labels (`pod`, `instance`, `endpoint`, ...). Without it, a
+        # kube-state-metrics rollout changes `pod`, which resolves the alert on
+        # the old series and re-fires it on the new one — a fresh mail for an
+        # unchanged failure. KSM rolled 11 times in the last 30 days, so this is
+        # routine, not hypothetical.
+        #
+        # Why `ready=~"False|Unknown"` and not `ready="False"`: Flux sets Ready
+        # False on failure but Unknown while a reconcile is in flight, and an
+        # object wedged in Unknown is stuck just as hard as one that is False —
+        # this repo has already had a Progressing CephCluster deadlock the whole
+        # apps layer (see rook-servicemonitors.yaml). Matching the two values
+        # explicitly, rather than `ready!="True"`, keeps an object whose `ready`
+        # label is absent entirely (freshly created, condition not yet written)
+        # out of the alert.
+        #
+        # Why `suspended!="true"` and not an `unless` join: `suspended` is a
+        # label on the same Info series, so one matcher does what the withdrawn
+        # draft needed a second vector for. The `!=` form also covers kinds where
+        # the label is absent. There is no deleted-object guard because KSM stops
+        # emitting the series when the object goes away — under `gotk_*` that
+        # guard had to be written, here it is structural.
+        #
+        # for: 2h — 31 of the 34 ImageRepositories scan at `interval: 1h`
+        # (2 at 30m: shelly-manager, vsftpd; 1 at 10m: wiki), so 91% of the
+        # population cannot recover from a transient registry failure faster
+        # than one hour whatever this value says. Anything under 1h mails about
+        # a single self-healing missed scan. 2h means two consecutive failed
+        # scans. Detection latency added by going through KSM is one 30s scrape.
+        - alert: FluxImageAutomationNotReady
+          expr: >-
+            max by (customresource_kind, name, exported_namespace) (
+              gotk_resource_info{
+                customresource_kind=~"ImageRepository|ImagePolicy|ImageUpdateAutomation",
+                ready=~"False|Unknown",
+                suspended!="true"
+              }
+            )
+          for: 2h
+          labels:
+            severity: warning
+          annotations:
+            summary: "Flux {{ $labels.customresource_kind }}/{{ $labels.name }} has not been Ready for 2h"
+            description: >-
+              An image-automation object has been not-Ready for two reconcile
+              intervals. Nothing is broken in the running cluster, but the image
+              it governs has stopped moving, security rebuilds included.
+              `flux get image all -n {{ $labels.exported_namespace }}` gives the
+              reason; the usual causes are registry credentials, a rate limit, or
+              a tag format the policy's `pattern`/`extract` no longer matches.
+              flux-notifications/alerts.yaml already mailed once at the moment of
+              transition — this is the "still broken two hours later" signal that
+              edge-triggered mail cannot give.
+
+        # The exact negation of the matcher above, so a kind added to
+        # helmrelease-prometheus.yaml's customResourceState config is covered by
+        # one of these two rules without anyone remembering to widen a list.
+        #
+        # for: 1h, derived from the intervals actually in this repo: all 33
+        # Kustomizations are at 10m (six retries inside the hour), all 25
+        # HelmReleases at 30m (two), and the flux-system GitRepository at 1m0s
+        # (sixty). The dominant non-failure source of ready="False" is a
+        # `dependsOn` wait, whose `reason` is not a label and so cannot be
+        # filtered out; an hour clears a whole-cluster reconcile chain.
+        #
+        # KNOWN TRADE, stated rather than hidden: this matcher also picks up the
+        # 17 HelmRepository and 4 OCIRepository objects, and every one of those is
+        # at `interval: 24h`. A failed index or manifest fetch there cannot
+        # self-heal inside an hour, so a transient failure yields up to two mails
+        # (repeat_interval is 12h) before the next scheduled reconcile. That is
+        # accepted as correct — a chart source stale for an hour is worth knowing
+        # — and it is real information rather than a blip. If it turns out to be
+        # the noisy one, the fix is a third rule for the 24h-interval kinds, not
+        # a global bump of this value, which would make Kustomization alerts
+        # useless.
+        - alert: FluxResourceNotReady
+          expr: >-
+            max by (customresource_kind, name, exported_namespace) (
+              gotk_resource_info{
+                customresource_kind!~"ImageRepository|ImagePolicy|ImageUpdateAutomation",
+                ready=~"False|Unknown",
+                suspended!="true"
+              }
+            )
+          for: 1h
+          labels:
+            severity: critical
+          annotations:
+            summary: "Flux {{ $labels.customresource_kind }}/{{ $labels.name }} has not been Ready for 1h"
+            description: >-
+              A Flux object outside image automation has been not-Ready for an
+              hour, so the desired state in git is not reaching the cluster.
+              `flux get all -A --status-selector ready=false` lists them with a
+              reason. For a Kustomization or HelmRelease this means a rollout is
+              stuck; for a HelmRepository or OCIRepository it means the chart or
+              artifact index is stale and dependent releases will not upgrade.
+
+        # THE DEAD MAN'S SWITCH, and the reason it is safe here and was not in
+        # #992. This asserts that the 34 ImagePolicy series exist — and they
+        # exist only because this change created them. Merged against `gotk_*`
+        # alone, as the withdrawn draft had it, this rule would have been
+        # permanently true and mailed admin@renner.uno every 12h forever.
+        #
+        # Its value is catching the removal of the customResourceState block, or
+        # a malformed edit to it, which would otherwise take all three rules
+        # above silently blind. `KubeStateMetricsDown` and `TargetDown` overlap it
+        # when the whole pod is down; neither notices a config that parses but
+        # emits nothing.
+        #
+        # for: 30m rides out the kube-state-metrics pod roll that applying this
+        # change itself causes, and any later roll.
+        - alert: FluxImagePolicyMetricsMissing
+          expr: absent(gotk_resource_info{customresource_kind="ImagePolicy"})
+          for: 30m
+          labels:
+            severity: critical
+          annotations:
+            summary: "Flux per-object readiness metrics have stopped being exported"
+            description: >-
+              kube-state-metrics is exporting no ImagePolicy series, so
+              FluxImageAutomationNotReady and FluxResourceNotReady cannot fire and
+              Flux readiness is unmonitored. Check the `customResourceState` block
+              in apps/monitoring/helmrelease-prometheus.yaml, then the
+              kube-state-metrics pod log — a custom-resource-state config that
+              fails to parse is logged there and yields no series.
+
+    # The Ready-but-frozen ImagePolicy signal: a policy whose `filterTags` or
+    # semver range excludes everything new keeps resolving the tag it resolved
+    # months ago, holds Ready=True the whole time, and no readiness rule can see
+    # it. `tika.yaml`'s own comment names the case.
+    #
+    # THIS GROUP DELIBERATELY CONTAINS NO ALERT, AND THAT IS A MEASURED DECISION,
+    # NOT AN OMISSION. Every candidate threshold was tested against the real
+    # release history and the real registries on 2026-09-26:
+    #
+    #   * A flat "frozen for N days" duration does not exist. Over the 191
+    #     automated bump commits in the 147 days to 2026-09-25, gatus went 127
+    #     days between bumps and paperless-ngx 126, both provably healthy — they
+    #     bumped. Six policies have never bumped at all. Any N below 127 fires on
+    #     the healthy pair; any N above 147 is longer than the whole observation
+    #     window.
+    #
+    #   * The "tag count grew >= 10%" comparator from this card's first design is
+    #     wrong, and the registries say so. Of the 18 policies frozen 14 days or
+    #     more, a 10% threshold fires on exactly three — gotenberg (+87 tags/14d
+    #     on 526), valkey (+68 on 301), eclipse-mosquitto (+13 on 90) — and all
+    #     three churn variant tags their filters reject on purpose. It does NOT
+    #     fire on tika (+4 on 136 = 3.0%), which is the canonical case this card
+    #     was opened for.
+    #
+    #   * Relaxing to "grew by >= 1 tag" does catch tika, but then also fires on
+    #     those same three plus velero-plugin-for-aws, whose `1.13.x` pin is
+    #     deliberate. Alertmanager routes every severity to the same email
+    #     receiver with repeat_interval: 12h, so that is four standing mails
+    #     every twelve hours, indefinitely, for four non-faults.
+    #
+    #   * `.status.lastScanResult.latestTags` looks like the way out and is not:
+    #     the CRD defines it as "the first 10 tags when sorting all the tags in
+    #     descending ALPHABETICAL order", which is not newest-first for any tag
+    #     scheme in this repo.
+    #
+    # So the recorded series below materialise the candidate set and, more to the
+    # point, START THE HISTORY that a threshold has to be derived from — a real
+    # time series of tagCount per repository, which has never existed here. The
+    # alert is deferred to a follow-up card that reads 14+ days of it. Recording
+    # rules cannot mail anyone, so this costs nothing while it warms up.
+    #
+    # interval: 15m is LOAD-BEARING and is declared here rather than inherited.
+    # Global evaluationInterval is 30s and no other rule group in this directory
+    # sets one (45 of the 46 loaded groups run at 30s). The `>= 1250` warm-up
+    # guard below counts samples, so at 30s it would be satisfied in 10.4 hours
+    # instead of 13 days and the frozen series would read true for policies that
+    # had not been observed frozen at all.
+    - name: flux-image-policy-staleness.rules
+      interval: 15m
+      rules:
+        # Stage one exists only to strip the KSM target labels. Counting distinct
+        # tags means counting distinct series, and `gotk_resource_info` carries
+        # `pod` and `instance` — there were 11 distinct kube-state-metrics pod
+        # identities in the last 30 days, so counting the raw series would report
+        # "several tags seen" after every roll, including the roll this change
+        # itself causes, and would keep reporting it for the whole window.
+        - record: flux:image_policy_tag:info
+          expr: >-
+            max by (name, exported_namespace, source_name, tag) (
+              gotk_resource_info{customresource_kind="ImagePolicy"}
+            )
+
+        # Distinct resolved tags seen in a trailing day. A tag change retires one
+        # series and starts another, so `count by (...)` over `count_over_time`
+        # is a count of tag values, and `== 1` means the tag did not move.
+        - record: flux:image_policy_tags_seen:1d
+          expr: >-
+            count by (name, exported_namespace, source_name) (
+              count_over_time(flux:image_policy_tag:info[1d])
+            )
+
+        # The repository side, keyed by `source_name` so it joins on the actual
+        # `spec.imageRepositoryRef.name` reference rather than on the convention
+        # that a policy is named after its repository. All 34 pairs match by name
+        # today, so this is drift insurance, not a bug fix.
+        - record: flux:image_repository_tag_count:1d
+          expr: >-
+            max by (source_name, exported_namespace) (
+              label_replace(
+                max_over_time(gotk_image_repository_tag_count[1d]),
+                "source_name", "$1", "name", "(.+)"
+              )
+            )
+
+        # The candidate set: resolved tag unmoved for 14 days while the
+        # repository's tag count is genuinely higher than it was.
+        #
+        # 14d, not the 30d this card was first designed around, because 30d does
+        # not fit. `retentionSize: 30GB` binds long before `retention: 90d`:
+        # (time() - prometheus_tsdb_lowest_timestamp/1000)/86400 measured 26.63
+        # days on 2026-09-26, at 84.5% of the size cap and growing ~1.09 GB/day
+        # (REN-107). A 30d window can never be satisfied. 14d uses 53% of that
+        # floor.
+        #
+        # >= 1250 is the warm-up guard and its arithmetic is: at interval 15m a
+        # full 14d window holds 14*24*4 = 1344 samples, so 1250 is 93% of a full
+        # window = 13.02 days of history, leaving ~23h of slack for Prometheus
+        # restarts and KSM rolls. Without it the series would read true on day
+        # one, when "unmoved for 14 days" is simply not yet knowable. 13.02 days
+        # against a 26.63-day floor is 2.0x margin; re-check it with the
+        # lowest_timestamp query above if retention is ever cut.
+        #
+        # `last_over_time - min_over_time >= 1` and not a max/min ratio: a ratio
+        # is direction-agnostic and is satisfied by a tag count that DROPPED
+        # (registry pruning, an edited exclusionList, a partial scan) exactly as
+        # by one that grew. Comparing the current value against the window
+        # minimum only counts net growth.
+        - record: flux:image_policy_tag_frozen:14d
+          expr: >-
+            (
+              max_over_time(flux:image_policy_tags_seen:1d[14d]) == 1
+              and
+              count_over_time(flux:image_policy_tags_seen:1d[14d]) >= 1250
+            )
+            and on (source_name, exported_namespace) (
+              (
+                last_over_time(flux:image_repository_tag_count:1d[1h])
+                - min_over_time(flux:image_repository_tag_count:1d[14d])
+              ) >= 1
+            )

kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml

@@ -0,0 +1,119 @@
+apiVersion: monitoring.coreos.com/v1
+kind: PrometheusRule
+metadata:
+  name: home-assistant
+  namespace: monitoring
+  labels:
+    # kube-prometheus-stack's Prometheus resource selects on this label;
+    # without it the rule is created but never loaded into Prometheus.
+    release: kube-prometheus-stack
+spec:
+  groups:
+    - name: home-assistant.rules
+      rules:
+        # The four services the house stops working without (REN-244). Before
+        # this group the only signal for any of them was the cluster-wide
+        # KubePodNotReady: 15m, severity warning, and not named per service.
+        #
+        # SHAPE. One alert per service, with `or` joining the ways it can be
+        # down, so one outage is one alert and not one per symptom. All four are
+        # `for: 5m`, severity critical. Alertmanager has one receiver today
+        # (email to the owner, apps/monitoring/helmrelease-prometheus.yaml), so
+        # "critical" does not change where these go today. It does keep them
+        # out of reach of the inhibit rules, which only silence warning/info.
+        #
+        # 5m IS SHORTER THAN A SLOW HA UPGRADE. Home Assistant can take up to
+        # 10 min to come back after an automatic image bump. REN-244 asks for
+        # detection in 5 min on purpose; a HomeAssistantDown that resolves on
+        # its own a few minutes later, right after an upgrade, is that case.
+        #
+        # Tests: tests/home-assistant_test.yaml.
+
+        - alert: HomeAssistantDown
+          # `absent` covers the target vanishing entirely (pod gone, Service
+          # has no endpoints during a Recreate rollout); `== 0` covers a target
+          # that exists but does not answer /api/prometheus.
+          #
+          # `max by (job, namespace)` gives both branches the same label set.
+          # Without it the alert changes identity when a pod that was gone
+          # comes back failing, and the new identity restarts the `for` timer.
+          expr: |
+            max by (job, namespace) (up{job="home-assistant", namespace="home-assistant"}) == 0
+            or absent(up{job="home-assistant", namespace="home-assistant"})
+          for: 5m
+          labels:
+            severity: critical
+          annotations:
+            summary: "Home Assistant has been down for 5 minutes"
+            description: "The home-assistant scrape target in namespace home-assistant is down or missing. A restart after an automatic image upgrade can take up to 10 minutes, so if an upgrade just rolled out, give it until then before acting. Rollback: kubernetes/RUNBOOK-home-assistant-image-rollback.md."
+
+        - alert: HomeAssistantRecorderDBUnhealthy
+          # ha-pg, the recorder's CNPG cluster. Fires when no instance is the
+          # primary, or when fewer than all instances have Postgres up.
+          #
+          # cnpg_collector_up, not `up`: the exporter lives in the instance
+          # manager, which keeps serving /metrics while Postgres itself is down.
+          # `up` would read 1 through exactly the outage this rule is for.
+          #
+          # `role` comes from podTargetLabels on the cnpg-ha-pg PodMonitor
+          # (cnpg-podmonitors.yaml). It flips on failover, so the instance count
+          # is taken across roles. Both branches come out as {} (`absent` of a
+          # comparison carries no labels), so it is one alert whichever of them
+          # is true, and a failover does not start a second one.
+          #
+          # THE 2 IS `spec.instances` in
+          # kubernetes/apps/home-assistant/ha-pg-cluster.yaml. Change both
+          # together, or this pages forever (more instances is fine; fewer is
+          # not).
+          expr: |
+            count(cnpg_collector_up{job="home-assistant/cnpg-ha-pg"} == 1) < 2
+            or absent(cnpg_collector_up{job="home-assistant/cnpg-ha-pg", role="primary"} == 1)
+          for: 5m
+          labels:
+            severity: critical
+            namespace: home-assistant
+          annotations:
+            summary: "Recorder database ha-pg has no primary or is missing an instance"
+            description: "In CNPG cluster home-assistant/ha-pg either no instance reports role=primary with Postgres up, or fewer than 2 instances have Postgres up. Without a primary the Home Assistant recorder cannot write history. Check `kubectl -n home-assistant get cluster ha-pg` and the pods' logs."
+
+        - alert: MosquittoDown
+          # Pod readiness only: there is no broker exporter, and no mosquitto_*
+          # or mqtt_* series in Prometheus.
+          #
+          # The pod regex is the Deployment's name shape
+          # (mosquitto-<replicaset hash>-<suffix>) and not `mosquitto-.*` on
+          # purpose: mosquitto-bootstrap.yaml runs six Jobs named
+          # mosquitto-*-credentials-bootstrap, and a bootstrap pod that is
+          # briefly Ready must not hide a broker that is down.
+          expr: |
+            absent(kube_pod_status_ready{namespace="home-assistant", pod=~"mosquitto-[a-z0-9]{6,10}-[a-z0-9]{5}", condition="true"} == 1)
+          for: 5m
+          labels:
+            severity: critical
+            # `absent` of a comparison carries no labels; set it for routing.
+            namespace: home-assistant
+          annotations:
+            summary: "Mosquitto (MQTT broker) has not been ready for 5 minutes"
+            description: "No mosquitto pod in namespace home-assistant is Ready. Zigbee2MQTT, Shelly and every other MQTT device lose their link to Home Assistant while this lasts, so Zigbee2MQTTDownOrCoordinatorOffline may fire alongside it for the same cause."
+
+        - alert: Zigbee2MQTTDownOrCoordinatorOffline
+          # Two ways Zigbee stops working: the pod is not Ready, or it runs but
+          # has lost the coordinator / MQTT bridge. The second is read from
+          # Home Assistant's own bridge sensor, so it disappears when HA is down;
+          # HomeAssistantDown covers that case and this rule stays silent
+          # instead of paging twice.
+          #
+          # Both branches come out as {} (`absent` of a comparison carries no
+          # labels; the sensor branch is reduced with `max`), for the same
+          # reason as in HomeAssistantDown: pod-not-ready turning into
+          # bridge-offline is one outage and keeps one `for` timer.
+          expr: |
+            absent(kube_pod_status_ready{namespace="home-assistant", pod=~"zigbee2mqtt-[a-z0-9]{6,10}-[a-z0-9]{5}", condition="true"} == 1)
+            or max(homeassistant_binary_sensor_state{entity="binary_sensor.zigbee2mqtt_bridge_connection_state"}) == 0
+          for: 5m
+          labels:
+            severity: critical
+            namespace: home-assistant
+          annotations:
+            summary: "Zigbee2MQTT is down or its coordinator/bridge is offline"
+            description: "Either the zigbee2mqtt pod in namespace home-assistant is not Ready, or Home Assistant's binary_sensor.zigbee2mqtt_bridge_connection_state reads off. If MosquittoDown is firing too, fix the broker first; the bridge sensor goes off whenever MQTT is gone."

kubernetes/apps/monitoring-rules/prometheusrule-openbao.yaml

@@ -57,10 +57,14 @@ spec:
           # the ClusterSecretStore goes InvalidProviderConfig and no
           # ExternalSecret can refresh.
           #
-          # No window: below quorum the secret store is down and every
-          # ExternalSecret stalls with it, so waiting only delays the page.
+          # Short window: readiness is only a proxy for quorum. A one-second
+          # node NotReady flap makes the node lifecycle controller mark that
+          # node's pods not-ready for ~30s while raft keeps its leader and
+          # serves (issue #920, during a rolling reboot with one replica
+          # already evicted). A really sealed or crashed majority persists,
+          # and existing Secrets keep their last values, so 2m costs little.
           expr: max by (namespace, statefulset) (kube_statefulset_status_replicas_ready{namespace="openbao", statefulset="openbao"}) < 2
-          for: 0m
+          for: 2m
           labels:
             severity: critical
           annotations:

kubernetes/apps/monitoring-rules/tests/README.md

@@ -0,0 +1,127 @@
+# promtool unit suites for the PrometheusRules in this directory
+
+`promtool test rules` suites, run on every pull request that touches
+`kubernetes/**` by the `promtool` step in
+[`.woodpecker/ci-manifests-test.yml`](../../../../.woodpecker/ci-manifests-test.yml)
+via [`scripts/ci/promtool-rules.sh`](../../../../scripts/ci/promtool-rules.sh).
+
+## Why this exists
+
+Before this, CI proved that every `PrometheusRule` in the tree assembled into
+valid YAML and nothing else. It did not check that an expression parses, that a
+label join matches anything, or that a `for` window behaves the way its comment
+claims.
+
+That was not theoretical. On [REN-129](/REN/issues/REN-129) the surge rule in
+`prometheusrule-cnpg-backup.yaml` was written `for: 2m` and reviewed at that
+value. Against Iris's seven measured nights it fires on six of them — two owner
+emails a night, 04:16 and 04:22 local, for a WAL backlog that clears itself in
+under ten minutes. It was caught by a human writing the arithmetic out by hand,
+twice, in two separate runs, and throwing the work away both times because
+there was nowhere in the repo to put it. [REN-152](/REN/issues/REN-152) then
+retired the depth rule in favour of one persistence rule (`> 5` for `20m`).
+
+## Layout
+
+| File | What it is |
+| --- | --- |
+| `cnpg-backup_test.yaml` | The suite. **Generated** — see below. |
+| `generate-cnpg-backup-test.py` | Regenerates it from the REN-129 measurement table. |
+| `external-secrets_test.yaml` | Hand-written suite for `prometheusrule-external-secrets.yaml` (REN-103). |
+| `home-assistant_test.yaml` | Hand-written suite for `prometheusrule-home-assistant.yaml` (REN-245). |
+| `ceph_test.yaml` | Hand-written suite for the `CephPoolLostRedundancyBlind` canary in `prometheusrule-ceph.yaml` (REN-93). |
+| `mutations.txt` | Negative controls. Each must make the suite fail. |
+
+## Adding a suite
+
+Name it `<something>_test.yaml` in this directory. Discovery is by that suffix,
+and `promtool-rules.sh` fails if it finds none, so a suite in the wrong place
+is a red pipeline rather than a silent gap.
+
+**`rule_files` entries are relative to this directory, not the repo root.**
+promtool joins `dirname(suite)` onto each entry, so a manifest in the parent
+directory is `../name.yaml`. Getting this wrong is not a test failure: promtool
+prints `WARNING: no file match pattern …`, loads **zero rules**, and every
+`exp_alerts: []` assertion then passes against an empty rule set. The first
+draft of the suite here shipped exactly that way — ninety silence assertions,
+green, testing nothing. `promtool-rules.sh` resolves every entry itself before
+running promtool and fails on a miss, because promtool has no flag that makes
+it fatal.
+
+`promtool` cannot read a `PrometheusRule` directly: it wants a bare `groups:`
+document, and a `PrometheusRule` carries that under `spec:`. The script strips
+`spec:` into a temporary mirrored tree with the same relative layout and runs
+there, so `../name.yaml` resolves in the mirror exactly as it reads here.
+Nothing is written back into the checkout.
+
+## Regenerating `cnpg-backup_test.yaml`
+
+```sh
+python3 kubernetes/apps/monitoring-rules/tests/generate-cnpg-backup-test.py \
+  > kubernetes/apps/monitoring-rules/tests/cnpg-backup_test.yaml
+```
+
+and commit the result. `--report` prints the reconstruction self-check table
+instead of the suite.
+
+The generator exists so the numbers in the suite are auditable: the seven
+measured nights are rebuilt from the `(peak, peak minute, fill, drain)` tuples
+Iris published on REN-129, not hand-typed, so a reader can check them against
+the measurement rather than taking them on trust. It is not a hermetic
+artifact, and **CI does not regenerate and diff it** — the promtool image has
+no Python, and byte-equality with a generator is not what stops a suite from
+being weakened. The mutation phase is.
+
+Two things the generator refuses to emit, both because they happened:
+
+- a window that does not cover the bursts. An earlier revision had `T0` at
+  `8 * 60 + 8` (08:08Z) instead of `2 * 60 + 8`, seven hours after every burst,
+  so all seven nights rebuilt as forty zeroes and their silence assertions
+  passed without touching a threshold. `assert_window_covers_bursts()` now
+  fails instead.
+- a night that clears the threshold *sooner* than the measured one, which would
+  be a weaker silence test than the night it claims to stand for. Longer is
+  fine and expected — a linear decay at the published mean drain rate holds the
+  tail above 5 a minute or two longer than the real trace on four of the seven
+  nights, which makes the suite harder to pass, not easier.
+
+## Negative controls
+
+A suite that passes proves the rules match the assertions. It does not prove
+the assertions reach the rules — both failure modes above were green. So
+`promtool-rules.sh` runs a second phase: for each entry in `mutations.txt` it
+breaks one value the rules depend on and requires the suites to **fail**. A
+mutation nothing notices is reported as a hole in the suite.
+
+This is also the answer to "why is that case in there". `backlog-for-12m` is
+silent on all seven reconstructed nights and fires only on the raw 2026-09-11
+worst-night case, so deleting that case makes `for: 20m` reducible to `12m`
+with the suite still green. The control is what makes the case load-bearing.
+
+Mutations are `name|manifest-path|sed-program`, applied to the stripped copy in
+the mirror (two spaces shallower than the manifest on disk — anchor on
+`^[[:space:]]*`). A program that matches nothing is a failure, not a skip:
+otherwise renaming a rule quietly turns its control into a no-op.
+
+## Reading `for` margins
+
+Margin is **pending** time, which is one minute less than the run length. An
+alert goes pending at the first evaluation where the expression is true and
+fires when `now - activeAt >= for`, so N consecutive one-minute samples above a
+threshold accumulate N-1 minutes. A trace with 5 samples above 5 does not fire
+`for: 5m`; it fires `for: 4m`. The `>5 pend` column in `--report` is the number
+to compare a `for` against — not the sample count next to it.
+
+## Running it locally
+
+```sh
+# promtool must be 3.x: the cluster runs kube-prometheus-stack >= 72.0.0, and
+# alert `for` evaluation is version-sensitive, so a 2.x promtool would be
+# checking a different evaluator than production. The script enforces this.
+sh scripts/ci/promtool-rules.sh
+```
+
+The script's own guards — the discovery floors, the `rule_files` resolution, the
+mutation phase — are tested offline and without promtool by
+`scripts/ci/test-promtool-rules.sh`, which runs on every pull request via
+`.woodpecker/ci-scripts-test.yml`.

