From 03f5a7030f165e535e4c3e2717a9ca2678c9dea2 Mon Sep 17 00:00:00 2001
From: huskcass <333684917+huskcass@users.noreply.github.com>
Date: Thu, 22 Oct 2026 04:56:06 -0700
Subject: [PATCH 1/2] test: load every unit and envtest case from YAML files

Each case now lives in its own file under a testdata/ directory next
to its package (testdata/<suite>/<case-name>.yaml). The Go test code
keeps only the loaders, runners, fixtures and assertions.

internal/testutil/cases walks a directory for *.yaml, decodes each
file strictly with sigs.k8s.io/yaml (unknown fields fail), orders the
files by relative path and fails on an empty directory. cases.Run
starts one subtest per file; cases.Entries builds one Ginkgo entry
per file.

Covered: autoscaling (guards, helpers, custom provider, LLMScaler
envtest tables), serving (LLMService envtest specs, predicate, repo,
values), chart (version, repo) and helm (client, namespace guard).
Case counts per table and spec names are unchanged; per-package
coverage is unchanged. DEVELOPMENT.md explains how to add a case.
---
 DEVELOPMENT.md                                |  20 +
 internal/chart/repo_test.go                   | 425 +++++---
 ...a-missing-index-is-reported-by-status.yaml |   7 +
 .../a-refused-read-says-it-needs-a-login.yaml |  23 +
 ...y-asking-for-basic-auth-needs-a-login.yaml |  14 +
 .../versions-auth-sends-basic-auth.yaml       |  50 +
 .../versions-from-a-local-chart.yaml          |  11 +
 .../versions-from-a-repo-index.yaml           |  25 +
 .../versions-from-an-oci-registry.yaml        |  32 +
 .../chart/testdata/version/bare/exact.yaml    |   7 +
 .../chart/testdata/version/bare/range.yaml    |   3 +
 .../chart/testdata/version/match/caret.yaml   |   9 +
 .../testdata/version/match/newest-first.yaml  |  14 +
 .../chart/testdata/version/parse/banana.yaml  |   2 +
 .../chart/testdata/version/parse/empty.yaml   |   2 +
 .../chart/testdata/version/parse/latest.yaml  |   2 +
 .../resolve-refuses/named-out-of-range.yaml   |   5 +
 .../version/resolve-refuses/nil-lister.yaml   |   4 +
 .../resolve-refuses/nothing-in-range.yaml     |   5 +
 .../resolve/exact-moves-running-off.yaml      |   5 +
 .../resolve/exact-needs-no-registry.yaml      |   4 +
 .../version/resolve/named-in-range.yaml       |   5 +
 .../version/resolve/newest-in-range.yaml      |   7 +
 .../resolve/running-kept-in-range.yaml        |   5 +
 .../version/resolve/running-out-of-range.yaml |   8 +
 internal/chart/version_test.go                | 183 ++--
 .../autoscaling/customprovider_test.go        | 466 ++++-----
 .../autoscaling/guards_cases_test.go          | 404 --------
 .../controller/autoscaling/guards_test.go     | 300 +++---
 .../autoscaling/llmscaler_casefile_test.go    | 399 ++++++++
 .../llmscaler_controller_cases_test.go        | 722 --------------
 .../autoscaling/llmscaler_controller_test.go  |  77 +-
 .../llmscaler_helpers_cases_test.go           | 199 ----
 .../autoscaling/llmscaler_helpers_test.go     | 520 +++++-----
 .../autoscaling/llmscaler_scenario_test.go    |   8 -
 ...ld-default-defertimeoutseconds-to-600.yaml |   6 +
 ...-custom-scaler-with-no-customprovider.yaml |   6 +
 ...t-a-prometheus-scaler-with-no-metrics.yaml |   8 +
 ...-and-customprovider-path-to-decisions.yaml |  11 +
 ...en-the-decision-server-is-unreachable.yaml |  20 +
 ...the-count-the-decision-server-reports.yaml |  12 +
 ...rral-on-a-no-op-once-no-guard-matches.yaml |  25 +
 ...s-while-a-rollout-guard-still-matches.yaml |  64 ++
 ...eferral-when-the-guard-reason-changes.yaml |  40 +
 ...ded-the-timeout-on-a-fresh-reconciler.yaml |  23 +
 ...ady-fleet-once-the-deferral-times-out.yaml |  52 +
 ...-rollout-exceeds-the-deferral-timeout.yaml |  55 ++
 ...eferral-when-defertimeoutseconds-is-0.yaml |  22 +
 ...rced-write-pace-when-no-guard-matches.yaml |  33 +
 ...scale-down-step-after-one-sync-period.yaml |  41 +
 ...-step-until-a-sync-period-has-elapsed.yaml |  47 +
 ...a-fallback-for-the-scaling-ratio-only.yaml |  22 +
 ...o-minreplicas-when-the-metric-is-zero.yaml |  23 +
 ...le-down-to-zero-when-minreplicas-is-0.yaml |  26 +
 ...scale-the-deployment-based-on-metrics.yaml |  13 +
 ...m-zero-when-the-metric-turns-positive.yaml |  22 +
 ...g-while-the-deployment-is-rolling-out.yaml |  17 +
 ...n-when-checkterminatingreplicas-is-on.yaml |  32 +
 ...readyreplicas-is-back-at-specreplicas.yaml |  29 +
 ...while-checkterminatingreplicas-is-off.yaml |  22 +
 ...p-on-the-previous-scale-down-settling.yaml |  21 +
 ...ation-window-between-scale-down-steps.yaml |  29 +
 ...at-a-time-when-maxstepreplicas-is-set.yaml |  33 +
 ...ollout-fully-rolled-out-does-not-hold.yaml |  18 +
 ...eration-not-observed-holds-a-scale-up.yaml |  19 +
 .../rollout-non-deployment-is-never-held.yaml |  17 +
 ...ollout-old-template-pods-hold-a-no-op.yaml |  19 +
 ...plicas-behind-spec-holds-a-scale-down.yaml |  19 +
 .../guards/hold/settle-no-op-is-not-held.yaml |  13 +
 .../hold/settle-ready-back-at-spec.yaml       |  13 +
 ...ettle-ready-count-of-0-is-not-settled.yaml |  15 +
 .../hold/settle-ready-differs-from-spec.yaml  |  14 +
 .../hold/settle-scale-up-is-not-held.yaml     |  13 +
 ...rminating-count-of-0-with-the-gate-on.yaml |  16 +
 ...inating-not-reported-with-the-gate-on.yaml |  17 +
 ...le-terminating-pods-with-the-gate-off.yaml |  15 +
 ...tle-terminating-pods-with-the-gate-on.yaml |  17 +
 .../no-guard-holds-a-settled-scale-down.yaml  |  15 +
 ...out-wins-over-an-unsettled-scale-down.yaml |  17 +
 ...settle-holds-once-the-rollout-is-done.yaml |  17 +
 ...-deployment-ignores-the-rollout-guard.yaml |  24 +
 .../ready-count-of-0-is-not-settled.yaml      |  27 +
 .../target/rollout-holds-a-scale-up.yaml      |  26 +
 ...out-wins-over-an-unsettled-scale-down.yaml |  26 +
 .../scale-up-is-not-gated-on-settle.yaml      |  24 +
 .../settled-scale-down-may-proceed.yaml       |  24 +
 .../guards/target/unsettled-scale-down.yaml   |  27 +
 .../deployment-with-full-status.yaml          |  27 +
 .../generic-kind-with-full-status.yaml        |  26 +
 ...alid-typed-fields-behave-like-missing.yaml |  20 +
 .../leaderworkerset-with-full-status.yaml     |  26 +
 ...eneration-updated-and-total-read-as-0.yaml |  18 +
 .../missing-readyreplicas-reads-as-0.yaml     |  26 +
 .../missing-spec-replicas-reads-as-1.yaml     |  25 +
 ...g-terminatingreplicas-is-not-reported.yaml |  24 +
 .../spec-replicas-of-0-is-kept.yaml           |  21 +
 .../statefulset-with-full-status.yaml         |  26 +
 .../helpers/compute-desired/minus-inf.yaml    |   9 +
 .../testdata/helpers/compute-desired/nan.yaml |   9 +
 .../helpers/compute-desired/plus-inf.yaml     |   9 +
 .../positive-signal-wakes-the-fleet.yaml      |  14 +
 .../zero-signal-recommends-zero.yaml          |  14 +
 .../empty-serveraddress.yaml                  |   7 +
 .../empty-serviceid.yaml                      |   7 +
 ...rror-even-when-the-body-is-a-decision.yaml |   8 +
 .../a-body-that-is-not-json-is-an-error.yaml  |   7 +
 ...-int32-is-refused-rather-than-wrapped.yaml |   7 +
 ...on-for-another-service-does-not-match.yaml |   8 +
 ...on-with-no-active-count-is-not-a-zero.yaml |   7 +
 ...decision-resolves-to-its-active-count.yaml |   7 +
 .../a-negative-count-is-refused.yaml          |   7 +
 ...payload-with-no-apiversion-is-refused.yaml |   7 +
 .../an-empty-decision-list-is-not-a-zero.yaml |   7 +
 ...chema-is-refused-not-parsed-hopefully.yaml |   7 +
 ...licates-are-an-error-not-a-pick-first.yaml |   7 +
 ...lter-is-reported-when-nothing-matches.yaml |   8 +
 ...ter-picks-between-same-named-services.yaml |   8 +
 .../zero-is-a-decision-not-a-missing-one.yaml |   7 +
 .../a-configured-path-is-used.yaml            |  12 +
 ...thout-a-leading-slash-still-forms-one.yaml |  13 +
 ...slash-on-serveraddress-is-not-doubled.yaml |  14 +
 ...-empty-path-falls-back-to-the-default.yaml |  12 +
 ...ecision-server-payload-resolves-to-15.yaml |  21 +
 .../helpers/defer-timeout/explicit-0.yaml     |   6 +
 .../helpers/defer-timeout/explicit-30.yaml    |   6 +
 .../helpers/defer-timeout/nil-spec.yaml       |   6 +
 ...dations-clear-the-forced-write-record.yaml |  16 +
 .../deadline-already-passed.yaml              |   8 +
 .../deadline-due-this-instant.yaml            |   8 +
 .../deadline-later-than-sync-period.yaml      |   8 +
 .../deadline-sooner-than-sync-period.yaml     |   8 +
 .../deadline-under-the-1s-floor.yaml          |   8 +
 .../not-holding-stays-on-sync-period.yaml     |   8 +
 .../pacing-later-than-sync-period.yaml        |   9 +
 .../pacing-sooner-than-sync-period.yaml       |   9 +
 .../pacing-under-the-1s-floor.yaml            |   9 +
 .../timeout-0-stays-on-sync-period.yaml       |   8 +
 .../14s-into-a-15s-period.yaml                |   9 +
 .../exactly-one-sync-period.yaml              |   8 +
 .../forced-write-pace/just-recorded.yaml      |   8 +
 .../helpers/forced-write-pace/no-record.yaml  |   9 +
 .../forced-write-pace/zero-sync-period.yaml   |   9 +
 .../helpers/parse-target-value/empty.yaml     |   6 +
 ...f-parses-as-a-float-but-is-not-usable.yaml |   6 +
 ...n-parses-as-a-float-but-is-not-usable.yaml |   6 +
 .../negative-would-invert-scaling.yaml        |   6 +
 .../parse-target-value/not-a-number.yaml      |   6 +
 ...rcent-suffix-is-rejected-not-stripped.yaml |   6 +
 .../parse-target-value/plain-integer.yaml     |   6 +
 .../helpers/parse-target-value/ratio.yaml     |   6 +
 .../surrounding-whitespace.yaml               |   6 +
 .../zero-would-make-the-ratio-infinite.yaml   |   6 +
 .../59s-of-a-60s-timeout.yaml                 |   7 +
 .../exactly-the-timeout.yaml                  |   7 +
 .../past-defer-timeout/nil-deferredsince.yaml |   8 +
 .../past-defer-timeout/past-the-timeout.yaml  |   7 +
 .../helpers/past-defer-timeout/timeout-0.yaml |   7 +
 ...d-query-is-rejected-naming-the-series.yaml |   9 +
 .../empty-vector-is-not-zero.yaml             |   7 +
 ...ti-series-error-names-both-label-sets.yaml |  10 +
 .../single-series-resolves.yaml               |   6 +
 .../pins-the-published-condition-reason.yaml  |   5 +
 ...decision-above-maxreplicas-is-clamped.yaml |  14 +
 ...sion-of-zero-clamps-up-to-minreplicas.yaml |  14 +
 ...mprovider-holds-the-fleet-where-it-is.yaml |  15 +
 ...-decision-holds-the-fleet-where-it-is.yaml |  13 +
 ...sed-as-is-not-scaled-by-readyreplicas.yaml |  15 +
 ...tricprovider-still-queries-prometheus.yaml |  16 +
 .../scale-guard/caps-a-burst-scale-down.yaml  |   8 +
 .../scale-guard/leaves-a-no-op-untouched.yaml |   8 +
 .../leaves-scale-up-untouched.yaml            |   8 +
 ...ses-through-a-step-exactly-at-the-cap.yaml |   8 +
 .../passes-through-a-step-within-the-cap.yaml |   8 +
 .../unlimited-when-maxstep-is-0.yaml          |   8 +
 ...-holds-inside-the-window-then-expires.yaml |  19 +
 ...fter-a-scale-down-paces-the-next-step.yaml |  32 +
 .../window-0-makes-the-pin-a-no-op.yaml       |  12 +
 .../serving/llmservice_cases_test.go          | 892 +++++++++++++++++
 .../serving/llmservice_controller_test.go     | 922 +-----------------
 internal/controller/serving/predicate_test.go |  80 +-
 internal/controller/serving/repo_test.go      |  42 +-
 ...-a-matching-release-and-refuses-drift.yaml |  57 ++
 .../clamps-a-short-pending-wait-to-15s.yaml   |  21 +
 ...-adopt-a-release-that-is-not-deployed.yaml |  28 +
 ...-call-helm-when-the-spec-is-unchanged.yaml |  15 +
 ...ite-a-pending-release-it-does-not-own.yaml |  36 +
 ...er-when-the-release-is-already-absent.yaml |  16 +
 ...-the-chart-version-cannot-be-resolved.yaml |  19 +
 ...licts-only-for-the-current-generation.yaml |  27 +
 ...tall-and-a-transient-chart-list-error.yaml |  57 ++
 ...alls-and-records-a-controllerrevision.yaml |  44 +
 ...-never-installed-or-adopted-on-delete.yaml |  61 ++
 ...ved-version-while-the-range-allows-it.yaml |  25 +
 ...neration-while-install-is-in-progress.yaml |  26 +
 ...aves-status-unchanged-while-suspended.yaml |  25 +
 ...ans-a-delete-and-uninstalls-otherwise.yaml |  27 +
 ...esolves-once-per-new-re-resolve-value.yaml |  54 +
 .../controller/records-a-helm-failure.yaml    |  20 +
 ...es-as-a-controllerrevision-of-its-own.yaml |  41 +
 ...le-owned-pending-install-and-installs.yaml |  23 +
 ...ry-it-was-not-started-to-install-from.yaml |  26 +
 ...snapshot-under-a-reused-revision-name.yaml |  14 +
 ...-instead-of-drifting-a-failed-release.yaml |  35 +
