feat(stalwart): spread each node group over nodes and give each one a PDB

A two-replica node group was redundancy on paper: nothing kept both pods off
one node, and no PodDisruptionBudget kept a drain from taking both at once.
Both objects are now rendered per node group, gated on that group's COMPUTED
replica count being >= 2.

The distribution is a promise, not an effort: maxSkew 1 on
kubernetes.io/hostname with whenUnsatisfiable DoNotSchedule, selecting only
this group's own pods, so an "mx" and a "full" pod may share a node. The
selector comes from the chart (stalwart.groupLabels), never from the operator -
a spread constraint whose selector matches no pod is trivially satisfied,
distributes nothing and reports nothing.

nodeAffinityPolicy: Ignore is set explicitly, and that is the load-bearing
detail. Its API default is Honor, under which a narrowing nodeSelector shrinks
the set of counted domains until both replicas fit into one: skew 0, both pods
co-located, no Pending, no event, and a selector that is correct - so no check
can see it. With Ignore the same narrowing gives skew 2 and the second pod is
Pending. A render abort was the alternative and would also have forbidden a
nodeSelector spanning five nodes, which is legal.

A user entry colliding with the chart's own on (topologyKey,
whenUnsatisfiable) aborts the render naming both values, because the API
rejects the duplicate with a message about a list index and an apply-based path
merges the two over patchMergeKey: topologyKey silently, dropping one selector.

Minor rather than major, proved with a diff rather than argued: with the
shipped defaults (full: 1, mx: 0) and with ci/stalwart/test-values.yaml,
helm template renders byte-for-byte what 1.0.1+upv0.16.24 rendered - both
diffs empty.

New in CI, over every chart on every PR: ci/run-selector-match-test.sh holds
every rendered PodDisruptionBudget selector and every topologySpreadConstraints
labelSelector against the labels of the pod templates in the same render. Helm
itself is the YAML parser (ci/yaml2json), because the runners have neither
pyyaml nor yq and a hand-written parser could read nothing and report nothing.
podAffinity/podAntiAffinity are out of scope; matchExpressions-only selectors
are reported as not checked rather than counted as passing. Two anti-vacuum
guards: a chart with an overlay under ci/selector-match/ that renders no
selector fails, and a run that checks no selector at all fails.

Fixes #96

README.md

@@ -68,8 +68,8 @@ Every chart has a wiki page of its own — `https://git.jk-dev.de/jlab-cluster/h
 