kubernetes/apps/monitoring-rules/tests/ceph_test.yaml

@@ -0,0 +1,176 @@
+# promtool test rules for the CephPoolLostRedundancyBlind canary in
+# prometheusrule-ceph.yaml (REN-93).
+#
+# Run by the `promtool` step of .woodpecker/ci-manifests-test.yml through
+# scripts/ci/promtool-rules.sh, which strips `spec:` from the manifest into a
+# mirrored tree — hence `../` below, resolved from this file's directory. See
+# README.md here. The negative controls for this suite are the `ceph-*` entries
+# in mutations.txt.
+#
+# WHAT IT IS FOR. `promtool check rules` proves the canary parses. It cannot
+# prove the canary fires when the join breaks, or stays quiet when the join is
+# fine — every scenario below is well-formed PromQL in both the passing and the
+# failing case; only the synthetic data changes, and only evaluation sees it.
+#
+# `exp_annotations` must match the RENDERED annotations exactly, so the strings
+# below are the rule's with `{{ $labels.pool_id }}` substituted. Tedious to keep
+# in sync, and that is the point: an annotation edited into nonsense fails here
+# instead of in an email.
+rule_files:
+  - ../prometheusrule-ceph.yaml
+evaluation_interval: 1m
+tests:
+  # A. Healthy replicated pool, metadata intact. Neither alert may fire.
+  #    used/stored = 3.0, so well clear of the 1.5 threshold, and the join
+  #    resolves — so the canary has nothing to report either. This is the
+  #    scenario that fails if someone inverts a comparison.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '6000000+0x200'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_metadata{pool_id="1",type="replicated",name="rbd"}'
+        values: '1+0x200'
+    alert_rule_test:
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts: []
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancy
+        exp_alerts: []
+
+  # B. Real redundancy loss, metadata intact: used/stored = 1.0.
+  #    The critical alert must fire AND the canary must stay quiet. The second
+  #    half is the load-bearing one — a canary that also fires on a real
+  #    redundancy loss would mask the alert it exists to protect, and
+  #    `absent()` over an expression that IS producing series is the only thing
+  #    stopping that. Also pins the `for: 30m`: 120m is well past it.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_metadata{pool_id="1",type="replicated",name="rbd"}'
+        values: '1+0x200'
+    alert_rule_test:
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancy
+        exp_alerts:
+          - exp_labels:
+              severity: critical
+              pool_id: "1"
+            exp_annotations:
+              summary: "Ceph replicated pool 1 is running unreplicated"
+              description: "Replicated pool 1 has a replication factor below 1.5, meaning it is down to a single copy and one OSD loss destroys data. No replicated pool is meant to be unreplicated — it also blocks OSD maintenance for as long as it lasts, since `ceph osd ok-to-stop` refuses while its PGs would go offline. Erasure-coded pools are out of scope for this rule and are covered by POOL_NO_REDUNDANCY via CephHealthWarn (REN-88). Check with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail`."
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts: []
+
+  # C, D and E are the same failure seen three ways: the pool is genuinely at
+  # replication 1.0, and CephPoolLostRedundancy produces NOTHING. Each asserts
+  # `exp_alerts: []` for the critical alert — that empty list is the silent
+  # failure being surfaced, not an absence of interest — and asserts the canary
+  # fires instead. Remove the canary from the rule file and all three fail.
+  #
+  # C. `type` renamed to `pool_type` by a Rook or Ceph upgrade.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_metadata{pool_id="1",pool_type="replicated",name="rbd"}'
+        values: '1+0x200'
+    alert_rule_test:
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancy
+        exp_alerts: []
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "CephPoolLostRedundancy cannot fire — its metrics no longer join"
+              description: "The CephPoolLostRedundancy alert has been unable to produce any series for an hour, so a `critical` storage alert is silently dark. It joins ceph_pool_bytes_used / ceph_pool_stored against ceph_pool_metadata{type=\"replicated\"} on `pool_id`; one of those metrics, the `pool_id` label, the `type` label, or the `replicated` value has stopped matching — most likely a Rook or Ceph upgrade, or a relabel rule. Check what the exporter emits now with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail` and by querying `ceph_pool_metadata` directly, then fix the join in kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml — do not delete this canary (REN-93)."
+
+  # D. Label VALUE changed rather than the label name: `replicated` ->
+  #    `replica`. Every label the rule names still exists, which is why a check
+  #    on label presence alone would not catch this.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_metadata{pool_id="1",type="replica",name="rbd"}'
+        values: '1+0x200'
+    alert_rule_test:
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancy
+        exp_alerts: []
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "CephPoolLostRedundancy cannot fire — its metrics no longer join"
+              description: "The CephPoolLostRedundancy alert has been unable to produce any series for an hour, so a `critical` storage alert is silently dark. It joins ceph_pool_bytes_used / ceph_pool_stored against ceph_pool_metadata{type=\"replicated\"} on `pool_id`; one of those metrics, the `pool_id` label, the `type` label, or the `replicated` value has stopped matching — most likely a Rook or Ceph upgrade, or a relabel rule. Check what the exporter emits now with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail` and by querying `ceph_pool_metadata` directly, then fix the join in kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml — do not delete this canary (REN-93)."
+
+  # E. `pool_id` renamed to `pool` on the METADATA side only. Both metrics
+  #    exist, both label values are unchanged, and `type="replicated"` still
+  #    matches — only the `on(pool_id)` join key is gone from one side, so the
+  #    join matches nothing. The subtlest of the three, and the one a human
+  #    reading the expression is least likely to spot.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x200'
+      - series: 'ceph_pool_metadata{pool="1",type="replicated",name="rbd"}'
+        values: '1+0x200'
+    alert_rule_test:
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancy
+        exp_alerts: []
+      - eval_time: 120m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "CephPoolLostRedundancy cannot fire — its metrics no longer join"
+              description: "The CephPoolLostRedundancy alert has been unable to produce any series for an hour, so a `critical` storage alert is silently dark. It joins ceph_pool_bytes_used / ceph_pool_stored against ceph_pool_metadata{type=\"replicated\"} on `pool_id`; one of those metrics, the `pool_id` label, the `type` label, or the `replicated` value has stopped matching — most likely a Rook or Ceph upgrade, or a relabel rule. Check what the exporter emits now with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail` and by querying `ceph_pool_metadata` directly, then fix the join in kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml — do not delete this canary (REN-93)."
+
+  # F. mgr FAILOVER, which is the noise case the canary's `for: 1h` exists for.
+  #    All ceph_* metrics stop at t=30m and never return; with the 5m lookback
+  #    they read as present until t=35m, so the canary goes pending at ~t=36m.
+  #    It must NOT be firing at 75m (a failover-length gap — the exporter moved
+  #    through 17 mgr instances in 30d) and MUST be firing by 95m. Shorten the
+  #    `for:` and the 75m assertion fails; lengthen it and the 95m one does.
+  #    Both eval_times were checked against promtool 3.15.0 rather than derived,
+  #    because the exact ActiveAt depends on lookback and evaluation alignment.
+  - interval: 1m
+    input_series:
+      - series: 'ceph_pool_bytes_used{pool_id="1"}'
+        values: '6000000+0x30'
+      - series: 'ceph_pool_stored{pool_id="1"}'
+        values: '2000000+0x30'
+      - series: 'ceph_pool_metadata{pool_id="1",type="replicated",name="rbd"}'
+        values: '1+0x30'
+    alert_rule_test:
+      - eval_time: 75m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts: []
+      - eval_time: 95m
+        alertname: CephPoolLostRedundancyBlind
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "CephPoolLostRedundancy cannot fire — its metrics no longer join"
+              description: "The CephPoolLostRedundancy alert has been unable to produce any series for an hour, so a `critical` storage alert is silently dark. It joins ceph_pool_bytes_used / ceph_pool_stored against ceph_pool_metadata{type=\"replicated\"} on `pool_id`; one of those metrics, the `pool_id` label, the `type` label, or the `replicated` value has stopped matching — most likely a Rook or Ceph upgrade, or a relabel rule. Check what the exporter emits now with `kubectl -n rook-ceph exec deploy/rook-ceph-tools -- ceph osd pool ls detail` and by querying `ceph_pool_metadata` directly, then fix the join in kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml — do not delete this canary (REN-93)."

kubernetes/apps/monitoring-rules/tests/cnpg-backup_test.yaml

@@ -0,0 +1,431 @@
+# GENERATED by generate-cnpg-backup-test.py — see README.md in this
+# directory before editing by hand.
+#
+# Unit tests for kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml.
+# Every assertion below is either one of Iris's seven measured nights
+# (REN-129) rebuilt from its published (peak, fill, drain) numbers, the
+# worst night in Prometheus' retention as raw samples (REN-152), an
+# adversarial combination of the extremes of those numbers, or a failure
+# this cluster has not had but the rule exists to catch.
+#
+# `rule_files` entries are resolved RELATIVE TO THIS FILE'S DIRECTORY, not
+# to the working directory promtool was invoked from — promtool joins
+# dirname(testfile) onto each entry. So this is `../`, and it resolves in
+# the mirrored tree scripts/ci/promtool-rules.sh builds, where the
+# `spec:`-stripped manifests sit in the same layout as the real ones.
+#
+# A path that does NOT resolve is only a WARNING to promtool: it loads zero
+# rules, every exp_alerts: [] assertion passes against nothing, and the
+# suite exits 0. promtool-rules.sh resolves every entry itself before
+# running and fails on a miss; that check is the only thing standing
+# between a typo here and a green pipeline that tests nothing.
+rule_files:
+  - ../prometheusrule-cnpg-backup.yaml
+evaluation_interval: 1m
+
+tests:
+  # ---------------------------------------------------------------------
+  # 1-7. THE SEVEN MEASURED NIGHTS. The rule must stay silent through
+  # every one of them. The retired depth rule (`> 64`) shipped for review
+  # at `for: 2m` and fired on 6 of these 7 nights; REN-152 removed it.
+  # ---------------------------------------------------------------------
+  # 2026-09-20: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 106 at 02:15Z, fill 35.3/min, drain 13.2/min -> 10 samples above 5 (9m pending vs for: 20m).
+  - name: nightly burst 2026-09-20 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 0 35 71 106 93 80 66 53 40 27 14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 106
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:22Z, backlog 14
+      - eval_time: 14m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:24Z, backlog 0
+      - eval_time: 16m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-21: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 100 at 02:15Z, fill 33.3/min, drain 14.3/min -> 9 samples above 5 (8m pending vs for: 20m).
+  - name: nightly burst 2026-09-21 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 0 33 67 100 86 71 57 43 28 14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 100
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:21Z, backlog 14
+      - eval_time: 13m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:23Z, backlog 0
+      - eval_time: 15m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-22: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 101 at 02:15Z, fill 33.7/min, drain 14.4/min -> 9 samples above 5 (8m pending vs for: 20m).
+  - name: nightly burst 2026-09-22 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 0 34 67 101 87 72 58 43 29 15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 101
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:21Z, backlog 15
+      - eval_time: 13m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:23Z, backlog 0
+      - eval_time: 15m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-23: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 119 at 02:15Z, fill 39.7/min, drain 14.9/min -> 10 samples above 5 (9m pending vs for: 20m).
+  - name: nightly burst 2026-09-23 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 0 40 79 119 104 89 74 59 44 30 15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 119
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:22Z, backlog 15
+      - eval_time: 14m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:24Z, backlog 0
+      - eval_time: 16m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-24: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 56 at 02:14Z, fill 18.7/min, drain 14.0/min -> 6 samples above 5 (5m pending vs for: 20m).
+  - name: nightly burst 2026-09-24 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 19 37 56 42 28 14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:14Z, backlog 56
+      - eval_time: 6m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:17Z, backlog 14
+      - eval_time: 9m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:19Z, backlog 0
+      - eval_time: 11m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-25: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 91 at 02:15Z, fill 22.8/min, drain 13.0/min -> 10 samples above 5 (9m pending vs for: 20m).
+  - name: nightly burst 2026-09-25 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 23 46 68 91 78 65 52 39 26 13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 91
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:21Z, backlog 13
+      - eval_time: 13m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:23Z, backlog 0
+      - eval_time: 15m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # 2026-09-26: the nightly Home Assistant recorder purge at 04:12 local.
+  # peak 113 at 02:15Z, fill 28.5/min, drain 14.1/min -> 11 samples above 5 (10m pending vs for: 20m).
+  - name: nightly burst 2026-09-26 — recorder purge, benign, self-clearing
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 28 57 86 113 99 85 71 57 42 28 14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 113
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:22Z, backlog 14
+      - eval_time: 14m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:24Z, backlog 0
+      - eval_time: 16m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # ---------------------------------------------------------------------
+  # 8. THE ADVERSARIAL ENVELOPE. A night that did not happen, built from
+  # the worst fill (39.7/min, 09-23), the highest peak (119, 09-23) and
+  # the slowest drain (13.0/min, 09-25) in the week. 11 samples above 5,
+  # 10m pending against `for: 20m`.
+  # ---------------------------------------------------------------------
+  # Synthetic: max fill 39.7 + max peak 119 + min drain 13.0.
+  # peak 119 at 02:15Z, fill 39.7/min, drain 13.0/min -> 11 samples above 5 (10m pending vs for: 20m).
+  - name: adversarial envelope — worst measured fill and peak, slowest drain
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08Z .. 02:47Z, one sample per minute
+        values: '0 0 0 0 0 40 79 119 106 93 80 67 54 41 28 15 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # The cumulative archived counter on the same metric name. The
+        # rules must never read it; see mutations.txt:no-value-selector.
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      # 02:15Z, backlog 119
+      - eval_time: 7m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:23Z, backlog 15
+      - eval_time: 15m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:26Z, backlog 0
+      - eval_time: 18m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # ---------------------------------------------------------------------
+  # 9. THE WORST NIGHT ON RECORD, 2026-09-11, as raw 30s samples (see
+  # WORST_NIGHT in the generator). Peak 141, 26 samples above 5 = 12.5m
+  # of true expression. At `evaluation_interval: 1m` the rule sees it
+  # true from 02:13Z to 02:25Z: 12m pending against `for: 20m`,
+  # 8m of margin. This is the case that makes the margin
+  # load-bearing: at `for: 12m` it fires at 02:25Z, and nothing else in
+  # the suite notices (mutations.txt:backlog-for-12m).
+  # ---------------------------------------------------------------------
+  - name: worst night on record 2026-09-11 — peak 141, raw samples
+    interval: 30s
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # 02:08:00Z .. 02:47:00Z, one sample per 30s
+        values: '0 0 0 0 0 0 0 0 0 28 53 80 110 141 136 133 130 125 117 117 115 112 102 91 81 75 68 60 52 45 37 29 22 15 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        values: '148000 148007 148014 148021 148028 148035 148042 148049 148056 148063 148070 148077 148084 148091 148098 148105 148112 148119 148126 148133 148140 148147 148154 148161 148168 148175 148182 148189 148196 148203 148210 148217 148224 148231 148238 148245 148252 148259 148266 148273 148280 148287 148294 148301 148308 148315 148322 148329 148336 148343 148350 148357 148364 148371 148378 148385 148392 148399 148406 148413 148420 148427 148434 148441 148448 148455 148462 148469 148476 148483 148490 148497 148504 148511 148518 148525 148532 148539 148546'
+    alert_rule_test:
+      # 02:14Z, backlog 110
+      - eval_time: 6m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:25Z, backlog 8
+      - eval_time: 17m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      # 02:47Z, backlog 0
+      - eval_time: 39m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # ---------------------------------------------------------------------
+  # 10. THE FAILURE THE RULE EXISTS FOR: an archiver that stopped with
+  # 70 segments (1.1 GiB) queued. Not a burst — it does not drain. The
+  # backlog crosses 5 at t=5m, so `for: 20m` elapses at t=25m.
+  # The assertions bracket the transition: one minute early nothing is
+  # firing, on the minute it is.
+  # ---------------------------------------------------------------------
+  - name: stalled archiver at 70 segments — must fire
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        values: '0 0 0 0 0 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        # Flat, because a stalled archiver archives nothing. The rule does
+        # not read this; it is here so the input stays realistic.
+        values: '148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000 148000'
+    alert_rule_test:
+      - eval_time: 24m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      - eval_time: 25m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts:
+          # No `role` label: proof that `max by (namespace, pod)` aggregated it away. See mutations.txt:no-max-by.
+          - exp_labels:
+              severity: warning
+              namespace: home-assistant
+              pod: ha-pg-1
+            exp_annotations:
+              summary: "CNPG WAL archiving is falling behind on home-assistant/ha-pg-1"
+              description: "70 WAL segments have been waiting to archive for over 20 minutes. NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local) is normal and drains itself within about 13 minutes, and this rule is tuned to stay silent through it — so if this fired, the backlog did NOT clear and the archiver is stuck. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n home-assistant logs ha-pg-1 -c plugin-barman-cloud`."
+
+  # ---------------------------------------------------------------------
+  # 11. THE STALL A DEPTH THRESHOLD CANNOT SEE: 8 segments, stuck. It
+  # would never reach any number tuned above the nightly envelope, and
+  # point-in-time recovery is just as broken at 8 segments as at 70. This
+  # is why REN-152 keyed the rule on persistence and retired the depth rule.
+  # ---------------------------------------------------------------------
+  - name: low stall at 8 segments — must fire
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        values: '0 0 0 0 0 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8'
+    alert_rule_test:
+      - eval_time: 24m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      - eval_time: 25m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+              namespace: home-assistant
+              pod: ha-pg-1
+            exp_annotations:
+              summary: "CNPG WAL archiving is falling behind on home-assistant/ha-pg-1"
+              description: "8 WAL segments have been waiting to archive for over 20 minutes. NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local) is normal and drains itself within about 13 minutes, and this rule is tuned to stay silent through it — so if this fired, the backlog did NOT clear and the archiver is stuck. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n home-assistant logs ha-pg-1 -c plugin-barman-cloud`."
+
+  # ---------------------------------------------------------------------
+  # 12. PINNED AT EXACTLY 5 for the whole window. `> 5` is strict, so
+  # this is silence. A rule written `>= 5` would page on it, and a rule
+  # that had lost its `value="ready"` selector would page on the `done`
+  # counter regardless of what `ready` does — which is what the second
+  # series here is for.
+  # ---------------------------------------------------------------------
+  - name: pinned at exactly 5 — strict > means silence
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        values: '5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="done"}'
+        values: '148000 148014 148028 148042 148056 148070 148084 148098 148112 148126 148140 148154 148168 148182 148196 148210 148224 148238 148252 148266 148280 148294 148308 148322 148336 148350 148364 148378 148392 148406 148420 148434 148448 148462 148476 148490 148504 148518 148532 148546'
+    alert_rule_test:
+      - eval_time: 39m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+
+  # ---------------------------------------------------------------------
+  # 13. FAILOVER DURING A STALL — the case `max by (namespace, pod)` was
+  # added for. The backlog starts at t=5m on the primary; at t=11m the
+  # instance is promoted and `role` changes, which is a NEW SERIES to
+  # Prometheus and restarts the `for` timer on the raw expression.
+  # Aggregating `role` away keeps one continuous series, so the page
+  # still lands 20m after the backlog started (t=25m) rather
+  # than 20m after the failover. Without `max by` this case is
+  # silent at t=25m — see mutations.txt:no-max-by.
+  # ---------------------------------------------------------------------
+  - name: role flip mid-stall — for timer must not restart
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        # t=0..10m, then this series ends.
+        values: '0 0 0 0 0 70 70 70 70 70 70'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="replica",value="ready"}'
+        # Starts at t=11m. `_` is a gap, not a zero.
+        values: '_ _ _ _ _ _ _ _ _ _ _ 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70'
+    alert_rule_test:
+      - eval_time: 24m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts: []
+      - eval_time: 25m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+              namespace: home-assistant
+              pod: ha-pg-1
+            exp_annotations:
+              summary: "CNPG WAL archiving is falling behind on home-assistant/ha-pg-1"
+              description: "70 WAL segments have been waiting to archive for over 20 minutes. NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local) is normal and drains itself within about 13 minutes, and this rule is tuned to stay silent through it — so if this fired, the backlog did NOT clear and the archiver is stuck. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n home-assistant logs ha-pg-1 -c plugin-barman-cloud`."
+
+  # ---------------------------------------------------------------------
+  # 14. ONE STALLED INSTANCE AMONG HEALTHY ONES. There are 12 CNPG
+  # instances in this cluster. `by (namespace, pod)` has to keep them
+  # apart: exactly one alert, naming the pod that is actually stuck, so
+  # the 04:00 reader is not sent to look at the wrong Postgres.
+  # ---------------------------------------------------------------------
+  - name: one stalled pod among healthy ones — exactly one alert
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-1",container="postgres",role="primary",value="ready"}'
+        values: '0 0 0 0 0 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70 70'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="home-assistant",pod="ha-pg-2",container="postgres",role="replica",value="ready"}'
+        values: '0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+      - series: 'cnpg_collector_pg_wal_archive_status{namespace="immich",pod="immich-pg-1",container="postgres",role="primary",value="ready"}'
+        values: '0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+    alert_rule_test:
+      - eval_time: 25m
+        alertname: CNPGWalArchiveBacklog
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+              namespace: home-assistant
+              pod: ha-pg-1
+            exp_annotations:
+              summary: "CNPG WAL archiving is falling behind on home-assistant/ha-pg-1"
+              description: "70 WAL segments have been waiting to archive for over 20 minutes. NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local) is normal and drains itself within about 13 minutes, and this rule is tuned to stay silent through it — so if this fired, the backlog did NOT clear and the archiver is stuck. Point-in-time recovery is broken for anything after the last archived segment, and pg_wal will grow until the PVC fills. Check the barman-cloud sidecar logs and the ObjectStore credentials: `kubectl -n home-assistant logs ha-pg-1 -c plugin-barman-cloud`."
+
+  # ---------------------------------------------------------------------
+  # 15. THE EPOCH-ZERO GUARD on CNPGBackupFailed. The metric is a unix
+  # timestamp and reads 0 when no failure was ever recorded, so the rule
+  # carries a `> 0` guard to stop that default being read as "failed in
+  # 1970". Held at 0 for the whole window: silence. Nothing else in this
+  # suite exercises that rule, and a dropped guard would page on every
+  # instance in the cluster forever.
+  # ---------------------------------------------------------------------
+  - name: never-failed backup reads timestamp 0 — must not fire
+    interval: 1m
+    input_series:
+      - series: 'cnpg_collector_last_failed_backup_timestamp{namespace="home-assistant",pod="ha-pg-1"}'
+        values: '0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0'
+    alert_rule_test:
+      - eval_time: 39m
+        alertname: CNPGBackupFailed
+        exp_alerts: []