 ...le-owned-pending-upgrade-and-upgrades.yaml |  34 +
 ...d-release-when-applied-is-applyfailed.yaml |  33 +
 ...ollback-that-has-no-deployed-revision.yaml |  28 +
 ...installs-an-adopted-release-on-delete.yaml |  25 +
 ...eployed-release-whose-status-was-lost.yaml |  24 +
 ...se-that-still-has-a-deployed-revision.yaml |  28 +
 .../upgrades-and-bounds-history.yaml          |  34 +
 ...its-out-a-fresh-owned-pending-upgrade.yaml |  42 +
 .../testdata/predicate/annotation-change.yaml |   6 +
 .../serving/testdata/predicate/create.yaml    |   4 +
 .../serving/testdata/predicate/delete.yaml    |   4 +
 .../predicate/deletion-timestamp.yaml         |   5 +
 .../testdata/predicate/generation-change.yaml |   5 +
 .../predicate/status-only-ignored.yaml        |   5 +
 .../serving/testdata/repo/empty.yaml          |   7 +
 ...tps-modelsphere-github-io-helm-charts.yaml |   7 +
 .../repo/https-modelsphere-github-io.yaml     |   7 +
 .../serving/testdata/repo/nil-allowlist.yaml  |   5 +
 .../oci-ghcr-io-modelsphere-charts-slash.yaml |   7 +
 .../repo/oci-ghcr-io-modelsphere-charts.yaml  |   7 +
 .../oci-ghcr-io-modelsphere-x-charts.yaml     |   7 +
 .../repo/oci-ghcr-io-modelsphere.yaml         |   7 +
 .../repo/oci-ghcr-io-other-charts.yaml        |   7 +
 .../values/hash/stable-and-sensitive.yaml     |  32 +
 .../serving/testdata/values/merge/flat.yaml   |  23 +
 .../serving/testdata/values/merge/nested.yaml |  12 +
 internal/controller/serving/values_test.go    | 169 ++--
 internal/helm/guard_test.go                   | 150 +--
 internal/helm/helm_test.go                    | 368 +++----
 .../get-install-upgrade-uninstall.yaml        |  58 ++
 .../history-newest-first.yaml                 |  56 ++
 .../pending-install-failed.yaml               |  38 +
 .../rollback-pending-rollback.yaml            |  26 +
 internal/helm/testdata/guard/apply/hook.yaml  |  14 +
 .../helm/testdata/guard/apply/template.yaml   |  12 +
 .../helm/testdata/guard/apply/upgrade.yaml    |  10 +
 .../guard/namespace/after-a-good-one.yaml     |  12 +
 .../guard/namespace/another-namespace.yaml    |   8 +
 .../guard/namespace/cluster-scoped.yaml       |   7 +
 .../guard/namespace/comments-only.yaml        |   5 +
 .../guard/namespace/inside-a-list.yaml        |  15 +
 .../guard/namespace/namespace-object.yaml     |   7 +
 .../testdata/guard/namespace/no-kind.yaml     |   4 +
 .../guard/namespace/no-namespace.yaml         |   7 +
 .../guard/namespace/own-namespace.yaml        |   8 +
 .../guard/namespace/unknown-kind.yaml         |   7 +
 internal/testutil/cases/cases.go              | 236 +++++
 internal/testutil/cases/cases_test.go         | 152 +++
 .../testutil/cases/testdata/bad-yaml/x.yaml   |   1 +
 .../cases/testdata/name-mismatch/nope.yaml    |   6 +
 .../testutil/cases/testdata/no-yaml/note.txt  |   1 +
 internal/testutil/cases/testdata/ok/a/c.yaml  |   9 +
 internal/testutil/cases/testdata/ok/b.yaml    |   6 +
 .../cases/testdata/unknown-field/bad.yaml     |   7 +
 257 files changed, 7136 insertions(+), 3452 deletions(-)
 create mode 100644 internal/chart/testdata/repo/a-missing-index-is-reported-by-status/a-missing-index-is-reported-by-status.yaml
 create mode 100644 internal/chart/testdata/repo/a-refused-read-says-it-needs-a-login/a-refused-read-says-it-needs-a-login.yaml
 create mode 100644 internal/chart/testdata/repo/a-registry-asking-for-basic-auth-needs-a-login/a-registry-asking-for-basic-auth-needs-a-login.yaml
 create mode 100644 internal/chart/testdata/repo/versions-auth-sends-basic-auth/versions-auth-sends-basic-auth.yaml
 create mode 100644 internal/chart/testdata/repo/versions-from-a-local-chart/versions-from-a-local-chart.yaml
 create mode 100644 internal/chart/testdata/repo/versions-from-a-repo-index/versions-from-a-repo-index.yaml
 create mode 100644 internal/chart/testdata/repo/versions-from-an-oci-registry/versions-from-an-oci-registry.yaml
 create mode 100644 internal/chart/testdata/version/bare/exact.yaml
 create mode 100644 internal/chart/testdata/version/bare/range.yaml
 create mode 100644 internal/chart/testdata/version/match/caret.yaml
 create mode 100644 internal/chart/testdata/version/match/newest-first.yaml
 create mode 100644 internal/chart/testdata/version/parse/banana.yaml
 create mode 100644 internal/chart/testdata/version/parse/empty.yaml
 create mode 100644 internal/chart/testdata/version/parse/latest.yaml
 create mode 100644 internal/chart/testdata/version/resolve-refuses/named-out-of-range.yaml
 create mode 100644 internal/chart/testdata/version/resolve-refuses/nil-lister.yaml
 create mode 100644 internal/chart/testdata/version/resolve-refuses/nothing-in-range.yaml
 create mode 100644 internal/chart/testdata/version/resolve/exact-moves-running-off.yaml
 create mode 100644 internal/chart/testdata/version/resolve/exact-needs-no-registry.yaml
 create mode 100644 internal/chart/testdata/version/resolve/named-in-range.yaml
 create mode 100644 internal/chart/testdata/version/resolve/newest-in-range.yaml
 create mode 100644 internal/chart/testdata/version/resolve/running-kept-in-range.yaml
 create mode 100644 internal/chart/testdata/version/resolve/running-out-of-range.yaml
 delete mode 100644 internal/controller/autoscaling/guards_cases_test.go
 create mode 100644 internal/controller/autoscaling/llmscaler_casefile_test.go
 delete mode 100644 internal/controller/autoscaling/llmscaler_controller_cases_test.go
 delete mode 100644 internal/controller/autoscaling/llmscaler_helpers_cases_test.go
 create mode 100644 internal/controller/autoscaling/testdata/controller/crd-default/should-default-defertimeoutseconds-to-600.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-custom-scaler-with-no-customprovider.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-prometheus-scaler-with-no-metrics.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/custom-defaults/should-default-metricprovider-to-prometheus-and-customprovider-path-to-decisions.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/custom-reconcile/should-hold-replicas-when-the-decision-server-is-unreachable.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/custom-reconcile/should-scale-the-deployment-to-the-count-the-decision-server-reports.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-status/should-clear-a-deferral-on-a-no-op-once-no-guard-matches.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-deferredsince-across-no-op-syncs-while-a-rollout-guard-still-matches.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-one-deferral-when-the-guard-reason-changes.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-deferral-that-already-exceeded-the-timeout-on-a-fresh-reconciler.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-down-of-an-unready-fleet-once-the-deferral-times-out.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-up-once-a-stuck-rollout-exceeds-the-deferral-timeout.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/deferral-timeout/should-never-force-a-deferral-when-defertimeoutseconds-is-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-clear-deferredsince-and-the-forced-write-pace-when-no-guard-matches.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-force-the-next-stuck-scale-down-step-after-one-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-not-take-a-second-forced-step-until-a-sync-period-has-elapsed.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/metric-scaling/should-keep-the-ready-replica-fallback-for-the-scaling-ratio-only.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-minreplicas-when-the-metric-is-zero.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-zero-when-minreplicas-is-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/metric-scaling/should-successfully-scale-the-deployment-based-on-metrics.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/metric-scaling/should-wake-the-fleet-from-zero-when-the-metric-turns-positive.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/rollout-guard/should-defer-scaling-while-the-deployment-is-rolling-out.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/settle-guard/should-also-wait-for-the-drain-when-checkterminatingreplicas-is-on.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/settle-guard/should-defer-a-scale-down-until-readyreplicas-is-back-at-specreplicas.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/settle-guard/should-ignore-still-draining-pods-while-checkterminatingreplicas-is-off.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/settle-guard/should-not-gate-scale-up-on-the-previous-scale-down-settling.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-hold-for-the-stabilization-window-between-scale-down-steps.yaml
 create mode 100644 internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-walk-down-one-step-at-a-time-when-maxstepreplicas-is-set.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/rollout-fully-rolled-out-does-not-hold.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/rollout-generation-not-observed-holds-a-scale-up.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/rollout-non-deployment-is-never-held.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/rollout-old-template-pods-hold-a-no-op.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/rollout-updated-replicas-behind-spec-holds-a-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-no-op-is-not-held.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-ready-back-at-spec.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-ready-count-of-0-is-not-settled.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-ready-differs-from-spec.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-scale-up-is-not-held.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-terminating-count-of-0-with-the-gate-on.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-terminating-not-reported-with-the-gate-on.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-off.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-on.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/order/no-guard-holds-a-settled-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/order/rollout-wins-over-an-unsettled-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/order/settle-holds-once-the-rollout-is-done.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/non-deployment-ignores-the-rollout-guard.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/ready-count-of-0-is-not-settled.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/rollout-holds-a-scale-up.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/rollout-wins-over-an-unsettled-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/scale-up-is-not-gated-on-settle.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/settled-scale-down-may-proceed.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/target/unsettled-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/deployment-with-full-status.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/generic-kind-with-full-status.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/invalid-typed-fields-behave-like-missing.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/leaderworkerset-with-full-status.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/missing-generation-observedgeneration-updated-and-total-read-as-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/missing-readyreplicas-reads-as-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/missing-spec-replicas-reads-as-1.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/missing-terminatingreplicas-is-not-reported.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/spec-replicas-of-0-is-kept.yaml