   Note: the registry **accepts a re-upload of an already published version** (HTTP 201) and replaces the stored artifact, so a republish of an existing version is still silent — see #160. That is also what makes a manual `workflow_dispatch` of any `publish-<chart>.yml` a harmless way to prove the shared block works.
 - `change-app-version.yml` — manually dispatched app-version bump; the chart must be listed in its `chart-name` choice list.
-- `ci/helm-dependencies.sh <chart-dir>` — not a workflow but the step they all share: fetches a chart's declared dependencies into `charts/`, and returns silently for the charts that declare none (all but three: `ytdlp-web-player`, `stalwart`, `paperless-ngx`). It runs in **five** places, because every helm command this repo builds on exits 1 while a declared dependency is missing: before lint and package in `validate-charts.yml`, before package in the shared publish block (which is why that block names this script as a prerequisite of its callers — see above), and inside `ci/run-install-test.sh`, `ci/run-quantity-render-test.sh` and `ci/run-trivy-config.sh`. See "Chart dependencies" in the style guidelines for what that costs.
-- `validate-charts.yml` — PR pipeline: detects which charts a PR touches and, for exactly those, runs `helm lint --strict`, `helm package`, and a chart-version bump check (`template-chart` is exempt from the version check). PRs without chart changes skip the build matrix. After the build stage, an **install-test** job (the same changed-charts matrix, minus the charts listed in `ci/excluded-charts.txt`) installs each changed chart into an ephemeral [k3d](https://k3d.io) cluster on the runner and requires every workload to become Ready — see below. Both matrices follow the changed chart directories, so a PR that only changes the harness under `ci/` exercises nothing by itself; the Trivy scan has a fixed-sample fallback for that case, the install-test has none — bump the affected chart's patch version to pull it into the matrix, or run the all-charts install-test after the merge. The **resource-quantity render test** is the exception to that rule: it runs over every chart on every PR — see below.
+- `ci/helm-dependencies.sh <chart-dir>` — not a workflow but the step they all share: fetches a chart's declared dependencies into `charts/`, and returns silently for the charts that declare none (all but three: `ytdlp-web-player`, `stalwart`, `paperless-ngx`). It runs in **six** places, because every helm command this repo builds on exits 1 while a declared dependency is missing: before lint and package in `validate-charts.yml`, before package in the shared publish block (which is why that block names this script as a prerequisite of its callers — see above), and inside `ci/run-install-test.sh`, `ci/run-quantity-render-test.sh`, `ci/run-selector-match-test.sh` and `ci/run-trivy-config.sh`. See "Chart dependencies" in the style guidelines for what that costs.
+- `validate-charts.yml` — PR pipeline: detects which charts a PR touches and, for exactly those, runs `helm lint --strict`, `helm package`, and a chart-version bump check (`template-chart` is exempt from the version check). PRs without chart changes skip the build matrix. After the build stage, an **install-test** job (the same changed-charts matrix, minus the charts listed in `ci/excluded-charts.txt`) installs each changed chart into an ephemeral [k3d](https://k3d.io) cluster on the runner and requires every workload to become Ready — see below. Both matrices follow the changed chart directories, so a PR that only changes the harness under `ci/` exercises nothing by itself; the Trivy scan has a fixed-sample fallback for that case, the install-test has none — bump the affected chart's patch version to pull it into the matrix, or run the all-charts install-test after the merge. The **resource-quantity render test** and the **selector-match render test** are the two exceptions to that rule: both run over every chart on every PR — see below.
 - `install-test.yml` — reusable workflow (`workflow_call`) holding the install-test **job** itself: the runner-label expression, the pinned `kubectl` with its skew check, `envsubst`, the containerised-job detection, the per-leg k3d cluster and the `freesailarr` VPN credentials. Its one input is a JSON array of chart names; its callers decide *which* charts are tested and whether the job runs at all, never *how*. Two callers: `validate-charts.yml` and `install-test-all-charts.yml`. **It is not copied for the second caller** — a second copy is a second chance to test differently, unnoticed, the same argument that gives `ci/excluded-charts.txt` exactly one parser.
 
   **It stays in this repository, and that is a decision, not an oversight** (Gitea #66, decision 4). It is reusable in form but not in substance: it is welded to the `ci/` fixtures, to the `ci/excluded-charts.txt` parser and to this repository's chart matrix. A building block that only one repository can use is not shared, it is merely moved somewhere else — and moving it would buy nothing while putting a tag between us and a file we change often. `publish-helm-chart.yml` was the opposite case: it takes a chart name and a registry owner and knows nothing about this repository, which is exactly why it could go. This note exists so the question is not asked again in half a year.
@@ -142,6 +142,25 @@ Two details worth knowing before touching it:
 - **The resources paths come from `values.schema.json`, not `values.yaml`** — it is JSON, so `python3` reads it without pyyaml (same runner constraint as `ci/run-trivy-config.sh`), and every chart-owned key must appear there anyway. A block counts when it is `#/definitions/resources` or carries `limitFactor`/`requests` itself, which is what keeps `minecraft-server`'s JVM-sizing `minecraft.resources` out. A chart whose templates render a resources block while its schema declares none **fails** the test with exactly that message: the schema is then out of sync with the chart.
 - **A chart that never reaches the helper is reported, not counted as passing.** `gotenberg` sets every computable limit explicitly in its fallback defaults, and since a `null` never deletes a default there, no override can free them — nothing is multiplied, so nothing is proven. Those charts appear under "helper not reached" in the job summary rather than as a green tick for a code path that was never entered.
 