kubernetes/apps/monitoring-rules/tests/external-secrets_test.yaml

@@ -0,0 +1,115 @@
+# promtool test rules for prometheusrule-external-secrets.yaml.
+#
+# Run by the `promtool` step of .woodpecker/ci-manifests-test.yml through
+# scripts/ci/promtool-rules.sh, which strips `spec:` from the manifest into a
+# mirrored tree — hence `../` below, resolved from this file's directory. See
+# README.md here. The negative controls for this suite are the `eso-*` entries
+# in mutations.txt.
+#
+# WHAT IT IS FOR. Every claim in the rule file that a reviewer would otherwise
+# have to take on trust is asserted here instead: that the tombstone series does
+# not alert, that `for: 15m` really is 15m, that the annotation names the
+# ExternalSecret's own namespace and not the controller's, and that `max`
+# survives a two-pod scrape where `sum` does not.
+#
+# THE REJECTED FORM, for the record. REN-100 and ESO's shipped dashboard both
+# use `sum(...) by (namespace, name) == 1`. Run against the identical input
+# series below, that expression produces exactly one alert,
+# {namespace="external-secrets", name="gitea-db"} — the wrong namespace on the
+# one it catches, and nothing at all for monitoring/loki-basic-auth, whose two
+# scraped pods sum to 2 and stop matching `== 1`. Both defects, one expression.
+rule_files:
+  - ../prometheusrule-external-secrets.yaml
+
+evaluation_interval: 1m
+
+tests:
+  # Series shapes copied from the live cluster after #994 landed: the metric's
+  # own `namespace` has already been shadowed to `exported_namespace` by the
+  # PodMonitor's target label, and `namespace` is the controller pod's.
+  - interval: 1m
+    input_series:
+      # A healthy ExternalSecret. ESO writes BOTH polarities of the Ready
+      # condition on every reconcile, the inactive one at 0 — which is why the
+      # rule needs `== 1` and not a bare selector.
+      - series: 'externalsecret_status_condition{condition="Ready",status="True",name="uptime-robot-api-key",exported_namespace="uptime-monitor",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '1+0x60'
+      - series: 'externalsecret_status_condition{condition="Ready",status="False",name="uptime-robot-api-key",exported_namespace="uptime-monitor",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '0+0x60'
+
+      # Starts failing at t=10m and stays failed.
+      - series: 'externalsecret_status_condition{condition="Ready",status="False",name="gitea-db",exported_namespace="gitea",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '0+0x9 1+0x50'
+      - series: 'externalsecret_status_condition{condition="Ready",status="True",name="gitea-db",exported_namespace="gitea",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '1+0x9 0+0x50'
+
+      # Tombstone left behind for an ExternalSecret that no longer exists. Held
+      # at 1 forever until the controller restarts, so a rule that matched
+      # `condition!="Ready"` or ignored `status` would page on it every day.
+      - series: 'externalsecret_status_condition{condition="Deleted",status="False",name="wiki-index-basic-auth",exported_namespace="ai",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '1+0x60'
+
+      # The same failing ExternalSecret reported by TWO controller pods, as
+      # happens for the seconds of a Deployment rollout when Prometheus still
+      # has the old target and already has the new one.
+      - series: 'externalsecret_status_condition{condition="Ready",status="False",name="loki-basic-auth",exported_namespace="monitoring",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-56c649799b-t65gf"}'
+        values: '0+0x9 1+0x50'
+      - series: 'externalsecret_status_condition{condition="Ready",status="False",name="loki-basic-auth",exported_namespace="monitoring",namespace="external-secrets",job="monitoring/external-secrets",pod="external-secrets-7d4f9c6b88-qq2xv"}'
+        values: '0+0x9 1+0x50'
+
+    alert_rule_test:
+      # 5m into the failure: pending, not firing. Asserts the window is real —
+      # a `for:` accidentally set to 0 passes every other case here and fails
+      # only this one.
+      - eval_time: 15m
+        alertname: ExternalSecretSyncFailing
+        exp_alerts: []
+
+      # 20m into the failure. Both failing ExternalSecrets fire, ONCE each; the
+      # healthy one and the tombstone stay silent. The annotations are asserted
+      # in full because the whole point of REN-103's label finding is that the
+      # text names `gitea`, not `external-secrets`.
+      - eval_time: 30m
+        alertname: ExternalSecretSyncFailing
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+              exported_namespace: gitea
+              name: gitea-db
+            exp_annotations:
+              summary: "ExternalSecret gitea/gitea-db is not syncing"
+              description: "ESO has been unable to sync this ExternalSecret for 15 minutes. The Secret it manages still holds its last successfully synced value, so consumers keep working — what breaks is that a rotation at the provider no longer propagates, which surfaces much later as a credential that stopped being accepted. Get the reason with `kubectl -n gitea get externalsecret gitea-db -o jsonpath='{.status.conditions}'`; the common causes are a key that was renamed or removed in OpenBao, a ClusterSecretStore that cannot authenticate, and a target Secret owned by something else."
+          - exp_labels:
+              severity: warning
+              exported_namespace: monitoring
+              name: loki-basic-auth
+            exp_annotations:
+              summary: "ExternalSecret monitoring/loki-basic-auth is not syncing"
+              description: "ESO has been unable to sync this ExternalSecret for 15 minutes. The Secret it manages still holds its last successfully synced value, so consumers keep working — what breaks is that a rotation at the provider no longer propagates, which surfaces much later as a credential that stopped being accepted. Get the reason with `kubectl -n monitoring get externalsecret loki-basic-auth -o jsonpath='{.status.conditions}'`; the common causes are a key that was renamed or removed in OpenBao, a ClusterSecretStore that cannot authenticate, and a target Secret owned by something else."
+
+      # Metrics are present, so the dead-man's-switch stays quiet.
+      - eval_time: 30m
+        alertname: ExternalSecretMetricsMissing
+        exp_alerts: []
+
+  # The scrape disappears entirely: PodMonitor deleted, selector broken by a
+  # chart bump, or the `monitoring` ingress grant reverted out of
+  # external-secrets-ns.yaml. ExternalSecretSyncFailing cannot fire on series
+  # that do not exist — this is the case that made REN-103 necessary, so it is
+  # the one the dead-man's-switch has to cover.
+  - interval: 1m
+    input_series:
+      - series: 'up{job="monitoring/external-secrets"}'
+        values: '0+0x60'
+    alert_rule_test:
+      - eval_time: 30m
+        alertname: ExternalSecretSyncFailing
+        exp_alerts: []
+      - eval_time: 30m
+        alertname: ExternalSecretMetricsMissing
+        exp_alerts:
+          - exp_labels:
+              severity: warning
+            exp_annotations:
+              summary: "No ExternalSecret metrics are reaching Prometheus"
+              description: "`externalsecret_status_condition` has no series. Either the ESO controller is gone or its scrape stopped being collected, and ExternalSecretSyncFailing is blind for as long as this is true. Check the PodMonitor `monitoring/external-secrets` still selects the controller pods, and that external-secrets-ns.yaml still admits `monitoring` on TCP/8080."