 create mode 100644 internal/controller/autoscaling/testdata/guards/workload-state/statefulset-with-full-status.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/compute-desired/minus-inf.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/compute-desired/nan.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/compute-desired/plus-inf.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/compute-desired/positive-signal-wakes-the-fleet.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/compute-desired/zero-signal-recommends-zero.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serveraddress.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serviceid.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-non-200/a-non-ok-reply-is-an-error-even-when-the-body-is-a-decision.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-request/a-configured-path-is-used.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-request/a-path-without-a-leading-slash-still-forms-one.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-request/a-trailing-slash-on-serveraddress-is-not-doubled.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-request/an-empty-path-falls-back-to-the-default.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-30.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/defer-timeout/nil-spec.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-memory/enddeferral-and-forgetrecommendations-clear-the-forced-write-record.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-already-passed.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-due-this-instant.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-later-than-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-sooner-than-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-under-the-1s-floor.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/not-holding-stays-on-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-later-than-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-sooner-than-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-under-the-1s-floor.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/timeout-0-stays-on-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/forced-write-pace/14s-into-a-15s-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/forced-write-pace/exactly-one-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/forced-write-pace/just-recorded.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/forced-write-pace/no-record.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/forced-write-pace/zero-sync-period.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/empty.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/inf-parses-as-a-float-but-is-not-usable.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/nan-parses-as-a-float-but-is-not-usable.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/negative-would-invert-scaling.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/not-a-number.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/percent-suffix-is-rejected-not-stripped.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/plain-integer.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/ratio.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/surrounding-whitespace.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/parse-target-value/zero-would-make-the-ratio-infinite.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/past-defer-timeout/59s-of-a-60s-timeout.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/past-defer-timeout/exactly-the-timeout.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/past-defer-timeout/nil-deferredsince.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/past-defer-timeout/past-the-timeout.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/past-defer-timeout/timeout-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/an-under-aggregated-query-is-rejected-naming-the-series.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/empty-vector-is-not-zero.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/multi-series-error-names-both-label-sets.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/single-series-resolves.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/reason-deferral-timed-out/pins-the-published-condition-reason.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-above-maxreplicas-is-clamped.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-of-zero-clamps-up-to-minreplicas.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-custom/a-missing-customprovider-holds-the-fleet-where-it-is.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-custom/an-unreadable-decision-holds-the-fleet-where-it-is.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-custom/the-decision-is-used-as-is-not-scaled-by-readyreplicas.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/recommend-default-prometheus/an-empty-metricprovider-still-queries-prometheus.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/caps-a-burst-scale-down.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-a-no-op-untouched.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-scale-up-untouched.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-exactly-at-the-cap.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-within-the-cap.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/scale-guard/unlimited-when-maxstep-is-0.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/stabilization/peak-holds-inside-the-window-then-expires.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/stabilization/pin-after-a-scale-down-paces-the-next-step.yaml
 create mode 100644 internal/controller/autoscaling/testdata/helpers/stabilization/window-0-makes-the-pin-a-no-op.yaml
 create mode 100644 internal/controller/serving/llmservice_cases_test.go
 create mode 100644 internal/controller/serving/testdata/controller/adopts-a-matching-release-and-refuses-drift.yaml
 create mode 100644 internal/controller/serving/testdata/controller/clamps-a-short-pending-wait-to-15s.yaml
 create mode 100644 internal/controller/serving/testdata/controller/does-not-adopt-a-release-that-is-not-deployed.yaml
 create mode 100644 internal/controller/serving/testdata/controller/does-not-call-helm-when-the-spec-is-unchanged.yaml
 create mode 100644 internal/controller/serving/testdata/controller/does-not-write-a-pending-release-it-does-not-own.yaml
 create mode 100644 internal/controller/serving/testdata/controller/drops-the-finalizer-when-the-release-is-already-absent.yaml
 create mode 100644 internal/controller/serving/testdata/controller/fails-when-the-chart-version-cannot-be-resolved.yaml
 create mode 100644 internal/controller/serving/testdata/controller/honors-force-conflicts-only-for-the-current-generation.yaml
 create mode 100644 internal/controller/serving/testdata/controller/installs-again-after-a-failed-install-and-a-transient-chart-list-error.yaml
 create mode 100644 internal/controller/serving/testdata/controller/installs-and-records-a-controllerrevision.yaml
 create mode 100644 internal/controller/serving/testdata/controller/keeps-a-release-it-never-installed-or-adopted-on-delete.yaml
 create mode 100644 internal/controller/serving/testdata/controller/keeps-a-resolved-version-while-the-range-allows-it.yaml
 create mode 100644 internal/controller/serving/testdata/controller/keeps-observedgeneration-while-install-is-in-progress.yaml
 create mode 100644 internal/controller/serving/testdata/controller/leaves-status-unchanged-while-suspended.yaml
 create mode 100644 internal/controller/serving/testdata/controller/orphans-a-delete-and-uninstalls-otherwise.yaml
 create mode 100644 internal/controller/serving/testdata/controller/re-resolves-once-per-new-re-resolve-value.yaml
 create mode 100644 internal/controller/serving/testdata/controller/records-a-helm-failure.yaml
 create mode 100644 internal/controller/serving/testdata/controller/records-a-rollback-to-earlier-values-as-a-controllerrevision-of-its-own.yaml
 create mode 100644 internal/controller/serving/testdata/controller/recovers-a-stale-owned-pending-install-and-installs.yaml
 create mode 100644 internal/controller/serving/testdata/controller/refuses-a-chart-repository-it-was-not-started-to-install-from.yaml
 create mode 100644 internal/controller/serving/testdata/controller/replaces-its-own-stale-snapshot-under-a-reused-revision-name.yaml
 create mode 100644 internal/controller/serving/testdata/controller/requeues-a-helm-get-error-instead-of-drifting-a-failed-release.yaml
 create mode 100644 internal/controller/serving/testdata/controller/rolls-back-a-stale-owned-pending-upgrade-and-upgrades.yaml
 create mode 100644 internal/controller/serving/testdata/controller/still-drifts-a-foreign-failed-release-when-applied-is-applyfailed.yaml
 create mode 100644 internal/controller/serving/testdata/controller/uninstalls-a-stale-owned-pending-rollback-that-has-no-deployed-revision.yaml
 create mode 100644 internal/controller/serving/testdata/controller/uninstalls-an-adopted-release-on-delete.yaml
 create mode 100644 internal/controller/serving/testdata/controller/upgrades-an-owned-deployed-release-whose-status-was-lost.yaml
 create mode 100644 internal/controller/serving/testdata/controller/upgrades-an-owned-failed-release-that-still-has-a-deployed-revision.yaml
 create mode 100644 internal/controller/serving/testdata/controller/upgrades-and-bounds-history.yaml
 create mode 100644 internal/controller/serving/testdata/controller/waits-out-a-fresh-owned-pending-upgrade.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/annotation-change.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/create.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/delete.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/deletion-timestamp.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/generation-change.yaml
 create mode 100644 internal/controller/serving/testdata/predicate/status-only-ignored.yaml
 create mode 100644 internal/controller/serving/testdata/repo/empty.yaml
 create mode 100644 internal/controller/serving/testdata/repo/https-modelsphere-github-io-helm-charts.yaml
 create mode 100644 internal/controller/serving/testdata/repo/https-modelsphere-github-io.yaml
 create mode 100644 internal/controller/serving/testdata/repo/nil-allowlist.yaml
 create mode 100644 internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts-slash.yaml
 create mode 100644 internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts.yaml
 create mode 100644 internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-x-charts.yaml
 create mode 100644 internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere.yaml
 create mode 100644 internal/controller/serving/testdata/repo/oci-ghcr-io-other-charts.yaml
 create mode 100644 internal/controller/serving/testdata/values/hash/stable-and-sensitive.yaml
 create mode 100644 internal/controller/serving/testdata/values/merge/flat.yaml
 create mode 100644 internal/controller/serving/testdata/values/merge/nested.yaml
 create mode 100644 internal/helm/testdata/client/get-install-upgrade-uninstall/get-install-upgrade-uninstall.yaml
 create mode 100644 internal/helm/testdata/client/history-newest-first/history-newest-first.yaml
 create mode 100644 internal/helm/testdata/client/pending-install-failed/pending-install-failed.yaml
 create mode 100644 internal/helm/testdata/client/rollback-pending-rollback/rollback-pending-rollback.yaml
 create mode 100644 internal/helm/testdata/guard/apply/hook.yaml
 create mode 100644 internal/helm/testdata/guard/apply/template.yaml
 create mode 100644 internal/helm/testdata/guard/apply/upgrade.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/after-a-good-one.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/another-namespace.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/cluster-scoped.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/comments-only.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/inside-a-list.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/namespace-object.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/no-kind.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/no-namespace.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/own-namespace.yaml
 create mode 100644 internal/helm/testdata/guard/namespace/unknown-kind.yaml
 create mode 100644 internal/testutil/cases/cases.go
 create mode 100644 internal/testutil/cases/cases_test.go
 create mode 100644 internal/testutil/cases/testdata/bad-yaml/x.yaml
 create mode 100644 internal/testutil/cases/testdata/name-mismatch/nope.yaml
 create mode 100644 internal/testutil/cases/testdata/no-yaml/note.txt
 create mode 100644 internal/testutil/cases/testdata/ok/a/c.yaml
 create mode 100644 internal/testutil/cases/testdata/ok/b.yaml
 create mode 100644 internal/testutil/cases/testdata/unknown-field/bad.yaml

diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md
index 74f0d95..103fa9a 100644
--- a/DEVELOPMENT.md
+++ b/DEVELOPMENT.md
@@ -137,6 +137,26 @@ The release workflow packages and publishes `dist/chart`.
 
 More information can be found via the [Kubebuilder Documentation](https://book.kubebuilder.io/introduction.html)
 
+## File-based tests
+
+Unit and envtest cases live in YAML under a `testdata/` directory next to the package, one file per case. The Go tests load those files, stand up fixtures, and check the result. Kind-based end-to-end tests under `test/e2e` are unchanged.
+
+```
+internal/<pkg>/testdata/<suite>/<case-name>.yaml
+```
+
+The suite directory groups a table (`guards/hold`, `controller/metric-scaling`). The file name is the case name in kebab-case. A document has:
+
+- `name`, optional. When set it must equal the file name without `.yaml`.
+- `description`, the sentence `go test -v` and Ginkgo print. When it is empty, the file name is used.
+- `inputs` and `expected`, decoded into a Go struct. An unknown field, a duplicate key, or a value of the wrong type fails the load.
+
+Durations are Go duration strings (`30s`, `1m1s`, `-2h`). A nil pointer is an omitted field or `null`. A pointer to zero is `0` or `"0s"`, not an omission. Times relative to a fixed origin are offsets from that origin; the runner adds them.
+
+The loader is `internal/testutil/cases`. It reads every `*.yaml` under the directory, in path order, and rejects an empty directory. `cases.Run` starts one subtest per file. Ginkgo tables use `cases.Entries`, one entry per file.
+
+To add a case, copy a nearby file, change the inputs and the expected result, and run `make test`.
+
 ## License
 
 Copyright 2026.
diff --git a/internal/chart/repo_test.go b/internal/chart/repo_test.go
index 00f06b8..b2d658e 100644
--- a/internal/chart/repo_test.go
+++ b/internal/chart/repo_test.go
@@ -4,6 +4,7 @@ import (
 	"encoding/base64"
 	"errors"
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -11,181 +12,321 @@ import (
 	"reflect"
 	"strings"
 	"testing"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-func TestVersionsFromARepoIndex(t *testing.T) {
-	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/charts/index.yaml" {
-			http.NotFound(w, r)
-			return
-		}
-		_, _ = fmt.Fprint(w, "apiVersion: v1\nentries:\n  sglang:\n    - version: 0.8.0\n    - version: 0.7.1\n  vllm:\n    - version: 1.0.0\n")
-	}))
-	t.Cleanup(srv.Close)
+// chartCall is one Versions or VersionsAuth read.
+// repoKind is https, oci, or local. A local read writes Chart.yaml from
+// chartYAML and passes localRepo through; the directory wins over that URL.
+// {server}, {repo}, {indexURL}, and {basic} in expected errors are filled
+// once the server exists. {basic} is "Basic " plus base64(username:password).
+type chartCall struct {
+	Fn        string `json:"fn"`
+	RepoKind  string `json:"repoKind"`
+	RepoPath  string `json:"repoPath,omitempty"`
+	Chart     string `json:"chart"`
+	Username  string `json:"username,omitempty"`
+	Password  string `json:"password,omitempty"`
+	LocalRepo string `json:"localRepo,omitempty"`
+	ChartYAML string `json:"chartYAML,omitempty"`
 
-	got, err := Versions(t.Context(), srv.URL+"/charts/", "", "sglang")
-	if err != nil || !reflect.DeepEqual(got, []string{v080, v071}) {
-		t.Fatalf("got %v, %v", got, err)
-	}
-	if _, err := Versions(t.Context(), srv.URL+"/charts", "", "nope"); err == nil || !strings.Contains(err.Error(), `no chart "nope"`) {
-		t.Errorf("a chart the index lacks: %v", err)
-	}
+	Versions         *[]string `json:"versions,omitempty"`
+	ErrorExact       string    `json:"errorExact,omitempty"`
+	ErrorContains    []string  `json:"errorContains,omitempty"`
+	ErrorNotContains []string  `json:"errorNotContains,omitempty"`
+	NeedsLogin       bool      `json:"needsLogin,omitempty"`
+	CapturedAuth     string    `json:"capturedAuth,omitempty"`
+	CapturedEquals   string    `json:"capturedEquals,omitempty"`
 }
 
-// A public registry refuses the first read with a Bearer challenge and hands
-// out an anonymous token; tags come back a page at a time.
-func TestVersionsFromAnOCIRegistry(t *testing.T) {
-	var srv *httptest.Server
-	srv = httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		switch {
-		case r.URL.Path == tokenPath:
-			if r.URL.Query().Get("scope") != "repository:org/charts/sglang:pull" {
-				http.Error(w, "bad scope "+r.URL.Query().Get("scope"), http.StatusBadRequest)
-				return
-			}
-			_, _ = fmt.Fprint(w, `{"token":"t"}`)
-		case r.Header.Get("Authorization") != "Bearer t":
-			w.Header().Set("WWW-Authenticate", `Bearer realm="`+srv.URL+`/token",service="reg",scope="repository:org/charts/sglang:pull"`)
-			w.WriteHeader(http.StatusUnauthorized)
-		case r.URL.Path == "/v2/org/charts/sglang/tags/list" && r.URL.Query().Get("last") == "":
-			w.Header().Set("Link", `</v2/org/charts/sglang/tags/list?last=0.7.1&n=2>; rel="next"`)
-			_, _ = fmt.Fprint(w, `{"tags":["0.7.0","0.7.1"]}`)
-		case r.URL.Path == "/v2/org/charts/sglang/tags/list":
-			_, _ = fmt.Fprint(w, `{"tags":["0.8.0_build.1","sha256-abc.sig"]}`)
-		default:
-			http.NotFound(w, r)
-		}
-	}))
-	t.Cleanup(srv.Close)
-	saved := client
-	client = srv.Client()
-	t.Cleanup(func() { client = saved })
+// chartRoute is one httptest response. The first matching route wins.
+// A route with no path matches every path still unmatched.
+// No match answers 404 with body "404 page not found\n".
+// query selects the route. requireQuery on a selected route answers 400
+// "bad scope ..." when the query does not match, as the token endpoint does.
+// requireAuth skips the route when Authorization differs.
+// captureAuth records Authorization when the path and query match.
+type chartRoute struct {
+	Path         string            `json:"path"`
+	Query        map[string]string `json:"query,omitempty"`
+	Status       int               `json:"status,omitempty"`
+	Body         string            `json:"body,omitempty"`
+	Headers      map[string]string `json:"headers,omitempty"`
+	Challenge    string            `json:"challenge,omitempty"`
+	RealmPath    string            `json:"realmPath,omitempty"`
+	RealmName    string            `json:"realmName,omitempty"`
+	Service      string            `json:"service,omitempty"`
+	Scope        string            `json:"scope,omitempty"`
+	RequireAuth  string            `json:"requireAuth,omitempty"`
+	RequireQuery map[string]string `json:"requireQuery,omitempty"`
+	CaptureAuth  string            `json:"captureAuth,omitempty"`
+}
 
-	repo := "oci://" + strings.TrimPrefix(srv.URL, "https://") + "/org/charts"
-	got, err := Versions(t.Context(), repo, "", "sglang")
-	want := []string{"0.7.0", v071, "0.8.0+build.1", "sha256-abc.sig"}
-	if err != nil || !reflect.DeepEqual(got, want) {
-		t.Fatalf("got %v, %v; want %v", got, err, want)
-	}
+type repoCase struct {
+	Name          string       `json:"name,omitempty"`
+	Description   string       `json:"description,omitempty"`
+	TLS           bool         `json:"tls,omitempty"`
+	ReplaceClient bool         `json:"replaceClient,omitempty"`
+	Username      string       `json:"username,omitempty"`
+	Password      string       `json:"password,omitempty"`
+	Routes        []chartRoute `json:"routes,omitempty"`
+	Calls         []chartCall  `json:"calls"`
+}
+
+func TestVersionsFromARepoIndex(t *testing.T) {
+	cases.Run(t, "testdata/repo/versions-from-a-repo-index", runRepoCase)
+}
+
+func TestVersionsFromAnOCIRegistry(t *testing.T) {
+	cases.Run(t, "testdata/repo/versions-from-an-oci-registry", runRepoCase)
 }
 
 func TestVersionsFromALocalChart(t *testing.T) {
-	dir := t.TempDir()
-	if err := os.MkdirAll(filepath.Join(dir, "sglang"), 0o755); err != nil {
-		t.Fatal(err)
-	}
-	if err := os.WriteFile(filepath.Join(dir, "sglang", "Chart.yaml"), []byte("name: sglang\nversion: 0.7.4\n"), 0o644); err != nil {
-		t.Fatal(err)
-	}
-	got, err := Versions(t.Context(), "https://ignored.example.com", dir, "sglang")
-	if err != nil || !reflect.DeepEqual(got, []string{"0.7.4"}) {
-		t.Fatalf("got %v, %v; want the chart directory's own version", got, err)
-	}
+	cases.Run(t, "testdata/repo/versions-from-a-local-chart", runRepoCase)
 }
 
-// A private registry, or a chart that is not there: the token endpoint refuses,
-// and the error says so in words, with what the registry said.
 func TestARefusedReadSaysItNeedsALogin(t *testing.T) {
-	var srv *httptest.Server
-	srv = httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path == tokenPath {
-			w.WriteHeader(http.StatusForbidden)
-			_, _ = fmt.Fprint(w, `{"errors":[{"code":"DENIED","message":"requested access to the resource is denied"}]}`)
-			return
+	cases.Run(t, "testdata/repo/a-refused-read-says-it-needs-a-login", runRepoCase)
+}
+
+func TestARegistryAskingForBasicAuthNeedsALogin(t *testing.T) {
+	cases.Run(t, "testdata/repo/a-registry-asking-for-basic-auth-needs-a-login", runRepoCase)
+}
+
+func TestVersionsAuthSendsBasicAuth(t *testing.T) {
+	cases.Run(t, "testdata/repo/versions-auth-sends-basic-auth", runRepoCase)
+}
+
+func TestAMissingIndexIsReportedByStatus(t *testing.T) {
+	cases.Run(t, "testdata/repo/a-missing-index-is-reported-by-status", runRepoCase)
+}
+
+func runRepoCase(t *testing.T, tc repoCase) {
+	t.Helper()
+	if len(tc.Calls) == 0 {
+		t.Fatal("no calls")
+	}
+	if tc.ReplaceClient && !tc.TLS {
+		t.Fatal("replaceClient is only for a TLS server")
+	}
+	basic := ""
+	if tc.Username != "" || tc.Password != "" {
+		basic = basicHeader(tc.Username, tc.Password)
+	}
+	needsServer := false
+	for _, call := range tc.Calls {
+		if call.RepoKind == "https" || call.RepoKind == "oci" {
+			needsServer = true
 		}
-		w.Header().Set("WWW-Authenticate", `Bearer realm="`+srv.URL+`/token",service="reg",scope="repository:org/charts/sglang:pull"`)
-		w.WriteHeader(http.StatusUnauthorized)
-	}))
-	t.Cleanup(srv.Close)
+	}
+	var srv *httptest.Server
+	serverURL := ""
+	captured := map[string]string{}
 	saved := client
-	client = srv.Client()
-	t.Cleanup(func() { client = saved })
+	if needsServer {
+		handler := repoHandler(&serverURL, tc.Routes, captured, basic)
+		if tc.TLS {
+			srv = httptest.NewTLSServer(handler)
+		} else {
+			srv = httptest.NewServer(handler)
+		}
+		serverURL = srv.URL
+		t.Cleanup(srv.Close)
+		if tc.ReplaceClient {
+			client = srv.Client()
+			t.Cleanup(func() { client = saved })
+		}
+	}
+	for _, call := range tc.Calls {
+		runChartCall(t, srv, serverURL, basic, captured, call)
+	}
+	if !tc.ReplaceClient && client != saved {
+		t.Errorf("package client changed")
+	}
+}
 