+### Selector-match render test (all charts, every PR)
+
+`ci/run-selector-match-test.sh [chart …]` (no argument = all charts) renders every chart with the same fixtures the install-test uses and checks that **every label selector the render contains really matches the labels of a pod template in that same render**. Two kinds of object are covered, and both of them fail *silently* when their selector matches nothing:
+
+| Object | What a non-matching selector does | Who notices today |
+|---|---|---|
+| `topologySpreadConstraints` | the constraint is a condition over **zero** pods and is therefore trivially satisfied: nothing is distributed, no event, no log line | nobody — and with `whenUnsatisfiable: DoNotSchedule` the object reads like a guarantee |
+| `PodDisruptionBudget` | the budget does its arithmetic over zero pods and never blocks an eviction, so `kubectl drain` takes every replica at once | nobody — `kubectl get pdb` shows an object that looks like protection |
+
+Neither is a render error, neither fails the install-test, and neither shows up in the diff of the PR that causes it: the chart that breaks it is usually a chart that **renamed a label somewhere else**. Charts answer the first half of that by keeping the selector and the pod labels in one helper (`stalwart.groupLabels`, `keycloak`'s `app` label) — but one definition only helps while nobody writes a second one, and that is the half a test can hold. Introduced with Gitea #96, where the chart-owned spread constraint made the trivially-satisfied case reachable by design rather than by accident.
+
+Three things to know before touching it:
+
+- **Helm is the YAML parser, through the one-template chart `ci/yaml2json`.** The runners have neither `pyyaml` nor `yq` (measured; the same constraint `ci/run-trivy-config.sh` and `ci/run-quantity-render-test.sh` work under), and a YAML reader written in `awk` was the obvious alternative and the wrong one: its failure mode is a parser that quietly reads nothing, finds no selector, reports no mismatch and leaves the test green for a chart it never looked at. So the rendered manifest stream is handed to that chart as a **values file** and comes back as JSON, which `python3` reads from the standard library. The one textual step is the document split on a line that is exactly `---` in column one, and getting *that* wrong does not produce a false pass — the fragment stops being valid YAML and `helm` exits 1.
+- **What is deliberately not checked, named so the gap is not mistaken for coverage.** `podAffinity`/`podAntiAffinity` are out of scope: their semantics are relative to pods that exist, so a term matching nothing is a legitimate configuration rather than a mistake. And a selector expressed **only** through `matchExpressions` is **reported as "not checked", never counted as passing** — evaluating set-based requirements means reimplementing label-selector semantics, which is a second parser by another name.
+- **Two anti-vacuum guards, because this test can be green over nothing.** Both objects are rendered only above a replica threshold, and most charts' CI fixtures sit below it. `ci/selector-match/<chart>.yaml` is a render-only overlay applied **last**, whose only job is to push a chart over that threshold — and a chart that *has* one and still renders no selector **fails**, since that overlay existing is the whole claim that something was checked. On top of that the run fails if **no selector at all** was checked anywhere. Today `stalwart` is exercised through its overlay (two replicas in both node groups plus an `externalServices.valkey.host`, without which the chart's own "cluster with no in-memory store" abort fires) and `keycloak` through its ordinary install-test fixture, which already runs two replicas.
+
+Measured: 30 charts, **5 selectors verified in ~14 s**, no cluster. The fail path was exercised rather than assumed — renaming `node-group` to `nodegroup` in a hand-written copy of the PDB selector produced two `FAIL` lines naming both the selector and the labels actually present, and a non-zero exit; the same was done for the spread constraint's selector.
+
 ### Install-test (PR pipeline stage 2)
 