kubernetes/apps/monitoring-rules/tests/generate-cnpg-backup-test.py

@@ -0,0 +1,555 @@
+#!/usr/bin/env python3
+"""Regenerate cnpg-backup_test.yaml from the REN-129 measurements.
+
+The committed YAML is what CI runs; this script exists so the numbers in it are
+auditable and reproducible from the measurement table rather than being a
+hand-typed artifact nobody can re-derive. Run it from the repo root:
+
+    python3 kubernetes/apps/monitoring-rules/tests/generate-cnpg-backup-test.py \
+      > kubernetes/apps/monitoring-rules/tests/cnpg-backup_test.yaml
+
+Nothing in CI regenerates and diffs this — the promtool image has no Python, and
+the defence against a hand-weakened suite is the mutation phase of
+scripts/ci/promtool-rules.sh, not byte-equality with a generator. See README.md.
+
+THE TRACE MODEL. Iris measured seven consecutive nights of
+`cnpg_collector_pg_wal_archive_status{value="ready",pod="ha-pg-1"}` at 1-minute
+steps and published, per night, a (peak, peak minute, fill seg/min, drain
+seg/min) tuple. Those four numbers are consistent with one shape, which is the
+shape every night has: flat 0, a linear ramp at `fill` per minute to `peak`,
+then a linear decay at `drain` per minute back to 0. So a night is rebuilt as
+
+    ramp length  = round(peak / fill)          minutes
+    ramp start   = peak minute - ramp length   (value 0 at that minute)
+    during ramp  = round(fill * minutes since ramp start)
+    at peak      = peak                        (exact, not rounded from fill)
+    during decay = max(0, round(peak - drain * minutes since peak))
+
+FIDELITY. The reconstruction reproduces Iris's published `Duration >5` column
+exactly on 2026-09-20, -21 and -22, and runs one minute long on -23 and -24 and
+two minutes long on -25 and -26. Every discrepancy is in the same direction:
+the rebuild is never SHORTER above the threshold than the measurement. A linear
+decay at the published mean drain rate holds the tail above 5 for slightly
+longer than the real trace did, so each night is a marginally harder silence
+test than the night it stands for. That is the safe direction for this suite,
+whose whole job is to catch a rule that fires on a benign night. `--report`
+prints the comparison; do not "fix" it toward the published column without
+checking which way each night moves.
+
+These are reconstructions, not the raw samples. They are faithful in the only
+dimensions the rules read: how high the backlog goes, and how many consecutive
+minutes it spends above each threshold. Do not read them as a recording.
+"""
+
+import argparse
+import sys
+
+# Iris's 7-night measurement, REN-129. `dur_gt5` is her published "Duration >5"
+# column and is an INPUT TO THE SELF-CHECK ONLY — the traces are built from
+# peak/fill/drain/peak_min, and dur_gt5 is what the result is compared against.
+NIGHTS = [
+    # date        peak  peak_min  fill  drain  dur_gt5
+    ("2026-09-20", 106, "02:15", 35.3, 13.2, 10),
+    ("2026-09-21", 100, "02:15", 33.3, 14.3, 9),
+    ("2026-09-22", 101, "02:15", 33.7, 14.4, 9),
+    ("2026-09-23", 119, "02:15", 39.7, 14.9, 9),
+    ("2026-09-24", 56, "02:14", 18.7, 14.0, 5),
+    ("2026-09-25", 91, "02:15", 22.8, 13.0, 8),
+    ("2026-09-26", 113, "02:15", 28.5, 14.1, 9),
+]
+
+# The adversarial envelope: worst fill, worst peak, slowest drain observed in the
+# week, combined into a night that did not happen but is inside what the cluster
+# has demonstrated it can do. Nothing here is extrapolated past the measurements.
+ENVELOPE = ("envelope", 119, "02:15", 39.7, 13.0, None)
+
+# The worst night in Prometheus' 27-night retention (2026-08-30..09-26):
+# 2026-09-11, peak 141, the longest excursion above 5 on record. Not rebuilt
+# from a model like NIGHTS — these are the raw samples, read back through the
+# API server's services/proxy with
+#   max by (namespace, pod) (cnpg_collector_pg_wal_archive_status{
+#     value="ready", pod="ha-pg-1"})
+# over 02:08:00Z..02:47:00Z at the real 30s scrape step (REN-152). 26
+# consecutive samples above 5 = 12.5 minutes pending, which is the number the
+# `for: 20m` margin is measured against, and why mutations.txt:backlog-for-12m
+# is caught here and nowhere else.
+WORST_NIGHT = ("2026-09-11", [
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 28, 53, 80, 110, 141, 136, 133, 130, 125, 117,
+    117, 115, 112, 102, 91, 81, 75, 68, 60, 52, 45, 37, 29, 22, 15, 8, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
+])
+
+FOR_MIN = 20  # `for:` on CNPGWalArchiveBacklog, in minutes
+
+# Series labels. `container`, `role` and `instance` are on the real series and
+# are carried here on purpose: they are what `max by (namespace, pod)` has to
+# aggregate away, and a test whose input is already aggregated cannot see whether
+# the rule does that. `value="done"` is the cumulative counter that sits on the
+# same metric name — it is in the input so that dropping the `value="ready"`
+# selector from the rules is a detectable mutation and not a silent no-op.
+NS = "home-assistant"
+POD = "ha-pg-1"
+
+
+def labels(pod=POD, value="ready", role="primary", ns=NS):
+    return (
+        "cnpg_collector_pg_wal_archive_status{"
+        f'namespace="{ns}",pod="{pod}",container="postgres",'
+        f'role="{role}",value="{value}"'
+        "}"
+    )
+
+
+# Test window: t=0 is 02:08Z, 40 one-minute samples, so every night's burst sits
+# inside it with lead-in and tail. promtool eval_times are relative to t=0, so
+# every assertion below carries its UTC clock time in a comment.
+#
+# This MUST stay aligned with the peak minutes in NIGHTS. An earlier revision
+# had `8 * 60 + 8` here (08:08Z), which put the whole window seven hours after
+# every burst: all seven nights rebuilt as 40 zeroes and their ~90 silence
+# assertions passed without exercising a single threshold. The
+# assert_window_covers_bursts() call in main() exists so that cannot recur
+# silently.
+T0 = 2 * 60 + 8  # 02:08Z in minutes since midnight
+SAMPLES = 40
+
+
+def hhmm(minute):
+    return f"{minute // 60:02d}:{minute % 60:02d}"
+
+
+def parse_hhmm(s):
+    h, m = s.split(":")
+    return int(h) * 60 + int(m)
+
+
+def trace(peak, peak_min, fill, drain, n=SAMPLES, t0=T0):
+    """Rebuild one night. See THE TRACE MODEL in the module docstring."""
+    ramp_len = max(1, round(peak / fill))
+    ramp_start = peak_min - ramp_len
+    out = []
+    for m in range(t0, t0 + n):
+        if m <= ramp_start:
+            v = 0
+        elif m < peak_min:
+            v = round(fill * (m - ramp_start))
+        elif m == peak_min:
+            v = peak
+        else:
+            v = max(0, round(peak - drain * (m - peak_min)))
+        out.append(v)
+    return out
+
+
+def longest_run_above(vals, thr):
+    """(length, index of last sample) of the longest run of vals > thr."""
+    best = (0, None)
+    run = 0
+    for i, v in enumerate(vals):
+        if v > thr:
+            run += 1
+            if run > best[0]:
+                best = (run, i)
+        else:
+            run = 0
+    return best
+
+
+def pending_minutes(vals, thr):
+    """Max minutes a `for` timer accumulates on `> thr` over this trace.
+
+    This is NOT the run length, and the difference is the whole margin
+    argument. Prometheus makes an alert pending at the first evaluation where
+    the expression is true and fires it when `now - activeAt >= for`, so a run
+    of N consecutive 1-minute samples above the threshold accumulates N-1
+    minutes of pending time, not N. A trace that spends 5 samples above 64
+    therefore does NOT fire `for: 5m`; it fires `for: 4m`. Quote this number,
+    not the run length, when arguing about a `for` window.
+    """
+    n, _ = longest_run_above(vals, thr)
+    return max(0, n - 1)
+
+
+def assert_window_covers_bursts():
+    """Refuse to emit a suite whose trace window misses the bursts.
+
+    Guards the T0 bug described above T0: if the window and the peak minutes
+    ever drift apart again, every measured night rebuilds as flat zero and the
+    suite passes while testing nothing. A night that is flat 0 cannot be a
+    silence test for a burst, so treat it as a generator fault, not a result.
+    """
+    for date, peak, pm, fill, drain, _dur in NIGHTS + [ENVELOPE]:
+        vals = trace(peak, parse_hhmm(pm), fill, drain)
+        if max(vals) != peak:
+            raise SystemExit(
+                f"generator fault: {date} rebuilt with max {max(vals)}, expected "
+                f"peak {peak}. The T0={T0} window ({hhmm(T0)}Z.."
+                f"{hhmm(T0 + SAMPLES - 1)}Z) does not cover peak minute {pm}Z. "
+                "Fix T0 or SAMPLES; do not emit a suite of flat-zero nights."
+            )
+
+
+def counter(start, step, n=SAMPLES):
+    return [start + step * i for i in range(n)]
+
+
+def fmt(vals):
+    return " ".join(str(v) for v in vals)
+
+
+def eval_minutes(vals):
+    """The minutes where a `for` window is closest to elapsing, so the tightest
+    places to assert silence. For each threshold: the last sample of the longest
+    run above it (the moment maximum pending time has accumulated). Plus the
+    peak, plus two minutes after the backlog has settled back to 0."""
+    picks = set()
+    _, last = longest_run_above(vals, 5)
+    if last is not None:
+        picks.add(last)
+    picks.add(max(range(len(vals)), key=lambda i: vals[i]))
+    settled = next(
+        (i for i in range(len(vals) - 1, -1, -1) if vals[i] > 0),
+        len(vals) - 1,
+    )
+    picks.add(min(settled + 2, len(vals) - 1))
+    return sorted(picks)
+
+
+L = []
+
+
+def w(line=""):
+    L.append(line)
+
+
+def burst_case(name, date, peak, peak_min_s, fill, drain, why):
+    peak_min = parse_hhmm(peak_min_s)
+    vals = trace(peak, peak_min, fill, drain)
+    n5, _ = longest_run_above(vals, 5)
+    w(f"  # {why}")
+    w(
+        f"  # peak {peak} at {peak_min_s}Z, fill {fill}/min, drain {drain}/min"
+        f" -> {n5} samples above 5 ({pending_minutes(vals, 5)}m pending vs"
+        f" for: {FOR_MIN}m)."
+    )
+    w(f"  - name: {name}")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels()}'")
+    w(f"        # {hhmm(T0)}Z .. {hhmm(T0 + SAMPLES - 1)}Z, one sample per minute")
+    w(f"        values: '{fmt(vals)}'")
+    w(f"      - series: '{labels(value='done')}'")
+    w("        # The cumulative archived counter on the same metric name. The")
+    w("        # rules must never read it; see mutations.txt:no-value-selector.")
+    w(f"        values: '{fmt(counter(148_000, 14))}'")
+    w("    alert_rule_test:")
+    for i in eval_minutes(vals):
+        clock = hhmm(T0 + i)
+        w(f"      # {clock}Z, backlog {vals[i]}")
+        w(f"      - eval_time: {i}m")
+        w("        alertname: CNPGWalArchiveBacklog")
+        w("        exp_alerts: []")
+    w("")
+    return vals, n5
+
+
+def desc(n, pod=POD, ns=NS):
+    """The rule's `description` as promtool renders it for `n` segments."""
+    return (
+        f"{n} WAL segments have been waiting to archive for over {FOR_MIN} minutes."
+        " NOTE ON TIMING: a nightly write burst on ha-pg-1 at 02:14Z (04:14 local)"
+        " is normal and drains itself within about 13 minutes, and this rule is"
+        " tuned to stay silent through it — so if this fired, the backlog did NOT"
+        " clear and the archiver is stuck. Point-in-time recovery is broken for"
+        " anything after the last archived segment, and pg_wal will grow until the"
+        " PVC fills. Check the barman-cloud sidecar logs and the ObjectStore"
+        f" credentials: `kubectl -n {ns} logs {pod} -c plugin-barman-cloud`."
+    )
+
+
+def firing(eval_time, n, pod=POD, ns=NS, note=None):
+    w(f"      - eval_time: {eval_time}")
+    w("        alertname: CNPGWalArchiveBacklog")
+    w("        exp_alerts:")
+    if note:
+        w(f"          # {note}")
+    w("          - exp_labels:")
+    w("              severity: warning")
+    w(f"              namespace: {ns}")
+    w(f"              pod: {pod}")
+    w("            exp_annotations:")
+    w(f'              summary: "CNPG WAL archiving is falling behind on {ns}/{pod}"')
+    w(f'              description: "{desc(n, pod, ns)}"')
+
+
+def silent(eval_time, alert="CNPGWalArchiveBacklog"):
+    w(f"      - eval_time: {eval_time}")
+    w(f"        alertname: {alert}")
+    w("        exp_alerts: []")
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument(
+        "--report",
+        action="store_true",
+        help="print the reconstruction self-check table instead of the suite",
+    )
+    args = ap.parse_args()
+
+    assert_window_covers_bursts()
+
+    if args.report:
+        # `pend` is what the `for` window is actually compared against; see
+        # pending_minutes(). `>5 pub` is Iris's published column and is the
+        # only number here that is not derived by this script.
+        print(f"{'night':12} {'peak':>5} {'fill':>6} {'drain':>6} "
+              f"{'>5 pub':>7} {'>5 n':>6} {'>5 pend':>8}")
+        for date, peak, pm, fill, drain, dur in NIGHTS + [ENVELOPE]:
+            vals = trace(peak, parse_hhmm(pm), fill, drain)
+            n5, _ = longest_run_above(vals, 5)
+            print(f"{date:12} {peak:5} {fill:6} {drain:6} "
+                  f"{('-' if dur is None else dur):>7} {n5:>6} "
+                  f"{pending_minutes(vals, 5):>8}")
+        wn, _ = longest_run_above(WORST_NIGHT[1], 5)
+        print(f"{WORST_NIGHT[0]:12} {max(WORST_NIGHT[1]):5} {'raw':>6} {'30s':>6} "
+              f"{'-':>7} {wn:>6} {max(0, wn - 1) / 2:>8}")
+        print()
+        print(f"for: {FOR_MIN}m on >5 must exceed every '>5 pend' above.")
+        return 0
+
+    w("# GENERATED by generate-cnpg-backup-test.py — see README.md in this")
+    w("# directory before editing by hand.")
+    w("#")
+    w("# Unit tests for kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml.")
+    w("# Every assertion below is either one of Iris's seven measured nights")
+    w("# (REN-129) rebuilt from its published (peak, fill, drain) numbers, the")
+    w("# worst night in Prometheus' retention as raw samples (REN-152), an")
+    w("# adversarial combination of the extremes of those numbers, or a failure")
+    w("# this cluster has not had but the rule exists to catch.")
+    w("#")
+    w("# `rule_files` entries are resolved RELATIVE TO THIS FILE'S DIRECTORY, not")
+    w("# to the working directory promtool was invoked from — promtool joins")
+    w("# dirname(testfile) onto each entry. So this is `../`, and it resolves in")
+    w("# the mirrored tree scripts/ci/promtool-rules.sh builds, where the")
+    w("# `spec:`-stripped manifests sit in the same layout as the real ones.")
+    w("#")
+    w("# A path that does NOT resolve is only a WARNING to promtool: it loads zero")
+    w("# rules, every exp_alerts: [] assertion passes against nothing, and the")
+    w("# suite exits 0. promtool-rules.sh resolves every entry itself before")
+    w("# running and fails on a miss; that check is the only thing standing")
+    w("# between a typo here and a green pipeline that tests nothing.")
+    w("rule_files:")
+    w("  - ../prometheusrule-cnpg-backup.yaml")
+    w("evaluation_interval: 1m")
+    w("")
+    w("tests:")
+    w("  # ---------------------------------------------------------------------")
+    w("  # 1-7. THE SEVEN MEASURED NIGHTS. The rule must stay silent through")
+    w("  # every one of them. The retired depth rule (`> 64`) shipped for review")
+    w("  # at `for: 2m` and fired on 6 of these 7 nights; REN-152 removed it.")
+    w("  # ---------------------------------------------------------------------")
+    checks = []
+    for date, peak, pm, fill, drain, dur in NIGHTS:
+        vals, n5 = burst_case(
+            f"nightly burst {date} — recorder purge, benign, self-clearing",
+            date, peak, pm, fill, drain,
+            f"{date}: the nightly Home Assistant recorder purge at 04:12 local.",
+        )
+        checks.append((date, dur, n5))
+
+    w("  # ---------------------------------------------------------------------")
+    w("  # 8. THE ADVERSARIAL ENVELOPE. A night that did not happen, built from")
+    w("  # the worst fill (39.7/min, 09-23), the highest peak (119, 09-23) and")
+    w("  # the slowest drain (13.0/min, 09-25) in the week. 11 samples above 5,")
+    w("  # 10m pending against `for: 20m`.")
+    w("  # ---------------------------------------------------------------------")
+    burst_case(
+        "adversarial envelope — worst measured fill and peak, slowest drain",
+        *ENVELOPE[:5],
+        why="Synthetic: max fill 39.7 + max peak 119 + min drain 13.0.",
+    )
+
+    date, raw = WORST_NIGHT
+    wn, wlast = longest_run_above(raw, 5)
+    peak_i = max(range(len(raw)), key=lambda i: raw[i])
+    w("  # ---------------------------------------------------------------------")
+    w(f"  # 9. THE WORST NIGHT ON RECORD, {date}, as raw 30s samples (see")
+    w("  # WORST_NIGHT in the generator). Peak 141, 26 samples above 5 = 12.5m")
+    w("  # of true expression. At `evaluation_interval: 1m` the rule sees it")
+    w(f"  # true from 02:13Z to 02:25Z: 12m pending against `for: {FOR_MIN}m`,")
+    w(f"  # {FOR_MIN - 12}m of margin. This is the case that makes the margin")
+    w("  # load-bearing: at `for: 12m` it fires at 02:25Z, and nothing else in")
+    w("  # the suite notices (mutations.txt:backlog-for-12m).")
+    w("  # ---------------------------------------------------------------------")
+    w(f"  - name: worst night on record {date} — peak 141, raw samples")
+    w("    interval: 30s")
+    w("    input_series:")
+    w(f"      - series: '{labels()}'")
+    w(f"        # {hhmm(T0)}:00Z .. {hhmm(T0 + 39)}:00Z, one sample per 30s")
+    w(f"        values: '{fmt(raw)}'")
+    w(f"      - series: '{labels(value='done')}'")
+    w(f"        values: '{fmt(counter(148_000, 7, len(raw)))}'")
+    w("    alert_rule_test:")
+    for i in (peak_i, wlast, len(raw) - 1):
+        # eval times in whole minutes: the group evaluates on the minute.
+        m = (i // 2)
+        w(f"      # {hhmm(T0 + m)}Z, backlog {raw[m * 2]}")
+        silent(f"{m}m")
+    w("")
+
+    # ---- 10. stalled archiver, well above threshold ------------------------
+    stalled = [0] * 5 + [70] * 35
+    fire_at = 5 + FOR_MIN
+    w("  # ---------------------------------------------------------------------")
+    w("  # 10. THE FAILURE THE RULE EXISTS FOR: an archiver that stopped with")
+    w("  # 70 segments (1.1 GiB) queued. Not a burst — it does not drain. The")
+    w(f"  # backlog crosses 5 at t=5m, so `for: {FOR_MIN}m` elapses at t={fire_at}m.")
+    w("  # The assertions bracket the transition: one minute early nothing is")
+    w("  # firing, on the minute it is.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: stalled archiver at 70 segments — must fire")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels()}'")
+    w(f"        values: '{fmt(stalled)}'")
+    w(f"      - series: '{labels(value='done')}'")
+    w("        # Flat, because a stalled archiver archives nothing. The rule does")
+    w("        # not read this; it is here so the input stays realistic.")
+    w(f"        values: '{fmt([148_000] * SAMPLES)}'")
+    w("    alert_rule_test:")
+    silent(f"{fire_at - 1}m")
+    firing(f"{fire_at}m", 70, note="No `role` label: proof that `max by (namespace, pod)` "
+           "aggregated it away. See mutations.txt:no-max-by.")
+    w("")
+
+    # ---- 11. low stall ------------------------------------------------------
+    low = [0] * 5 + [8] * 35
+    w("  # ---------------------------------------------------------------------")
+    w("  # 11. THE STALL A DEPTH THRESHOLD CANNOT SEE: 8 segments, stuck. It")
+    w("  # would never reach any number tuned above the nightly envelope, and")
+    w("  # point-in-time recovery is just as broken at 8 segments as at 70. This")
+    w("  # is why REN-152 keyed the rule on persistence and retired the depth rule.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: low stall at 8 segments — must fire")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels()}'")
+    w(f"        values: '{fmt(low)}'")
+    w("    alert_rule_test:")
+    silent(f"{fire_at - 1}m")
+    firing(f"{fire_at}m", 8)
+    w("")
+
+    # ---- 12. exactly at the threshold ------------------------------------
+    w("  # ---------------------------------------------------------------------")
+    w("  # 12. PINNED AT EXACTLY 5 for the whole window. `> 5` is strict, so")
+    w("  # this is silence. A rule written `>= 5` would page on it, and a rule")
+    w("  # that had lost its `value=\"ready\"` selector would page on the `done`")
+    w("  # counter regardless of what `ready` does — which is what the second")
+    w("  # series here is for.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: pinned at exactly 5 — strict > means silence")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels()}'")
+    w(f"        values: '{fmt([5] * SAMPLES)}'")
+    w(f"      - series: '{labels(value='done')}'")
+    w(f"        values: '{fmt(counter(148_000, 14))}'")
+    w("    alert_rule_test:")
+    silent("39m")
+    w("")
+
+    # ---- 13. role flip ---------------------------------------------------
+    w("  # ---------------------------------------------------------------------")
+    w("  # 13. FAILOVER DURING A STALL — the case `max by (namespace, pod)` was")
+    w("  # added for. The backlog starts at t=5m on the primary; at t=11m the")
+    w("  # instance is promoted and `role` changes, which is a NEW SERIES to")
+    w("  # Prometheus and restarts the `for` timer on the raw expression.")
+    w("  # Aggregating `role` away keeps one continuous series, so the page")
+    w(f"  # still lands {FOR_MIN}m after the backlog started (t={fire_at}m) rather")
+    w(f"  # than {FOR_MIN}m after the failover. Without `max by` this case is")
+    w(f"  # silent at t={fire_at}m — see mutations.txt:no-max-by.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: role flip mid-stall — for timer must not restart")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels(role='primary')}'")
+    w("        # t=0..10m, then this series ends.")
+    w(f"        values: '{fmt([0] * 5 + [70] * 6)}'")
+    w(f"      - series: '{labels(role='replica')}'")
+    w("        # Starts at t=11m. `_` is a gap, not a zero.")
+    w(f"        values: '{'_ ' * 11}{fmt([70] * 29)}'")
+    w("    alert_rule_test:")
+    silent(f"{fire_at - 1}m")
+    firing(f"{fire_at}m", 70)
+    w("")
+
+    # ---- 14. one stalled pod among healthy ones --------------------------
+    w("  # ---------------------------------------------------------------------")
+    w("  # 14. ONE STALLED INSTANCE AMONG HEALTHY ONES. There are 12 CNPG")
+    w("  # instances in this cluster. `by (namespace, pod)` has to keep them")
+    w("  # apart: exactly one alert, naming the pod that is actually stuck, so")
+    w("  # the 04:00 reader is not sent to look at the wrong Postgres.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: one stalled pod among healthy ones — exactly one alert")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: '{labels(pod='ha-pg-1')}'")
+    w(f"        values: '{fmt([0] * 5 + [70] * 35)}'")
+    w(f"      - series: '{labels(pod='ha-pg-2', role='replica')}'")
+    w(f"        values: '{fmt([0] * SAMPLES)}'")
+    w(f"      - series: '{labels(pod='immich-pg-1', ns='immich')}'")
+    w(f"        values: '{fmt([0] * SAMPLES)}'")
+    w("    alert_rule_test:")
+    firing(f"{fire_at}m", 70, pod="ha-pg-1")
+    w("")
+
+    # ---- 15. CNPGBackupFailed --------------------------------------------
+    w("  # ---------------------------------------------------------------------")
+    w("  # 15. THE EPOCH-ZERO GUARD on CNPGBackupFailed. The metric is a unix")
+    w("  # timestamp and reads 0 when no failure was ever recorded, so the rule")
+    w("  # carries a `> 0` guard to stop that default being read as \"failed in")
+    w("  # 1970\". Held at 0 for the whole window: silence. Nothing else in this")
+    w("  # suite exercises that rule, and a dropped guard would page on every")
+    w("  # instance in the cluster forever.")
+    w("  # ---------------------------------------------------------------------")
+    w("  - name: never-failed backup reads timestamp 0 — must not fire")
+    w("    interval: 1m")
+    w("    input_series:")
+    w(f"      - series: 'cnpg_collector_last_failed_backup_timestamp{{namespace=\"{NS}\",pod=\"{POD}\"}}'")
+    w(f"        values: '{fmt([0] * SAMPLES)}'")
+    w("    alert_rule_test:")
+    silent("39m", "CNPGBackupFailed")
+
+    print("\n".join(L))
+
+    # Fidelity self-check on stderr so it is visible when regenerating but does
+    # not contaminate the suite on stdout. Longer than measured is tolerated and
+    # expected (see FIDELITY in the module docstring); SHORTER is not, because a
+    # night that clears the threshold sooner than the real one is a weaker
+    # silence test than the night it claims to stand for.
+    short = [(d, m, r) for d, m, r in checks if r < m]
+    if short:
+        print(
+            "FAIL: reconstruction is SHORTER above 5 than the measurement on "
+            + ", ".join(f"{d} measured {m} rebuilt {r}" for d, m, r in short)
+            + " — these nights are weaker silence tests than the real ones. "
+            "Refusing to vouch for the suite.",
+            file=sys.stderr,
+        )
+        return 1
+    long_ = [(d, m, r) for d, m, r in checks if r > m]
+    if long_:
+        print(
+            f"NOTE: {len(long_)} of {len(checks)} nights rebuild LONGER above 5 "
+            "than the published column (conservative direction): "
+            + ", ".join(f"{d} {m}->{r}" for d, m, r in long_),
+            file=sys.stderr,
+        )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