-	repo := "oci://" + strings.TrimPrefix(srv.URL, "https://") + "/org/charts"
-	_, err := Versions(t.Context(), repo, "", "sglang")
-	var re *RepoError
-	if !errors.As(err, &re) || !re.NeedsLogin() {
-		t.Fatalf("got %v, want a RepoError needing a login", err)
+func runChartCall(t *testing.T, srv *httptest.Server, serverURL, basic string, captured map[string]string, call chartCall) {
+	t.Helper()
+	repo, path := call.source(t, srv)
+	indexURL := ""
+	if call.RepoKind == "https" {
+		indexURL = strings.TrimSuffix(repo, "/") + "/index.yaml"
 	}
-	for _, want := range []string{repo + "/sglang", "needs a login", "DENIED: requested access to the resource is denied"} {
-		if !strings.Contains(err.Error(), want) {
-			t.Errorf("%q does not say %q", err, want)
+	callBasic := basic
+	if call.Username != "" || call.Password != "" {
+		callBasic = basicHeader(call.Username, call.Password)
+	}
+	var (
+		got []string
+		err error
+	)
+	switch call.Fn {
+	case "versions":
+		if call.Username != "" || call.Password != "" {
+			t.Fatal("versions is anonymous; use versionsAuth")
 		}
+		got, err = Versions(t.Context(), repo, path, call.Chart)
+	case "versionsAuth":
+		got, err = VersionsAuth(t.Context(), repo, path, call.Chart, Credentials{
+			Username: call.Username,
+			Password: call.Password,
+		})
+	default:
+		t.Fatalf("fn %q", call.Fn)
 	}
-	if strings.Contains(err.Error(), tokenPath) {
-		t.Errorf("%q names the token endpoint rather than the repository", err)
+	switch {
+	case call.Versions != nil:
+		wantAuth, auth := "", ""
+		badAuth := false
+		if call.CapturedAuth != "" {
+			auth = captured[call.CapturedAuth]
+			wantAuth = subst(call.CapturedEquals, serverURL, repo, indexURL, callBasic)
+			badAuth = auth != wantAuth
+		}
+		if err != nil || !reflect.DeepEqual(got, *call.Versions) || badAuth {
+			t.Fatalf("got %v, %v; want %v; auth %q want %q", got, err, *call.Versions, auth, wantAuth)
+		}
+	case call.NeedsLogin || call.ErrorExact != "" || len(call.ErrorContains) > 0 || len(call.ErrorNotContains) > 0:
+		if call.NeedsLogin {
+			var re *RepoError
+			if err == nil || !errors.As(err, &re) || !re.NeedsLogin() {
+				t.Fatalf("got %v, want a RepoError needing a login", err)
+			}
+		}
+		if call.ErrorExact != "" {
+			want := subst(call.ErrorExact, serverURL, repo, indexURL, callBasic)
+			if err == nil || err.Error() != want {
+				t.Errorf("got %v", err)
+			}
+		}
+		for _, sub := range call.ErrorContains {
+			want := subst(sub, serverURL, repo, indexURL, callBasic)
+			if err == nil || !strings.Contains(err.Error(), want) {
+				t.Errorf("%q does not say %q", err, want)
+			}
+		}
+		for _, sub := range call.ErrorNotContains {
+			want := subst(sub, serverURL, repo, indexURL, callBasic)
+			if err != nil && strings.Contains(err.Error(), want) {
+				t.Errorf("%q names the token endpoint rather than the repository", err)
+			}
+		}
+	default:
+		t.Fatal("call has no expectation")
 	}
 }
 
-func TestARegistryAskingForBasicAuthNeedsALogin(t *testing.T) {
-	srv := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		w.Header().Set("WWW-Authenticate", `Basic realm="harbor"`)
-		w.WriteHeader(http.StatusUnauthorized)
-	}))
-	t.Cleanup(srv.Close)
-	saved := client
-	client = srv.Client()
-	t.Cleanup(func() { client = saved })
-
-	_, err := Versions(t.Context(), "oci://"+strings.TrimPrefix(srv.URL, "https://")+"/org", "", "sglang")
-	var re *RepoError
-	if !errors.As(err, &re) || !re.NeedsLogin() || !strings.Contains(err.Error(), "Basic auth") {
-		t.Fatalf("got %v", err)
+func (call chartCall) source(t *testing.T, srv *httptest.Server) (repo, path string) {
+	t.Helper()
+	switch call.RepoKind {
+	case "https":
+		if srv == nil {
+			t.Fatal("https repo needs a server")
+		}
+		return srv.URL + call.RepoPath, ""
+	case "oci":
+		if srv == nil {
+			t.Fatal("oci repo needs a server")
+		}
+		return "oci://" + strings.TrimPrefix(srv.URL, "https://") + call.RepoPath, ""
+	case "local":
+		if call.ChartYAML == "" {
+			t.Fatal("local chart needs chartYAML")
+		}
+		root := t.TempDir()
+		dir := filepath.Join(root, call.Chart)
+		if err := os.MkdirAll(dir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(filepath.Join(dir, "Chart.yaml"), []byte(call.ChartYAML), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		return call.LocalRepo, root
+	default:
+		t.Fatalf("repoKind %q", call.RepoKind)
+		return "", ""
 	}
 }
 
-func TestVersionsAuthSendsBasicAuth(t *testing.T) {
-	want := "Basic " + base64.StdEncoding.EncodeToString([]byte("user:s3cret"))
-	var tokenAuth string
-	var srv *httptest.Server
-	srv = httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		switch r.URL.Path {
-		case tokenPath:
-			tokenAuth = r.Header.Get("Authorization")
-			if tokenAuth != want {
-				http.Error(w, "no basic", http.StatusUnauthorized)
-				return
+func repoHandler(serverURL *string, routes []chartRoute, captured map[string]string, basic string) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		for i := range routes {
+			route := routes[i]
+			if route.Path != "" && r.URL.Path != route.Path {
+				continue
 			}
-			_, _ = fmt.Fprint(w, `{"token":"t"}`)
-		case "/v2/org/sglang/tags/list":
-			if r.Header.Get("Authorization") != "Bearer t" {
-				w.Header().Set("WWW-Authenticate", `Bearer realm="`+srv.URL+`/token",service="reg",scope="repository:org/sglang:pull"`)
-				w.WriteHeader(http.StatusUnauthorized)
-				return
+			if !queryMatch(r, route.Query) {
+				continue
 			}
-			_, _ = fmt.Fprint(w, `{"tags":["0.8.0"]}`)
-		case "/charts/index.yaml":
-			if r.Header.Get("Authorization") != want {
-				http.Error(w, "no basic", http.StatusUnauthorized)
+			if len(route.RequireQuery) > 0 && !queryMatch(r, route.RequireQuery) {
+				http.Error(w, "bad scope "+r.URL.Query().Get("scope"), http.StatusBadRequest)
 				return
 			}
-			_, _ = fmt.Fprint(w, "apiVersion: v1\nentries:\n  sglang:\n    - version: 0.8.0\n")
-		default:
-			http.NotFound(w, r)
+			if route.CaptureAuth != "" {
+				captured[route.CaptureAuth] = r.Header.Get("Authorization")
+			}
+			wantAuth := subst(route.RequireAuth, *serverURL, "", "", basic)
+			if route.RequireAuth != "" && r.Header.Get("Authorization") != wantAuth {
+				continue
+			}
+			writeRoute(w, route, *serverURL)
+			return
 		}
-	}))
-	t.Cleanup(srv.Close)
-	saved := client
-	client = srv.Client()
-	t.Cleanup(func() { client = saved })
+		http.NotFound(w, r)
+	})
+}
 
-	creds := Credentials{Username: "user", Password: "s3cret"}
-	idx, err := VersionsAuth(t.Context(), "https://"+strings.TrimPrefix(srv.URL, "https://")+"/charts", "", "sglang", creds)
-	if err != nil || !reflect.DeepEqual(idx, []string{v080}) {
-		t.Fatalf("index: got %v, %v", idx, err)
+func writeRoute(w http.ResponseWriter, route chartRoute, serverURL string) {
+	if route.Challenge != "" {
+		w.Header().Set("WWW-Authenticate", challengeHeader(route, serverURL))
+	}
+	for k, v := range route.Headers {
+		w.Header().Set(k, v)
 	}
-	repo := "oci://" + strings.TrimPrefix(srv.URL, "https://") + "/org"
-	got, err := VersionsAuth(t.Context(), repo, "", "sglang", creds)
-	if err != nil || !reflect.DeepEqual(got, []string{v080}) || tokenAuth != want {
-		t.Fatalf("oci: got %v, %v; token auth %q", got, err, tokenAuth)
+	status := route.Status
+	if status == 0 {
+		if route.Challenge != "" {
+			status = http.StatusUnauthorized
+		} else {
+			status = http.StatusOK
+		}
+	}
+	w.WriteHeader(status)
+	if route.Body != "" {
+		_, _ = io.WriteString(w, route.Body)
 	}
 }
 