 For every chart a PR changes, `ci/run-install-test.sh <chart>` runs against a throwaway k3d cluster: it applies the stub `OnePasswordItem` CRD plus the chart's fixtures, `helm install`s the chart with its CI values, and waits for all Deployments/StatefulSets to roll out (per-chart timeout). Since probes and security hardening are active on all workloads, "Ready" means the app actually answers. On failure the job dumps pod descriptions, events and container logs.
@@ -389,6 +408,9 @@ These guidelines are **binding for every current and future chart** in this repo
 - Implemented via a per-chart `<chart>.replicas` helper in `_helpers.tpl` (reference in `template-chart`); the helper uses an explicit nil check (`kindIs "invalid"`) because Helm's `default` would swallow an explicit `0`.
 - **A chart whose default is more than one replica ships a `PodDisruptionBudget`** — `policy/v1`, named `<release>-<chart>-pdb`, `minAvailable` as a value (default `1`), selector matching the workload's `app` label. `keycloak` (Gitea #83) is the first and sets the convention. The reason it is a rule and not a per-chart taste: a node drain is the most common reason a pod moves — it is what a node upgrade does — and without a budget the drain is free to take every replica at once, so the second replica buys nothing against the failure it was added for.
 - **And it is rendered only while the effective replica count is `>= 2`.** `minAvailable: 1` with exactly one replica forbids every voluntary eviction, so `kubectl drain` blocks forever on a pod nothing may move — and the message names the PDB, not the replica count, which sends the reader to the wrong file. The same applies to `scaleDown: true`, where there are no pods and the object would describe a promise about nothing. So the condition is on the **computed** replica count (the `<chart>.replicas` helper), not on the raw value, and `NOTES.txt` prints which of the two happened rather than leaving it to be inferred from `kubectl get pdb`.
+- **A chart whose replicas are a property of a *group* gets one PDB per group, with that group's own count** (Gitea #96; `stalwart` is the first, one node group per cluster role). The condition above then applies per group — with `full: 2` and `mx: 1` exactly one object is rendered — and `NOTES.txt` lists the outcome per group rather than once for the release. The name carries the group (`<release>-<chart>-<group>-pdb`), and so does the selector.
+- **A budget without a distribution rule is half an answer.** The PDB keeps a drain from taking every replica at once; it says nothing about where the replicas were put in the first place. Where a chart's replicas are meant to survive the loss of a node, it also offers a **per-group switch that renders the `topologySpreadConstraints` entry itself, with a selector the chart fills from its own pod labels** — never one the operator writes, because a spread constraint whose selector matches no pod is trivially satisfied and distributes nothing, with a green rollout and no message. `stalwart` sets the convention: `nodes.<group>.spreadAcrossNodes` (default `true`, gated on the same computed `>= 2`), `maxSkew: 1` on `kubernetes.io/hostname`, `whenUnsatisfiable: DoNotSchedule` so the distribution is a promise rather than an effort, and a generic per-group pass-through (`affinity`, `nodeSelector`, `tolerations`, `topologySpreadConstraints`) for everything the switch does not cover. Three traps come with it and all three belong in the chart's `values.yaml`: `nodeAffinityPolicy: Ignore` must be set **explicitly** (its API default is `Honor`, under which a narrowing `nodeSelector` shrinks the set of counted domains until both replicas fit in one, giving skew 0 — co-location with no `Pending`, no event, and a selector that is *correct*, so no CI check can see it); that field is **beta** and is pruned without a word below Kubernetes 1.26, which brings that exact gap back; and the replica ceiling is **a rule, not a number** — as many replicas per group as there are schedulable workers, because `nodeTaintsPolicy` defaults to `Ignore` and control-plane domains therefore hold the minimum at 0 permanently. That ceiling comes from `nodeTaintsPolicy` and is **not** a consequence of setting `nodeAffinityPolicy`; the two are documented separately for that reason. A user entry colliding with the chart's own on `(topologyKey, whenUnsatisfiable)` **aborts the render naming both values**, because the two alternatives are an API error pointing at a list index and a silent merge over `patchMergeKey: topologyKey` that drops one selector and goes green.
+- **Both selectors are held against the rendered pod labels on every PR** by `ci/run-selector-match-test.sh` — see "Selector-match render test" above. Needed because the shared-helper rule above is a convention, and a convention is the thing a copy breaks.
 
 ### 12. Values that must not lie (#68, #93, #99, #215)