kubernetes/apps/monitoring-rules/tests/home-assistant_test.yaml

@@ -0,0 +1,223 @@
+# promtool test rules for prometheusrule-home-assistant.yaml (REN-245).
+#
+# Run by the `promtool` step of .woodpecker/ci-manifests-test.yml through
+# scripts/ci/promtool-rules.sh — see README.md here. Negative controls are the
+# `ha-*` entries in mutations.txt.
+#
+# WHAT IT IS FOR. Each rule claims three things in its comments: it fires 5m
+# after the outage starts and not before, it stays silent on a healthy system,
+# and one outage is one alert even when the outage changes shape. Each claim
+# is asserted here. Series names and labels follow the live shapes recorded in
+# the REN-244 "findings" document.
+rule_files:
+  - ../prometheusrule-home-assistant.yaml
+
+evaluation_interval: 1m
+
+tests:
+  # --- HomeAssistantDown --------------------------------------------------
+  # Target answers for 10m, fails for 5m, disappears (Recreate rollout) for
+  # 3m, then comes back still failing. One outage from t=10m: the alert must
+  # fire at t=15m and still be the same alert after the target changed shape.
+  - interval: 1m
+    input_series:
+      - series: 'up{job="home-assistant",namespace="home-assistant",instance="10.0.0.1:8124",pod="home-assistant-6d9c7b5f4-abcde"}'
+        values: '1x9 0x4'
+      - series: 'up{job="home-assistant",namespace="home-assistant",instance="10.0.0.2:8124",pod="home-assistant-6d9c7b5f4-fghij"}'
+        values: '_x22 0x10'
+    alert_rule_test:
+      - eval_time: 14m
+        alertname: HomeAssistantDown
+        exp_alerts: []
+      - eval_time: 15m
+        alertname: HomeAssistantDown
+        exp_alerts:
+          - exp_labels: {severity: critical, job: home-assistant, namespace: home-assistant}
+            exp_annotations:
+              summary: "Home Assistant has been down for 5 minutes"
+              description: "The home-assistant scrape target in namespace home-assistant is down or missing. A restart after an automatic image upgrade can take up to 10 minutes, so if an upgrade just rolled out, give it until then before acting. Rollback: kubernetes/RUNBOOK-home-assistant-image-rollback.md."
+      # Old target gone (stale after 5m lookback), new one failing: same alert.
+      - eval_time: 25m
+        alertname: HomeAssistantDown
+        exp_alerts:
+          - exp_labels: {severity: critical, job: home-assistant, namespace: home-assistant}
+            exp_annotations:
+              summary: "Home Assistant has been down for 5 minutes"
+              description: "The home-assistant scrape target in namespace home-assistant is down or missing. A restart after an automatic image upgrade can take up to 10 minutes, so if an upgrade just rolled out, give it until then before acting. Rollback: kubernetes/RUNBOOK-home-assistant-image-rollback.md."
+
+  # Healthy HA stays silent.
+  - interval: 1m
+    input_series:
+      - series: 'up{job="home-assistant",namespace="home-assistant",instance="10.0.0.1:8124",pod="home-assistant-6d9c7b5f4-abcde"}'
+        values: '1x30'
+    alert_rule_test:
+      - eval_time: 30m
+        alertname: HomeAssistantDown
+        exp_alerts: []
+
+  # --- HomeAssistantRecorderDBUnhealthy ----------------------------------
+  # Two instances as in ha-pg-cluster.yaml. The replica's Postgres goes down at
+  # t=10m (its exporter keeps answering with cnpg_collector_up=0).
+  - interval: 1m
+    input_series:
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-2",role="primary",cluster="ha-pg"}'
+        values: '1x30'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-1",role="replica",cluster="ha-pg"}'
+        values: '1x9 0x20'
+    alert_rule_test:
+      - eval_time: 9m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts: []
+      - eval_time: 14m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts: []
+      - eval_time: 15m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Recorder database ha-pg has no primary or is missing an instance"
+              description: "In CNPG cluster home-assistant/ha-pg either no instance reports role=primary with Postgres up, or fewer than 2 instances have Postgres up. Without a primary the Home Assistant recorder cannot write history. Check `kubectl -n home-assistant get cluster ha-pg` and the pods' logs."
+
+  # Primary's Postgres goes down at t=10m. Both branches are true at once (no
+  # primary AND only one instance up); still exactly one alert.
+  - interval: 1m
+    input_series:
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-2",role="primary",cluster="ha-pg"}'
+        values: '1x9 0x20'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-1",role="replica",cluster="ha-pg"}'
+        values: '1x30'
+    alert_rule_test:
+      - eval_time: 15m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Recorder database ha-pg has no primary or is missing an instance"
+              description: "In CNPG cluster home-assistant/ha-pg either no instance reports role=primary with Postgres up, or fewer than 2 instances have Postgres up. Without a primary the Home Assistant recorder cannot write history. Check `kubectl -n home-assistant get cluster ha-pg` and the pods' logs."
+
+  # Both instances up but neither is primary (failover stuck): fires.
+  - interval: 1m
+    input_series:
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-2",role="replica",cluster="ha-pg"}'
+        values: '1x30'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-1",role="replica",cluster="ha-pg"}'
+        values: '1x30'
+    alert_rule_test:
+      - eval_time: 4m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts: []
+      - eval_time: 5m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Recorder database ha-pg has no primary or is missing an instance"
+              description: "In CNPG cluster home-assistant/ha-pg either no instance reports role=primary with Postgres up, or fewer than 2 instances have Postgres up. Without a primary the Home Assistant recorder cannot write history. Check `kubectl -n home-assistant get cluster ha-pg` and the pods' logs."
+
+  # Clean switchover at t=10m: ha-pg-1 takes primary, ha-pg-2 becomes replica.
+  # The role label flips, so both old series go stale and new ones start. Must
+  # stay silent throughout.
+  - interval: 1m
+    input_series:
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-2",role="primary",cluster="ha-pg"}'
+        values: '1x9'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-1",role="replica",cluster="ha-pg"}'
+        values: '1x9'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-1",role="primary",cluster="ha-pg"}'
+        values: '_x10 1x20'
+      - series: 'cnpg_collector_up{job="home-assistant/cnpg-ha-pg",namespace="home-assistant",pod="ha-pg-2",role="replica",cluster="ha-pg"}'
+        values: '_x10 1x20'
+    alert_rule_test:
+      - eval_time: 12m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts: []
+      - eval_time: 30m
+        alertname: HomeAssistantRecorderDBUnhealthy
+        exp_alerts: []
+
+  # --- MosquittoDown ------------------------------------------------------
+  # The broker goes not-Ready at t=10m while a credentials-bootstrap Job pod is
+  # Ready. The Job pod must not mask the broker.
+  - interval: 1m
+    input_series:
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="mosquitto-7f8b9c6d5-x2k9q",condition="true"}'
+        values: '1x9 0x20'
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="mosquitto-credentials-bootstrap-q7w2e",condition="true"}'
+        values: '1x30'
+    alert_rule_test:
+      - eval_time: 14m
+        alertname: MosquittoDown
+        exp_alerts: []
+      - eval_time: 15m
+        alertname: MosquittoDown
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Mosquitto (MQTT broker) has not been ready for 5 minutes"
+              description: "No mosquitto pod in namespace home-assistant is Ready. Zigbee2MQTT, Shelly and every other MQTT device lose their link to Home Assistant while this lasts, so Zigbee2MQTTDownOrCoordinatorOffline may fire alongside it for the same cause."
+
+  # --- Zigbee2MQTTDownOrCoordinatorOffline -------------------------------
+  # Pod Ready throughout; HA's bridge sensor reads 0 from t=10m (coordinator
+  # unplugged).
+  - interval: 1m
+    input_series:
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="zigbee2mqtt-5c4d8f7b6-m3n4p",condition="true"}'
+        values: '1x30'
+      - series: 'homeassistant_binary_sensor_state{entity="binary_sensor.zigbee2mqtt_bridge_connection_state",domain="binary_sensor",friendly_name="Zigbee2MQTT Bridge Connection state",job="home-assistant",namespace="home-assistant",instance="10.0.0.1:8124"}'
+        values: '1x9 0x20'
+    alert_rule_test:
+      - eval_time: 14m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts: []
+      - eval_time: 15m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Zigbee2MQTT is down or its coordinator/bridge is offline"
+              description: "Either the zigbee2mqtt pod in namespace home-assistant is not Ready, or Home Assistant's binary_sensor.zigbee2mqtt_bridge_connection_state reads off. If MosquittoDown is firing too, fix the broker first; the bridge sensor goes off whenever MQTT is gone."
+
+  # Pod not Ready from t=10m with the bridge sensor at 0; a replacement pod is
+  # Ready from t=18m but the bridge stays offline. One outage that changes
+  # shape: fires at t=15m as one alert, and is still that alert once only the
+  # sensor branch is true.
+  - interval: 1m
+    input_series:
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="zigbee2mqtt-5c4d8f7b6-m3n4p",condition="true"}'
+        values: '1x9 0x20'
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="zigbee2mqtt-5c4d8f7b6-r5s6t",condition="true"}'
+        values: '_x17 1x12'
+      - series: 'homeassistant_binary_sensor_state{entity="binary_sensor.zigbee2mqtt_bridge_connection_state",domain="binary_sensor",friendly_name="Zigbee2MQTT Bridge Connection state",job="home-assistant",namespace="home-assistant",instance="10.0.0.1:8124"}'
+        values: '1x9 0x20'
+    alert_rule_test:
+      - eval_time: 14m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts: []
+      - eval_time: 15m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Zigbee2MQTT is down or its coordinator/bridge is offline"
+              description: "Either the zigbee2mqtt pod in namespace home-assistant is not Ready, or Home Assistant's binary_sensor.zigbee2mqtt_bridge_connection_state reads off. If MosquittoDown is firing too, fix the broker first; the bridge sensor goes off whenever MQTT is gone."
+      - eval_time: 20m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts:
+          - exp_labels: {severity: critical, namespace: home-assistant}
+            exp_annotations:
+              summary: "Zigbee2MQTT is down or its coordinator/bridge is offline"
+              description: "Either the zigbee2mqtt pod in namespace home-assistant is not Ready, or Home Assistant's binary_sensor.zigbee2mqtt_bridge_connection_state reads off. If MosquittoDown is firing too, fix the broker first; the bridge sensor goes off whenever MQTT is gone."
+
+  # HA down: the bridge sensor goes stale and disappears, the Z2M pod is fine.
+  # HomeAssistantDown owns this; the Zigbee rule must stay silent.
+  - interval: 1m
+    input_series:
+      - series: 'kube_pod_status_ready{namespace="home-assistant",pod="zigbee2mqtt-5c4d8f7b6-m3n4p",condition="true"}'
+        values: '1x30'
+      - series: 'homeassistant_binary_sensor_state{entity="binary_sensor.zigbee2mqtt_bridge_connection_state",domain="binary_sensor",friendly_name="Zigbee2MQTT Bridge Connection state",job="home-assistant",namespace="home-assistant",instance="10.0.0.1:8124"}'
+        values: '1x9'
+    alert_rule_test:
+      - eval_time: 30m
+        alertname: Zigbee2MQTTDownOrCoordinatorOffline
+        exp_alerts: []