-func TestAMissingIndexIsReportedByStatus(t *testing.T) {
-	srv := httptest.NewServer(http.NotFoundHandler())
-	t.Cleanup(srv.Close)
-	_, err := Versions(t.Context(), srv.URL+"/charts", "", "sglang")
-	if err == nil || err.Error() != srv.URL+"/charts/index.yaml: 404 Not Found -- the repository says: 404 page not found" {
-		t.Errorf("got %v", err)
+func challengeHeader(route chartRoute, serverURL string) string {
+	switch route.Challenge {
+	case "bearer":
+		return fmt.Sprintf(`Bearer realm="%s%s",service="%s",scope="%s"`, serverURL, route.RealmPath, route.Service, route.Scope)
+	case "basic":
+		return fmt.Sprintf(`Basic realm="%s"`, route.RealmName)
+	default:
+		return route.Challenge
 	}
 }
+
+func queryMatch(r *http.Request, q map[string]string) bool {
+	for k, v := range q {
+		if r.URL.Query().Get(k) != v {
+			return false
+		}
+	}
+	return true
+}
+
+func basicHeader(user, pass string) string {
+	return "Basic " + base64.StdEncoding.EncodeToString([]byte(user+":"+pass))
+}
+
+func subst(s, server, repo, indexURL, basic string) string {
+	return strings.NewReplacer(
+		"{server}", server,
+		"{repo}", repo,
+		"{indexURL}", indexURL,
+		"{basic}", basic,
+	).Replace(s)
+}
diff --git a/internal/chart/testdata/repo/a-missing-index-is-reported-by-status/a-missing-index-is-reported-by-status.yaml b/internal/chart/testdata/repo/a-missing-index-is-reported-by-status/a-missing-index-is-reported-by-status.yaml
new file mode 100644
index 0000000..f3970e3
--- /dev/null
+++ b/internal/chart/testdata/repo/a-missing-index-is-reported-by-status/a-missing-index-is-reported-by-status.yaml
@@ -0,0 +1,7 @@
+description: missing index is reported by status
+calls:
+  - fn: versions
+    repoKind: https
+    repoPath: /charts
+    chart: sglang
+    errorExact: "{indexURL}: 404 Not Found -- the repository says: 404 page not found"
diff --git a/internal/chart/testdata/repo/a-refused-read-says-it-needs-a-login/a-refused-read-says-it-needs-a-login.yaml b/internal/chart/testdata/repo/a-refused-read-says-it-needs-a-login/a-refused-read-says-it-needs-a-login.yaml
new file mode 100644
index 0000000..a57557f
--- /dev/null
+++ b/internal/chart/testdata/repo/a-refused-read-says-it-needs-a-login/a-refused-read-says-it-needs-a-login.yaml
@@ -0,0 +1,23 @@
+description: refused read needs a login and does not name the token path
+tls: true
+replaceClient: true
+routes:
+  - path: /token
+    status: 403
+    body: '{"errors":[{"code":"DENIED","message":"requested access to the resource is denied"}]}'
+  - challenge: bearer
+    realmPath: /token
+    service: reg
+    scope: "repository:org/charts/sglang:pull"
+calls:
+  - fn: versions
+    repoKind: oci
+    repoPath: /org/charts
+    chart: sglang
+    needsLogin: true
+    errorContains:
+      - "{repo}/sglang"
+      - "needs a login"
+      - "DENIED: requested access to the resource is denied"
+    errorNotContains:
+      - /token
diff --git a/internal/chart/testdata/repo/a-registry-asking-for-basic-auth-needs-a-login/a-registry-asking-for-basic-auth-needs-a-login.yaml b/internal/chart/testdata/repo/a-registry-asking-for-basic-auth-needs-a-login/a-registry-asking-for-basic-auth-needs-a-login.yaml
new file mode 100644
index 0000000..3587ce9
--- /dev/null
+++ b/internal/chart/testdata/repo/a-registry-asking-for-basic-auth-needs-a-login/a-registry-asking-for-basic-auth-needs-a-login.yaml
@@ -0,0 +1,14 @@
+description: basic challenge needs a login
+tls: true
+replaceClient: true
+routes:
+  - challenge: basic
+    realmName: harbor
+calls:
+  - fn: versions
+    repoKind: oci
+    repoPath: /org
+    chart: sglang
+    needsLogin: true
+    errorContains:
+      - "Basic auth"
diff --git a/internal/chart/testdata/repo/versions-auth-sends-basic-auth/versions-auth-sends-basic-auth.yaml b/internal/chart/testdata/repo/versions-auth-sends-basic-auth/versions-auth-sends-basic-auth.yaml
new file mode 100644
index 0000000..d892728
--- /dev/null
+++ b/internal/chart/testdata/repo/versions-auth-sends-basic-auth/versions-auth-sends-basic-auth.yaml
@@ -0,0 +1,50 @@
+description: index and oci send basic auth
+tls: true
+replaceClient: true
+username: user
+password: s3cret
+routes:
+  - path: /token
+    requireAuth: "{basic}"
+    captureAuth: token
+    body: '{"token":"t"}'
+  - path: /token
+    status: 401
+    body: "no basic\n"
+  - path: /v2/org/sglang/tags/list
+    requireAuth: "Bearer t"
+    body: '{"tags":["0.8.0"]}'
+  - path: /v2/org/sglang/tags/list
+    challenge: bearer
+    realmPath: /token
+    service: reg
+    scope: "repository:org/sglang:pull"
+  - path: /charts/index.yaml
+    requireAuth: "{basic}"
+    body: |
+      apiVersion: v1
+      entries:
+        sglang:
+          - version: 0.8.0
+  - path: /charts/index.yaml
+    status: 401
+    body: "no basic\n"
+calls:
+  - fn: versionsAuth
+    repoKind: https
+    repoPath: /charts
+    chart: sglang
+    username: user
+    password: s3cret
+    versions:
+      - "0.8.0"
+  - fn: versionsAuth
+    repoKind: oci
+    repoPath: /org
+    chart: sglang
+    username: user
+    password: s3cret
+    versions:
+      - "0.8.0"
+    capturedAuth: token
+    capturedEquals: "{basic}"
diff --git a/internal/chart/testdata/repo/versions-from-a-local-chart/versions-from-a-local-chart.yaml b/internal/chart/testdata/repo/versions-from-a-local-chart/versions-from-a-local-chart.yaml
new file mode 100644
index 0000000..16ad794
--- /dev/null
+++ b/internal/chart/testdata/repo/versions-from-a-local-chart/versions-from-a-local-chart.yaml
@@ -0,0 +1,11 @@
+description: local Chart.yaml version, repo URL ignored
+calls:
+  - fn: versions
+    repoKind: local
+    localRepo: https://ignored.example.com
+    chart: sglang
+    chartYAML: |
+      name: sglang
+      version: 0.7.4
+    versions:
+      - "0.7.4"
diff --git a/internal/chart/testdata/repo/versions-from-a-repo-index/versions-from-a-repo-index.yaml b/internal/chart/testdata/repo/versions-from-a-repo-index/versions-from-a-repo-index.yaml
new file mode 100644
index 0000000..f895ac6
--- /dev/null
+++ b/internal/chart/testdata/repo/versions-from-a-repo-index/versions-from-a-repo-index.yaml
@@ -0,0 +1,25 @@
+description: index lists sglang and reports a missing chart
+routes:
+  - path: /charts/index.yaml
+    body: |
+      apiVersion: v1
+      entries:
+        sglang:
+          - version: 0.8.0
+          - version: 0.7.1
+        vllm:
+          - version: 1.0.0
+calls:
+  - fn: versions
+    repoKind: https
+    repoPath: /charts/
+    chart: sglang
+    versions:
+      - "0.8.0"
+      - "0.7.1"
+  - fn: versions
+    repoKind: https
+    repoPath: /charts
+    chart: nope
+    errorContains:
+      - 'no chart "nope"'
diff --git a/internal/chart/testdata/repo/versions-from-an-oci-registry/versions-from-an-oci-registry.yaml b/internal/chart/testdata/repo/versions-from-an-oci-registry/versions-from-an-oci-registry.yaml
new file mode 100644
index 0000000..171b44f
--- /dev/null
+++ b/internal/chart/testdata/repo/versions-from-an-oci-registry/versions-from-an-oci-registry.yaml
@@ -0,0 +1,32 @@
+description: bearer challenge, anonymous token, and paginated tags
+tls: true
+replaceClient: true
+routes:
+  - path: /token
+    requireQuery:
+      scope: "repository:org/charts/sglang:pull"
+    body: '{"token":"t"}'
+  - path: /v2/org/charts/sglang/tags/list
+    query:
+      last: ""
+    requireAuth: "Bearer t"
+    headers:
+      Link: '</v2/org/charts/sglang/tags/list?last=0.7.1&n=2>; rel="next"'
+    body: '{"tags":["0.7.0","0.7.1"]}'
+  - path: /v2/org/charts/sglang/tags/list
+    requireAuth: "Bearer t"
+    body: '{"tags":["0.8.0_build.1","sha256-abc.sig"]}'
+  - challenge: bearer
+    realmPath: /token
+    service: reg
+    scope: "repository:org/charts/sglang:pull"
+calls:
+  - fn: versions
+    repoKind: oci
+    repoPath: /org/charts
+    chart: sglang
+    versions:
+      - "0.7.0"
+      - "0.7.1"
+      - "0.8.0+build.1"
+      - sha256-abc.sig
diff --git a/internal/chart/testdata/version/bare/exact.yaml b/internal/chart/testdata/version/bare/exact.yaml
new file mode 100644
index 0000000..4297528
--- /dev/null
+++ b/internal/chart/testdata/version/bare/exact.yaml
@@ -0,0 +1,7 @@
+description: a bare version allows only itself
+spec: "0.7.1"
+exact: "0.7.1"
+allows:
+  - "0.7.1"
+refuses:
+  - "0.7.2"
diff --git a/internal/chart/testdata/version/bare/range.yaml b/internal/chart/testdata/version/bare/range.yaml
new file mode 100644
index 0000000..bc7a98f
--- /dev/null
+++ b/internal/chart/testdata/version/bare/range.yaml
@@ -0,0 +1,3 @@
+description: a range names no one version
+spec: ">=0.7.1"
+exact: ""
diff --git a/internal/chart/testdata/version/match/caret.yaml b/internal/chart/testdata/version/match/caret.yaml
new file mode 100644
index 0000000..4bb1976
--- /dev/null
+++ b/internal/chart/testdata/version/match/caret.yaml
@@ -0,0 +1,9 @@
+description: "^0.7.1 want 0.7.x only"
+spec: "^0.7.1"
+have:
+  - "0.7.1"
+  - "0.7.9"
+  - "0.8.0"
+want:
+  - "0.7.9"
+  - "0.7.1"
diff --git a/internal/chart/testdata/version/match/newest-first.yaml b/internal/chart/testdata/version/match/newest-first.yaml
new file mode 100644
index 0000000..19d3bfe
--- /dev/null
+++ b/internal/chart/testdata/version/match/newest-first.yaml
@@ -0,0 +1,14 @@
+description: newest first, no prerelease, no duplicates
+spec: ">=0.7.1"
+have:
+  - "0.7.0"
+  - "0.7.1"
+  - "0.10.0"
+  - latest
+  - "0.8.0"
+  - "0.9.0-rc1"
+  - "0.8.0"
+want:
+  - "0.10.0"
+  - "0.8.0"
+  - "0.7.1"
diff --git a/internal/chart/testdata/version/parse/banana.yaml b/internal/chart/testdata/version/parse/banana.yaml
new file mode 100644
index 0000000..2f3012e
--- /dev/null
+++ b/internal/chart/testdata/version/parse/banana.yaml
@@ -0,0 +1,2 @@
+description: ">=banana"
+spec: ">=banana"
diff --git a/internal/chart/testdata/version/parse/empty.yaml b/internal/chart/testdata/version/parse/empty.yaml
new file mode 100644
index 0000000..aea354b
--- /dev/null
+++ b/internal/chart/testdata/version/parse/empty.yaml
@@ -0,0 +1,2 @@
+description: empty
+spec: ""
diff --git a/internal/chart/testdata/version/parse/latest.yaml b/internal/chart/testdata/version/parse/latest.yaml
new file mode 100644
index 0000000..dde4415
--- /dev/null
+++ b/internal/chart/testdata/version/parse/latest.yaml
@@ -0,0 +1,2 @@
+description: latest
+spec: latest
diff --git a/internal/chart/testdata/version/resolve-refuses/named-out-of-range.yaml b/internal/chart/testdata/version/resolve-refuses/named-out-of-range.yaml
new file mode 100644
index 0000000..ebba9a5
--- /dev/null
+++ b/internal/chart/testdata/version/resolve-refuses/named-out-of-range.yaml
@@ -0,0 +1,5 @@
+description: a named version out of range
+spec: ">=0.7.1"
+want: "0.6.0"
+noList: true
+errorContains: "not in the catalog's"
diff --git a/internal/chart/testdata/version/resolve-refuses/nil-lister.yaml b/internal/chart/testdata/version/resolve-refuses/nil-lister.yaml
new file mode 100644
index 0000000..86dfdca
--- /dev/null
+++ b/internal/chart/testdata/version/resolve-refuses/nil-lister.yaml
@@ -0,0 +1,4 @@
+description: a range with nowhere to list
+spec: ">=0.7.1"
+listNil: true
+errorContains: "Name a chart version"
diff --git a/internal/chart/testdata/version/resolve-refuses/nothing-in-range.yaml b/internal/chart/testdata/version/resolve-refuses/nothing-in-range.yaml
new file mode 100644
index 0000000..23f9164
--- /dev/null
+++ b/internal/chart/testdata/version/resolve-refuses/nothing-in-range.yaml
@@ -0,0 +1,5 @@
+description: nothing in range
+spec: ">=0.7.1"
+list:
+  - "0.6.0"
+errorContains: "no chart version satisfies"
diff --git a/internal/chart/testdata/version/resolve/exact-moves-running-off.yaml b/internal/chart/testdata/version/resolve/exact-moves-running-off.yaml
new file mode 100644
index 0000000..752ef0c
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/exact-moves-running-off.yaml
@@ -0,0 +1,5 @@
+description: exact moves a running one off
+spec: "0.7.2"
+keep: "0.7.1"
+got: "0.7.2"
+noList: true
diff --git a/internal/chart/testdata/version/resolve/exact-needs-no-registry.yaml b/internal/chart/testdata/version/resolve/exact-needs-no-registry.yaml
new file mode 100644
index 0000000..b23c5a6
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/exact-needs-no-registry.yaml
@@ -0,0 +1,4 @@
+description: exact needs no registry
+spec: "0.7.1"
+got: "0.7.1"
+noList: true
diff --git a/internal/chart/testdata/version/resolve/named-in-range.yaml b/internal/chart/testdata/version/resolve/named-in-range.yaml
new file mode 100644
index 0000000..0be71f4
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/named-in-range.yaml
@@ -0,0 +1,5 @@
+description: named, in range
+spec: ">=0.7.1"
+want: "0.7.3"
+got: "0.7.3"
+noList: true
diff --git a/internal/chart/testdata/version/resolve/newest-in-range.yaml b/internal/chart/testdata/version/resolve/newest-in-range.yaml
new file mode 100644
index 0000000..37c6763
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/newest-in-range.yaml
@@ -0,0 +1,7 @@
+description: newest in range
+spec: ">=0.7.1 <0.9.0"
+got: "0.8.1"
+list:
+  - "0.7.1"
+  - "0.8.1"
+  - "0.9.0"
diff --git a/internal/chart/testdata/version/resolve/running-kept-in-range.yaml b/internal/chart/testdata/version/resolve/running-kept-in-range.yaml
new file mode 100644
index 0000000..477ea4e
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/running-kept-in-range.yaml
@@ -0,0 +1,5 @@
+description: running one kept while in range
+spec: ">=0.7.1"
+keep: "0.7.2"
+got: "0.7.2"
+noList: true
diff --git a/internal/chart/testdata/version/resolve/running-out-of-range.yaml b/internal/chart/testdata/version/resolve/running-out-of-range.yaml
new file mode 100644
index 0000000..a7f015f
--- /dev/null
+++ b/internal/chart/testdata/version/resolve/running-out-of-range.yaml
@@ -0,0 +1,8 @@
+description: running one out of range
+spec: ">=0.8.0"
+keep: "0.7.2"
+got: "0.9.0"
+list:
+  - "0.7.2"
+  - "0.8.0"
+  - "0.9.0"
diff --git a/internal/chart/version_test.go b/internal/chart/version_test.go
index 1892948..b50c627 100644
--- a/internal/chart/version_test.go
+++ b/internal/chart/version_test.go
@@ -6,13 +6,8 @@ import (
 	"reflect"
 	"strings"
 	"testing"
-)
 