kubernetes/apps/monitoring-rules/tests/mutations.txt

@@ -0,0 +1,143 @@
+# Negative controls for the promtool suites in this directory.
+#
+# A unit-test suite that passes proves the rules behave as asserted. It does not
+# prove the assertions are worth anything: a suite whose `rule_files` path does
+# not resolve, or whose traces are all zero, passes just as green. Both of those
+# happened while this suite was being written, and neither was visible from the
+# exit code.
+#
+# So scripts/ci/promtool-rules.sh runs a second phase. For each mutation below
+# it damages ONE value the rules depend on, re-runs every suite, and requires the
+# run to FAIL. A mutation the suite does not notice is a hole in the suite, and
+# CI reports it as such.
+#
+# FORMAT — three `|`-separated fields, no `|` inside any of them:
+#
+#   name | manifest path relative to the repo root | sed program
+#
+# The sed program is applied to the `spec:`-stripped copy in the mirrored tree,
+# so its indentation is two spaces shallower than the manifest on disk. Anchor
+# on `^[[:space:]]*` rather than a fixed indent. `#` in the first column is a
+# comment; blank lines are ignored.
+#
+# The script also requires each sed program to CHANGE the file. A mutation that
+# matches nothing is reported as a failure, not skipped — otherwise renaming a
+# rule or rewording an expression silently turns its control into a no-op and
+# the suite goes back to vouching for itself.
+#
+# `detected by` notes below are for the reader. The machine check is only "the
+# suite must fail"; nothing parses them.
+
+# The margin control, and the only reason case 9 is in the suite. The worst
+# night in Prometheus' retention (2026-09-11, peak 141) holds the expression
+# true for 12 minutes at a 1m evaluation step, so at `for: 12m` it fires. The
+# seven reconstructed nights peak at 10m pending and would stay silent — without
+# the raw worst-night case, `for: 20m` could be cut to 12m with the suite green.
+# Detected by: case 9, the worst night on record.
+backlog-for-12m|kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml|s/^\([[:space:]]*\)for: 20m$/\1for: 12m/
+
+# The value REN-129 changed, in the direction it changed it from. At 30m the
+# rule does not page until 35 minutes into a stall that started at t=5m, so the
+# assertion at t=25m sees nothing. This is the control that stops the 20m being
+# quietly walked back.
+# Detected by: cases 10, 11, 13, 14 — every case that expects a page.
+backlog-for-30m|kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml|s/^\([[:space:]]*\)for: 20m$/\1for: 30m/
+
+# Drop the aggregation, keeping the selector. `role` flips on failover; a
+# changed label is a new series to Prometheus and restarts the `for` timer, so
+# a backlog that persists across a promotion pages 20m after the promotion
+# instead of 20m after the backlog. Scoped to `expr:` lines because
+# `max by (namespace, pod)` also appears in this manifest's comments, where
+# rewriting it would mutate the documentation and nothing else.
+# Detected by: case 13, the role flip.
+no-max-by|kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml|/^[[:space:]]*expr:/s/max by (namespace, pod) (\(cnpg_collector_pg_wal_archive_status{value="ready"}\))/\1/
+
+# Drop the `value="ready"` selector. The same metric name also carries
+# `value="done"`, a monotonic counter in the hundreds of thousands, so the
+# rule would fire permanently on every CNPG instance in the cluster. The
+# measured-night cases carry that counter series in their input specifically so
+# this is detectable rather than a silent no-op.
+# Detected by: cases 1-9 and 12 — every case with a `done` series.
+no-value-selector|kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml|/^[[:space:]]*expr:/s/{value="ready"}//
+
+# Loosen the comparison. `> 5` is strict; `>= 5` pages
+# on a backlog sitting at exactly five segments, which is 80 MiB and normal.
+# Detected by: case 12, pinned at exactly 5.
+backlog-ge-5|kubernetes/apps/monitoring-rules/prometheusrule-cnpg-backup.yaml|/^[[:space:]]*expr:/s/> 5$/>= 5/
+
+# --- prometheusrule-external-secrets.yaml (REN-103) -------------------------
+# Scoped to the ExternalSecretSyncFailing expression with /max by/ or /== 1/
+# where it matters: the manifest's comments quote the rejected forms, and
+# rewriting those would mutate the documentation and nothing else.
+
+# The window. At 5m the failure that starts at t=10m is already firing at the
+# t=15m assertion, which expects pending.
+# Detected by: the eval_time 15m case.
+eso-for-5m|kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml|s/^\([[:space:]]*\)for: 15m$/\1for: 5m/
+
+# REN-100's aggregation. `sum` over the two-pod rollout case is 2, fails
+# `== 1`, and loki-basic-auth never fires.
+# Detected by: the eval_time 30m case, loki-basic-auth.
+eso-sum-not-max|kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml|s/^\([[:space:]]*\)max by (exported_namespace, name) ($/\1sum by (exported_namespace, name) (/
+
+# The label Argus found shadowed on #994. Grouping by `namespace` names the
+# controller's namespace on every alert.
+# Detected by: the eval_time 30m case, exp_labels and the rendered summary.
+eso-by-namespace|kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml|s/^\([[:space:]]*\)max by (exported_namespace, name) ($/\1max by (namespace, name) (/
+
+# Drop `== 1`. ESO writes the inactive Ready polarity at 0, so every healthy
+# ExternalSecret matches the bare selector and fires.
+# Detected by: the eval_time 30m case, uptime-robot-api-key.
+eso-no-eq-1|kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml|s/^\([[:space:]]*\)) == 1$/\1)/
+
+# Drop the condition selector. The Deleted tombstone is status="False" at 1
+# forever, so it pages until the controller restarts.
+# Detected by: the eval_time 30m case, wiki-index-basic-auth.
+eso-any-condition|kubernetes/apps/monitoring-rules/prometheusrule-external-secrets.yaml|s/^\([[:space:]]*\)externalsecret_status_condition{condition="Ready",status="False"}$/\1externalsecret_status_condition{status="False"}/
+
+# prometheusrule-home-assistant.yaml (REN-245). Detected by the matching
+# `eval_time: 14m` / `15m` pairs in home-assistant_test.yaml.
+ha-down-for-10m|kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml|/alert: HomeAssistantDown/,/for:/s/for: 5m/for: 10m/
+# Drops the label normalisation: the alert changes identity when the target
+# comes back failing, and the t=25m assertion loses it.
+ha-down-no-max|kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml|s/max by (job, namespace) (up{job="home-assistant", namespace="home-assistant"}) == 0/up{job="home-assistant", namespace="home-assistant"} == 0/
+# The instance floor pinned to spec.instances. At 3 the healthy two-instance
+# cluster pages. Detected by the switchover case, which must stay silent.
+ha-recorder-count-3|kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml|s/== 1) < 2$/== 1) < 3/
+# The loose regex that lets a Ready bootstrap Job pod mask a down broker.
+ha-mosquitto-loose-regex|kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml|s/pod=~"mosquitto-\[a-z0-9\]{6,10}-\[a-z0-9\]{5}"/pod=~"mosquitto-.*"/
+# Drops the bridge-sensor branch. Detected by the coordinator-offline case.
+ha-z2m-no-bridge|kubernetes/apps/monitoring-rules/prometheusrule-home-assistant.yaml|s/bridge_connection_state"}) == 0$/bridge_connection_state"}) == 2/
+
+# --- prometheusrule-ceph.yaml (REN-93) --------------------------------------
+# The canary and the critical rule it guards share the `{type="replicated"}`
+# join text, so these scope by the canary's `absent(` ... `)` block: inside it
+# for the canary, `!` for everything else. The annotation quotes the selector
+# with escaped quotes and so never matches.
+
+# The window, shortened. At 15m the canary pages on a mgr-failover-length gap.
+# Detected by: case F, the eval_time 75m silence.
+ceph-blind-for-15m|kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml|s/^\([[:space:]]*\)for: 1h$/\1for: 15m/
+
+# The window, lengthened. At 2h a real blindness that started at t=30m is
+# still not reported at 95m.
+# Detected by: case F, the eval_time 95m page.
+ceph-blind-for-2h|kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml|s/^\([[:space:]]*\)for: 1h$/\1for: 2h/
+
+# Drop the value filter from the canary only. It then watches a join the
+# critical rule does not use, and stays quiet through a renamed `type` label
+# or value while the critical rule is dark.
+# Detected by: cases C and D.
+ceph-blind-any-type|kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml|/^[[:space:]]*absent($/,/^[[:space:]]*)$/s/{type="replicated"}//
+
+# Invert the canary: the bare join instead of its absence. It then fires on
+# every healthy cluster and on every real redundancy loss, i.e. it masks the
+# alert it exists to protect.
+# Detected by: cases A and B.
+ceph-blind-inverted|kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml|s/^\([[:space:]]*\)absent($/\1(/
+
+# REN-88's gate, removed from the critical rule. Erasure-coded and renamed
+# pools then match the critical rule again. Proves this suite loads and asserts
+# the critical rule, not only the canary.
+# Detected by: cases C, D and E, which assert the critical rule is silent.
+ceph-critical-any-type|kubernetes/apps/monitoring-rules/prometheusrule-ceph.yaml|/^[[:space:]]*absent($/,/^[[:space:]]*)$/!s/{type="replicated"}//

kubernetes/apps/monitoring/helmrelease-blackbox-exporter.yaml

@@ -2,6 +2,7 @@ apiVersion: helm.toolkit.fluxcd.io/v2
 kind: HelmRelease
 metadata:
   name: blackbox-exporter
+  # anchor: namespace
   namespace: monitoring
 spec:
   interval: 30m

kubernetes/apps/monitoring/helmrelease-prometheus.yaml

@@ -17,7 +17,7 @@ spec:
   chart:
     spec:
       chart: kube-prometheus-stack
-      version: ">=72.0.0"
+      version: "91.7.0"
       sourceRef:
         kind: HelmRepository
         name: prometheus-community
@@ -91,12 +91,11 @@ spec:
         # Additional scrape targets not covered by ServiceMonitors.
         additionalScrapeConfigs:
           # Gitea exposes /metrics anonymously on its main HTTP port. The
-          # `gitea_*` metrics power the gitea-overview dashboard. Gitea now runs
-          # in-cluster (StatefulSet in apps ns; LXC 114 retired), so scrape the
-          # ClusterIP Service directly instead of the dead LXC hostname.
+          # `gitea_*` metrics power the gitea-overview dashboard. Scrapes the
+          # ClusterIP Service of the StatefulSet in the gitea namespace.
           - job_name: gitea
             static_configs:
-              - targets: ["gitea.apps.svc.cluster.local:3000"]
+              - targets: ["gitea.gitea.svc.cluster.local:3000"]
             metrics_path: /metrics
 