-const (
-	v071      = "0.7.1"
-	v072      = "0.7.2"
-	v080      = "0.8.0"
-	tokenPath = "/token"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
 func spec(t *testing.T, s string) Spec {
@@ -24,79 +19,145 @@ func spec(t *testing.T, s string) Spec {
 	return sp
 }
 
-// A bare version is that version and nothing else, as before ranges.
+// bareCase is one exact-version check. Exact is a pointer so a range can
+// expect the empty string and an omitted field stays distinct from that.
+type bareCase struct {
+	Name        string   `json:"name,omitempty"`
+	Description string   `json:"description,omitempty"`
+	Spec        string   `json:"spec"`
+	Exact       *string  `json:"exact,omitempty"`
+	Allows      []string `json:"allows,omitempty"`
+	Refuses     []string `json:"refuses,omitempty"`
+}
+
 func TestABareVersionIsExact(t *testing.T) {
-	sp := spec(t, v071)
-	if sp.Exact() != v071 {
-		t.Errorf("Exact = %q, want 0.7.1", sp.Exact())
-	}
-	if sp.Allows(v072) || !sp.Allows(v071) {
-		t.Error("0.7.1 must allow 0.7.1 and only it")
-	}
-	if spec(t, ">=0.7.1").Exact() != "" {
-		t.Error("a range names no one version")
-	}
+	cases.Run(t, "testdata/version/bare", func(t *testing.T, tc bareCase) {
+		sp := spec(t, tc.Spec)
+		if tc.Exact == nil {
+			t.Fatal("exact is omitted; set it to the one version or to an empty string")
+		}
+		if sp.Exact() != *tc.Exact {
+			if *tc.Exact == "" {
+				t.Error("a range names no one version")
+			} else {
+				t.Errorf("Exact = %q, want %s", sp.Exact(), *tc.Exact)
+			}
+		}
+		bad := false
+		for _, v := range tc.Allows {
+			if !sp.Allows(v) {
+				bad = true
+			}
+		}
+		for _, v := range tc.Refuses {
+			if sp.Allows(v) {
+				bad = true
+			}
+		}
+		if bad {
+			t.Errorf("%s must allow %v and not %v", tc.Spec, tc.Allows, tc.Refuses)
+		}
+	})
+}
+
+type parseCase struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Spec        string `json:"spec"`
 }
 
 func TestParseSpecRefusesWhatIsNotAVersion(t *testing.T) {
-	for _, s := range []string{"", "latest", ">=banana"} {
-		if _, err := ParseSpec(s); err == nil {
-			t.Errorf("%q: want an error", s)
+	cases.Run(t, "testdata/version/parse", func(t *testing.T, tc parseCase) {
+		if _, err := ParseSpec(tc.Spec); err == nil {
+			t.Errorf("%q: want an error", tc.Spec)
 		}
-	}
+	})
 }
 
-func TestMatchIsNewestFirstAndSkipsNonVersions(t *testing.T) {
-	got := spec(t, ">=0.7.1").Match([]string{"0.7.0", v071, "0.10.0", "latest", v080, "0.9.0-rc1", v080})
-	want := []string{"0.10.0", v080, v071}
-	if !reflect.DeepEqual(got, want) {
-		t.Errorf("got %v, want %v -- newest first, no prerelease, no duplicates", got, want)
-	}
-	if got := spec(t, "^0.7.1").Match([]string{v071, "0.7.9", v080}); !reflect.DeepEqual(got, []string{"0.7.9", v071}) {
-		t.Errorf("^0.7.1 matched %v, want 0.7.x only", got)
-	}
+type matchCase struct {
+	Name        string   `json:"name,omitempty"`
+	Description string   `json:"description,omitempty"`
+	Spec        string   `json:"spec"`
+	Have        []string `json:"have"`
+	Want        []string `json:"want"`
 }
 
-func list(vs ...string) Lister {
-	return func(context.Context) ([]string, error) { return vs, nil }
+func TestMatchIsNewestFirstAndSkipsNonVersions(t *testing.T) {
+	cases.Run(t, "testdata/version/match", func(t *testing.T, tc matchCase) {
+		got := spec(t, tc.Spec).Match(tc.Have)
+		if !reflect.DeepEqual(got, tc.Want) {
+			t.Errorf("got %v, want %v -- %s", got, tc.Want, tc.Description)
+		}
+	})
 }
 
-func noList(t *testing.T) Lister {
-	return func(context.Context) ([]string, error) {
-		t.Error("the repository must not be read")
-		return nil, errors.New("read")
-	}
+// resolveCase is one Resolve call.
+// Exactly one of noList, listNil, and list is set.
+// noList fails the test if the repository is read.
+// listNil passes a nil Lister.
+// list, including an empty array, is returned as the catalog.
+type resolveCase struct {
+	Name          string    `json:"name,omitempty"`
+	Description   string    `json:"description,omitempty"`
+	Spec          string    `json:"spec"`
+	Want          string    `json:"want,omitempty"`
+	Keep          string    `json:"keep,omitempty"`
+	Got           string    `json:"got,omitempty"`
+	NoList        bool      `json:"noList,omitempty"`
+	ListNil       bool      `json:"listNil,omitempty"`
+	List          *[]string `json:"list,omitempty"`
+	ErrorContains string    `json:"errorContains,omitempty"`
 }
 
-func TestResolve(t *testing.T) {
-	ctx := t.Context()
-	for _, tc := range []struct {
-		name, spec, want, keep, got string
-		list                        Lister
-	}{
-		{name: "named, in range", spec: ">=0.7.1", want: "0.7.3", got: "0.7.3", list: noList(t)},
-		{name: "exact needs no registry", spec: v071, got: v071, list: noList(t)},
-		{name: "exact moves a running one off", spec: v072, keep: v071, got: v072, list: noList(t)},
-		{name: "running one kept while in range", spec: ">=0.7.1", keep: v072, got: v072, list: noList(t)},
-		{name: "running one out of range", spec: ">=0.8.0", keep: v072, got: "0.9.0", list: list(v072, v080, "0.9.0")},
-		{name: "newest in range", spec: ">=0.7.1 <0.9.0", got: "0.8.1", list: list(v071, "0.8.1", "0.9.0")},
-	} {
-		got, err := Resolve(ctx, spec(t, tc.spec), tc.want, tc.keep, tc.list)
-		if err != nil || got != tc.got {
-			t.Errorf("%s: got %q, %v; want %q", tc.name, got, err, tc.got)
+func (tc resolveCase) lister(t *testing.T) Lister {
+	t.Helper()
+	n := 0
+	if tc.NoList {
+		n++
+	}
+	if tc.ListNil {
+		n++
+	}
+	if tc.List != nil {
+		n++
+	}
+	if n != 1 {
+		t.Fatal("set exactly one of noList, listNil, and list")
+	}
+	if tc.NoList {
+		return func(context.Context) ([]string, error) {
+			t.Error("the repository must not be read")
+			return nil, errors.New("read")
 		}
 	}
+	if tc.ListNil {
+		return nil
+	}
+	vs := *tc.List
+	return func(context.Context) ([]string, error) { return vs, nil }
 }
 
-func TestResolveRefuses(t *testing.T) {
-	ctx := t.Context()
-	if _, err := Resolve(ctx, spec(t, ">=0.7.1"), "0.6.0", "", noList(t)); err == nil || !strings.Contains(err.Error(), "not in the catalog's") {
-		t.Errorf("a named version out of range: %v", err)
+func runResolve(t *testing.T, tc resolveCase) {
+	t.Helper()
+	if tc.ErrorContains == "" && tc.Got == "" {
+		t.Fatal("set got or errorContains")
 	}
-	if _, err := Resolve(ctx, spec(t, ">=0.7.1"), "", "", list("0.6.0")); err == nil || !strings.Contains(err.Error(), "no chart version satisfies") {
-		t.Errorf("nothing in range: %v", err)
+	got, err := Resolve(t.Context(), spec(t, tc.Spec), tc.Want, tc.Keep, tc.lister(t))
+	if tc.ErrorContains != "" {
+		if err == nil || !strings.Contains(err.Error(), tc.ErrorContains) {
+			t.Errorf("%s: got %q, %v; want an error with %q", tc.Description, got, err, tc.ErrorContains)
+		}
+		return
 	}
-	if _, err := Resolve(ctx, spec(t, ">=0.7.1"), "", "", nil); err == nil || !strings.Contains(err.Error(), "Name a chart version") {
-		t.Errorf("a range with nowhere to list: %v", err)
+	if err != nil || got != tc.Got {
+		t.Errorf("%s: got %q, %v; want %q", tc.Description, got, err, tc.Got)
 	}
 }
+
+func TestResolve(t *testing.T) {
+	cases.Run(t, "testdata/version/resolve", runResolve)
+}
+
+func TestResolveRefuses(t *testing.T) {
+	cases.Run(t, "testdata/version/resolve-refuses", runResolve)
+}
diff --git a/internal/controller/autoscaling/customprovider_test.go b/internal/controller/autoscaling/customprovider_test.go
index 386d0cd..5b93e59 100644
--- a/internal/controller/autoscaling/customprovider_test.go
+++ b/internal/controller/autoscaling/customprovider_test.go
@@ -27,237 +27,205 @@ import (
 	"k8s.io/apimachinery/pkg/types"
 