     # Grafana
@@ -305,6 +304,10 @@ spec:
             - matchers:
                 - alertname = "KubeCPUOvercommit"
               receiver: 'null'
+            # Helper alert that only drives the info inhibit rule below.
+            - matchers:
+                - alertname = "InfoInhibitor"
+              receiver: 'null'
             # Flux notification-controller posts info-severity events
             # here (e.g. image-automation commits). Keep them visible
             # but don't repeat-spam every 12h on the same event.
@@ -314,6 +317,23 @@ spec:
               group_wait: 1m
               group_interval: 10m
               repeat_interval: 24h
+        # From kube-prometheus manifests/alertmanager-secret.yaml.
+        inhibit_rules:
+          - source_matchers:
+              - severity = "critical"
+            target_matchers:
+              - severity =~ "warning|info"
+            equal: ['namespace', 'alertname']
+          - source_matchers:
+              - severity = "warning"
+            target_matchers:
+              - severity = "info"
+            equal: ['namespace', 'alertname']
+          - source_matchers:
+              - alertname = "InfoInhibitor"
+            target_matchers:
+              - severity = "info"
+            equal: ['namespace']
         receivers:
           - name: email
             email_configs:
@@ -340,6 +360,285 @@ spec:
         limits:
           memory: 256Mi
 
+      # --- Flux per-object readiness (REN-102) ---------------------------------
+      #
+      # WHY THIS IS HERE AND NOT ON THE FLUX CONTROLLERS. `gotk_reconcile_condition`
+      # and `gotk_suspend_status` are still declared by fluxcd/pkg/runtime, but no
+      # controller populates them at the tags pinned in
+      # clusters/homelab/flux-system/gotk-components.yaml, and an unpopulated
+      # GaugeVec exports no series. Verified after the REN-97 scrape landed: the
+      # PodMonitor now yields 2,273 `gotk_*` series and
+      # `count({__name__=~"gotk_reconcile_condition|gotk_suspend_status"})` is still
+      # empty. Upstream moved per-object readiness to kube-state-metrics
+      # custom-resource-state, which is what this block turns on.
+      #
+      # DO NOT ADD `collectors: []` OR `--custom-resource-state-only=true`.
+      # Upstream's reference config (fluxcd/flux2-monitoring-example,
+      # monitoring/controllers/kube-prometheus-stack/kube-state-metrics-config.yaml)
+      # opens with exactly those two lines. Copying them replaces every built-in
+      # collector with the Flux ones: measured 2026-09-26, that would take
+      # `count({job="kube-state-metrics"})` from 27,792 series to ~183 and blind
+      # every `Kube*` alert in the chart. Both keys are deliberately absent below.
+      #
+      # ROLLS THE kube-state-metrics POD. Enabling customResourceState adds an arg
+      # and a volume to the Deployment, so the pod is replaced once on apply. That
+      # is the whole blast radius of this block — KSM only watches, it writes
+      # nothing, and no Flux object or workload is touched. During the roll (a few
+      # seconds) `kube_*` has a gap; `KubeStateMetricsDown` and `TargetDown` cover
+      # it if it does not come back.
+      #
+      # NOTE FOR LATER EDITS: the subchart puts no config checksum on the
+      # Deployment's pod template, so editing `config` below changes the ConfigMap
+      # WITHOUT rolling the pod, and KSM keeps serving the old config until it
+      # restarts for another reason. Any future change to this config needs a
+      # deliberate pod restart to take effect.
+      #
+      # RBAC: list + watch, nothing else. `get` is NOT needed — KSM drives
+      # reflectors, never point-reads. Also note the chart adds
+      # `customresourcedefinitions: [list, watch]` on its own via
+      # rbac.customResourceState.createClusterRoleRules (default true); KSM needs it
+      # to resolve the GVKs below. Both land in the existing
+      # kube-prometheus-stack-kube-state-metrics ClusterRole rather than a second
+      # binding to maintain. No write verb, no wildcard, no cluster-admin.
+      rbac:
+        extraRules:
+          - apiGroups:
+              - source.toolkit.fluxcd.io
+              - kustomize.toolkit.fluxcd.io
+              - helm.toolkit.fluxcd.io
+              - image.toolkit.fluxcd.io
+            # Trimmed to the kinds this cluster actually has. Omitted from
+            # upstream's list, each for a checked reason:
+            #   buckets, receivers        — zero objects exist in this repo.
+            #   helmcharts                — created by source-controller at
+            #     runtime, one per HelmRelease; a not-Ready HelmChart is always
+            #     mirrored by the HelmRelease that owns it, so alerting on both
+            #     is one failure reported twice.
+            #   alerts, providers         — notification.toolkit.fluxcd.io/v1beta3
+            #     Alert and Provider have NO status.conditions in the CRD at all
+            #     (checked in gotk-components.yaml), so there is no readiness to
+            #     report; upstream omits `ready` for them for the same reason. A
+            #     series carrying only `suspended` is not alertable.
+            resources:
+              - gitrepositories
+              - helmrepositories
+              - ocirepositories
+              - kustomizations
+              - helmreleases
+              - imagerepositories
+              - imagepolicies
+              - imageupdateautomations
+            verbs: ["list", "watch"]
+
+      customResourceState:
+        enabled: true
+        config:
+          spec:
+            resources:
+              # Label set is upstream's, minus the churning ones. Dropped
+              # deliberately: `revision` (status.artifact.revision /
+              # lastAppliedRevision — a git SHA, so every commit that reaches a
+              # Kustomization mints a NEW SERIES), and `url` / `endpoint`.
+              # Prometheus is at 84.5% of its 30GB retentionSize cap (REN-107);
+              # a label that changes several times a day across 33 Kustomizations
+              # is exactly what must not be added. Kept: `ready` and `suspended`
+              # (what the alerts read), `name` / `exported_namespace` (identity),
+              # `source_name` (the join key, see prometheusrule-flux.yaml), plus
+              # `image` and `tag` on the two image kinds, which are stable enough
+              # to ride the existing series.
+              #
+              # `exported_namespace` is upstream's own label name for this, not a
+              # Prometheus collision rename: the chart's KSM ServiceMonitor runs
+              # honorLabels, so `namespace` on a KSM series is already the
+              # OBJECT's namespace and nothing gets renamed. The rules use the
+              # same string because the config declares it.
+              #
+              # KSM adds `customresource_group`, `customresource_version` and
+              # `customresource_kind` to every custom-resource-state series by
+              # itself. The rules key off `customresource_kind`.
+              - groupVersionKind:
+                  group: kustomize.toolkit.fluxcd.io
+                  version: v1
+                  kind: Kustomization
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux Kustomization resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+                      source_name: [spec, sourceRef, name]
+              - groupVersionKind:
+                  group: helm.toolkit.fluxcd.io
+                  version: v2
+                  kind: HelmRelease
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux HelmRelease resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+              - groupVersionKind:
+                  group: source.toolkit.fluxcd.io
+                  version: v1
+                  kind: GitRepository
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux GitRepository resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+              - groupVersionKind:
+                  group: source.toolkit.fluxcd.io
+                  version: v1
+                  kind: HelmRepository
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux HelmRepository resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+              - groupVersionKind:
+                  group: source.toolkit.fluxcd.io
+                  version: v1
+                  kind: OCIRepository
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux OCIRepository resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+              - groupVersionKind:
+                  group: image.toolkit.fluxcd.io
+                  version: v1
+                  kind: ImageRepository
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux ImageRepository resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+                      image: [spec, image]
+                  # The repository-side half of the Ready-but-frozen signal. A
+                  # policy can hold Ready=True forever while its filter excludes
+                  # every new release, so readiness cannot see it; what can is
+                  # "the repository kept finding tags and the policy's answer
+                  # never moved". .status.lastScanResult.tagCount is a required
+                  # field on this CRD, so the series exists for every object.
+                  - name: "image_repository_tag_count"
+                    help: "Number of tags found by the ImageRepository's last scan."
+                    each:
+                      type: Gauge
+                      gauge:
+                        path: [status, lastScanResult]
+                        valueFrom: [tagCount]
+                    # `name` belongs HERE and not under `gauge`. KSM resolves
+                    # `gauge.labelsFromPath` RELATIVE TO `gauge.path`, so
+                    # [metadata, name] under there would be looked up inside
+                    # status.lastScanResult and silently yield no label. The
+                    # metric-level labelsFromPath resolves from the object root.
+                    labelsFromPath:
+                      name: [metadata, name]
+                      exported_namespace: [metadata, namespace]
+              - groupVersionKind:
+                  group: image.toolkit.fluxcd.io
+                  version: v1
+                  kind: ImagePolicy
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux ImagePolicy resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+                      # The resolved tag, as a label on the Info series that
+                      # already exists — so the policy-side half of the frozen
+                      # signal costs zero extra series at rest. It retires one
+                      # series and starts another only when the tag actually
+                      # changes: 191 automated bumps in 147 days across 34
+                      # policies is ~1.3 new series/day, which ages out with
+                      # retention.
+                      tag: [status, latestRef, tag]
+                      # The join key. `metadata.name` happens to equal
+                      # `spec.imageRepositoryRef.name` for all 34 pairs in this
+                      # repo today, but the REFERENCE is this field, and a policy
+                      # named differently from its repository would silently drop
+                      # out of a name-based join.
+                      source_name: [spec, imageRepositoryRef, name]
+              - groupVersionKind:
+                  group: image.toolkit.fluxcd.io
+                  version: v1
+                  kind: ImageUpdateAutomation
+                metricNamePrefix: gotk
+                errorLogV: 10
+                metrics:
+                  - name: "resource_info"
+                    help: "The current state of a Flux ImageUpdateAutomation resource."
+                    each:
+                      type: Info
+                      info:
+                        labelsFromPath:
+                          name: [metadata, name]
+                    labelsFromPath:
+                      exported_namespace: [metadata, namespace]
+                      ready: [status, conditions, "[type=Ready]", status]
+                      suspended: [spec, suspend]
+                      source_name: [spec, sourceRef, name]
+
     # Drop the apiserver's high-cardinality latency histograms.
     #
     # Measured 2026-08-20: Prometheus held 298,236 active series, of which

kubernetes/apps/navidrome/deployment.yaml

@@ -0,0 +1,130 @@
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: navidrome-config
+  namespace: navidrome
+data:
+  TZ: "Europe/Berlin"
+  ND_LOGLEVEL: "info"
+  ND_SCANNER_SCHEDULE: "0 * * * *"
+  ND_ENABLEINSIGHTSCOLLECTOR: "false"
+  # Identity from the navidrome-gate outpost. Traefik strips this header on
+  # the ungated /rest and /share routes.
+  #
+  # ND_EXTAUTH_TRUSTEDSOURCES BELOW IS NOT A NARROWING. fd10:244::/48 plus
+  # 10.244.0.0/16 is the entire cluster pod CIDR, so navidrome's own
+  # reverse-proxy trust check accepts X-Authentik-Username from ANY pod.
+  # The CiliumNetworkPolicy is the only real boundary, and it is wider than
+  # "the outpost and Traefik" — default-deny-ingress in
+  # kubernetes/infrastructure/network-policies/navidrome-ns.yaml uses
+  # `endpointSelector: {}` and admits, on every port:
+  #   - the navidrome namespace, every port
+  #   - traefik, every port
+  #   - the navidrome-gate outpost, every port
+  #   - the lidarr-gate outpost, every port - reaches navidrome too
+  #   - host / remote-node, every port - the kubelet sources the HTTP probes
+  #     from the host netns, so this one can never be removed while the pod
+  #     has HTTP probes
+  #   - blackbox-exporter on :4533 and :8686 only
+  # The whole-monitoring-namespace admit that used to be in this list was
+  # narrowed to blackbox-exporter on those two ports in REN-28. Rules are
+  # cited by what they admit, not by line number, so this list cannot go
+  # stale the next time that file is edited.
+  # The value stays at the pod CIDR ON PURPOSE - see REN-28. The address
+  # navidrome sees is the outpost's POD IP, which is not stable across
+  # reschedules, so a narrower literal here fails open-ended and silently
+  # the next time that pod moves. Tightening the CNP is the control that
+  # would actually do work.
+  ND_EXTAUTH_USERHEADER: "X-Authentik-Username"
+  ND_EXTAUTH_TRUSTEDSOURCES: "fd10:244::/48,10.244.0.0/16"
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: navidrome
+  namespace: navidrome
+  labels:
+    app.kubernetes.io/name: navidrome
+  annotations:
+    configmap.reloader.stakater.com/reload: navidrome-config
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: navidrome
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: navidrome
+    spec:
+      enableServiceLinks: false
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        runAsGroup: 1000
+        fsGroup: 1000
+        fsGroupChangePolicy: OnRootMismatch
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: navidrome
+          image: deluan/navidrome:0.64.2 # {"$imagepolicy": "flux-system:navidrome"}
+          ports:
+            - name: http
+              containerPort: 4533
+          envFrom:
+            - configMapRef:
+                name: navidrome-config
+          volumeMounts:
+            - name: data
+              mountPath: /data
+            - name: music
+              mountPath: /music
+              readOnly: true
+          # anchor: kubelet-probes
+          livenessProbe:
+            httpGet:
+              path: /ping
+              port: http
+            initialDelaySeconds: 30
+            periodSeconds: 30
+            timeoutSeconds: 10
+          readinessProbe:
+            httpGet:
+              path: /ping
+              port: http
+            initialDelaySeconds: 10
+            periodSeconds: 10
+            timeoutSeconds: 5
+          resources:
+            requests:
+              cpu: 100m
+              memory: 256Mi
+            limits:
+              memory: 1Gi
+          securityContext:
+            allowPrivilegeEscalation: false
+            capabilities:
+              drop: ["ALL"]
+      volumes:
+        - name: data
+          persistentVolumeClaim:
+            claimName: navidrome-data
+        - name: music
+          persistentVolumeClaim:
+            claimName: navidrome-music-bulk
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: navidrome
+  namespace: navidrome
+spec:
+  selector:
+    app.kubernetes.io/name: navidrome
+  ports:
+    - name: http
+      port: 80
+      targetPort: http

kubernetes/apps/navidrome/gate-service.yaml

@@ -0,0 +1,13 @@
+# ExternalName mirror so the navidrome-namespace Ingress can backend to the
+# navidrome-gate outpost in the authentik namespace.
+apiVersion: v1
+kind: Service
+metadata:
+  name: ak-outpost-navidrome-gate
+  namespace: navidrome
+spec:
+  type: ExternalName
+  externalName: ak-outpost-navidrome-gate.authentik.svc.cluster.local
+  ports:
+    - port: 9000
+      targetPort: 9000