 	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
 const (
 	// sampleServiceID is the serviceId from the decision server's own example.
 	sampleServiceID = "fallback-modelforge-01"
-
-	// otherNamespace stands in for a second namespace reporting the same
-	// serviceId, which is the case customProvider.namespace exists to resolve.
-	otherNamespace = "staging"
 )
 
-// decisionBody wraps decisions in the envelope the server sends, so each test
-// only has to spell out the part it is about. Format verbs in decisions are
-// filled from args.
-func decisionBody(decisions string, args ...any) string {
-	return fmt.Sprintf(`{"apiVersion":%q,"decisions":[%s]}`,
-		decisionsAPIVersion, fmt.Sprintf(decisions, args...))
+type decisionFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		ServiceID     string `json:"serviceId"`
+		Namespace     string `json:"namespace"`
+		Body          string `json:"body"`
+		Status        int    `json:"status"`
+		PinAPIVersion bool   `json:"pinAPIVersion"`
+	} `json:"inputs"`
+	Expected struct {
+		Replicas      int32  `json:"replicas"`
+		ErrorContains string `json:"errorContains"`
+	} `json:"expected"`
 }
 
-// TestQueryDecisionReplicas covers how a decision payload is resolved to a
-// replica count: the schema must be one we know, exactly one decision must
-// address us, and the count must be present and sane. Everything else is an
-// error the caller skips the sync on rather than a zero.
 func TestQueryDecisionReplicas(t *testing.T) {
-	tests := []struct {
-		name      string
-		namespace string
-		body      string
-		want      int32
-		wantErr   string
-	}{
-		{
-			name: "a matching decision resolves to its active count",
-			body: decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{"active":15}}`, decisionNamespace),
-			want: 15,
-		},
-		{
-			name: "zero is a decision, not a missing one",
-			body: decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{"active":0}}`, decisionNamespace),
-			want: 0,
-		},
-		{
-			name:    "a decision with no active count is not a zero",
-			body:    decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{}}`, decisionNamespace),
-			wantErr: "carries no replicas.active",
-		},
-		{
-			name:    "a negative count is refused",
-			body:    decisionBody(`{"serviceId":"svc","replicas":{"active":-1}}`),
-			wantErr: "out of range",
-		},
-		{
-			name:    "a count past int32 is refused rather than wrapped",
-			body:    decisionBody(`{"serviceId":"svc","replicas":{"active":4294967296}}`),
-			wantErr: "out of range",
-		},
-		{
-			name:    "an unknown schema is refused, not parsed hopefully",
-			body:    `{"apiVersion":"llmscaling.inference.x-k8s.io/v2","decisions":[{"serviceId":"svc","replicas":{"active":3}}]}`,
-			wantErr: "unsupported decision schema",
-		},
-		{
-			name:    "a payload with no apiVersion is refused",
-			body:    `{"decisions":[{"serviceId":"svc","replicas":{"active":3}}]}`,
-			wantErr: "unsupported decision schema",
-		},
-		{
-			name:    "an empty decision list is not a zero",
-			body:    decisionBody(``),
-			wantErr: `no decision for serviceId "svc"`,
-		},
-		{
-			// The server answers with what it likes; a decision for someone else
-			// must not drive this scaler just because it came back.
-			name:    "a decision for another service does not match",
-			body:    decisionBody(`{"namespace":%q,"serviceId":"other","replicas":{"active":9}}`, decisionNamespace),
-			wantErr: `no decision for serviceId "svc"`,
-		},
-		{
-			name: "the namespace filter picks between same-named services",
-			body: decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{"active":2}},
-			                    {"namespace":%q,"serviceId":"svc","replicas":{"active":7}}`,
-				otherNamespace, decisionNamespace),
-			namespace: decisionNamespace,
-			want:      7,
-		},
-		{
-			name: "duplicates are an error, not a pick-first",
-			body: decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{"active":2}},
-			                    {"namespace":%q,"serviceId":"svc","replicas":{"active":7}}`,
-				otherNamespace, decisionNamespace),
-			wantErr: "2 decisions match",
-		},
-		{
-			name:      "the namespace filter is reported when nothing matches",
-			body:      decisionBody(`{"namespace":%q,"serviceId":"svc","replicas":{"active":2}}`, otherNamespace),
-			namespace: decisionNamespace,
-			wantErr:   fmt.Sprintf("in namespace %q", decisionNamespace),
-		},
-		{
-			name:    "a body that is not JSON is an error",
-			body:    `<html>502 Bad Gateway</html>`,
-			wantErr: "failed to decode",
-		},
-	}
-
-	for _, tt := range tests {
-		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-			_, _ = w.Write([]byte(tt.body))
-		}))
-
-		got, err := queryDecisionReplicas(srv.URL, "", "svc", tt.namespace, nil)
-		switch {
-		case tt.wantErr == "" && err != nil:
-			t.Errorf("%s: unexpected error: %v", tt.name, err)
-		case tt.wantErr == "" && got != tt.want:
-			t.Errorf("%s: got %d, want %d", tt.name, got, tt.want)
-		case tt.wantErr != "" && err == nil:
-			t.Errorf("%s: expected an error containing %q, got value %d", tt.name, tt.wantErr, got)
-		case tt.wantErr != "" && !strings.Contains(err.Error(), tt.wantErr):
-			t.Errorf("%s: error = %q, want it to contain %q", tt.name, err, tt.wantErr)
-		}
-		srv.Close()
-	}
+	cases.Run(t, "testdata/helpers/decision-replicas", func(t *testing.T, tc decisionFile) {
+		assertDecision(t, tc)
+	})
 }
 
 // TestQueryDecisionReplicasSamplePayload runs a real reply from the decision
 // server through the client verbatim, so the parser is pinned against the wire
 // format and not only against the fixtures the table above builds.
 func TestQueryDecisionReplicasSamplePayload(t *testing.T) {
+	cases.Run(t, "testdata/helpers/decision-sample", func(t *testing.T, tc decisionFile) {
+		if tc.Inputs.PinAPIVersion && !strings.Contains(tc.Inputs.Body, decisionsAPIVersion) {
+			t.Fatalf("sample body must carry apiVersion %s", decisionsAPIVersion)
+		}
+		assertDecision(t, tc)
+	})
+}
+
+func assertDecision(t *testing.T, tc decisionFile) {
+	t.Helper()
+	status := tc.Inputs.Status
+	if status == 0 {
+		status = http.StatusOK
+	}
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-		_, _ = w.Write([]byte(`{
-  "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1",
-  "decisions": [
-    {
-      "namespace": "modelforge",
-      "serviceId": "fallback-modelforge-01",
-      "replicas": {
-        "active": 15
-      }
-    }
-  ]
-}`))
+		w.WriteHeader(status)
+		_, _ = w.Write([]byte(tc.Inputs.Body))
 	}))
 	defer srv.Close()
 
-	got, err := queryDecisionReplicas(srv.URL, "", sampleServiceID, "", nil)
-	if err != nil {
-		t.Fatalf("unexpected error: %v", err)
+	got, err := queryDecisionReplicas(srv.URL, "", tc.Inputs.ServiceID, tc.Inputs.Namespace, nil)
+	if tc.Expected.ErrorContains == "" {
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if got != tc.Expected.Replicas {
+			t.Fatalf("got %d, want %d", got, tc.Expected.Replicas)
+		}
+		return
 	}
-	if got != 15 {
-		t.Errorf("got %d replicas, want 15", got)
+	if err == nil || !strings.Contains(err.Error(), tc.Expected.ErrorContains) {
+		t.Fatalf("error = %v, want it to contain %q", err, tc.Expected.ErrorContains)
 	}
 }
 
+type decisionRequestFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		ServiceID     string            `json:"serviceId"`
+		Path          string            `json:"path"`
+		TrailingSlash bool              `json:"trailingSlash"`
+		Headers       map[string]string `json:"headers"`
+	} `json:"inputs"`
+	Expected struct {
+		Replicas       int32  `json:"replicas"`
+		Path           string `json:"path"`
+		ServiceID      string `json:"serviceId"`
+		Authorization  string `json:"authorization"`
+		PinDefaultPath bool   `json:"pinDefaultPath"`
+	} `json:"expected"`
+}
+
 // TestQueryDecisionReplicasRequest pins the request the server actually sees:
 // the documented path and query parameter, plus any configured headers.
 func TestQueryDecisionReplicasRequest(t *testing.T) {
-	tests := []struct {
-		name          string
-		path          string
-		trailingSlash bool
-		wantPath      string
-	}{
-		{name: "an empty path falls back to the default", wantPath: defaultDecisionsPath},
-		{name: "a configured path is used", path: "/v2/decisions", wantPath: "/v2/decisions"},
-		{name: "a path without a leading slash still forms one", path: "decisions", wantPath: defaultDecisionsPath},
-		{name: "a trailing slash on serverAddress is not doubled", path: defaultDecisionsPath, trailingSlash: true, wantPath: defaultDecisionsPath},
-	}
-
-	for _, tt := range tests {
+	cases.Run(t, "testdata/helpers/decision-request", func(t *testing.T, tc decisionRequestFile) {
+		if tc.Expected.PinDefaultPath && tc.Expected.Path != defaultDecisionsPath {
+			t.Fatalf("path %q, want default %q", tc.Expected.Path, defaultDecisionsPath)
+		}
 		var gotPath, gotServiceID, gotAuth string
 		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
 			gotPath = req.URL.Path
 			gotServiceID = req.URL.Query().Get("serviceId")
 			gotAuth = req.Header.Get("Authorization")
-			_, _ = w.Write([]byte(decisionBody(`{"serviceId":%q,"replicas":{"active":4}}`, sampleServiceID)))
+			_, _ = fmt.Fprintf(w, `{"apiVersion":%q,"decisions":[{"serviceId":%q,"replicas":{"active":%d}}]}`,
+				decisionsAPIVersion, tc.Inputs.ServiceID, tc.Expected.Replicas)
 		}))
+		defer srv.Close()
 
 		address := srv.URL
-		if tt.trailingSlash {
+		if tc.Inputs.TrailingSlash {
 			address += "/"
 		}
-		got, err := queryDecisionReplicas(address, tt.path, sampleServiceID, "",
-			map[string]string{"Authorization": "Bearer t"})
+		got, err := queryDecisionReplicas(address, tc.Inputs.Path, tc.Inputs.ServiceID, "", tc.Inputs.Headers)
 		if err != nil {
-			t.Errorf("%s: unexpected error: %v", tt.name, err)
+			t.Fatalf("unexpected error: %v", err)
 		}
-		if got != 4 {
-			t.Errorf("%s: got %d replicas, want 4", tt.name, got)
+		if got != tc.Expected.Replicas {
+			t.Fatalf("got %d replicas, want %d", got, tc.Expected.Replicas)
 		}
-		if gotPath != tt.wantPath {
-			t.Errorf("%s: path = %q, want %q", tt.name, gotPath, tt.wantPath)
+		if gotPath != tc.Expected.Path {
+			t.Fatalf("path = %q, want %q", gotPath, tc.Expected.Path)
 		}
-		if gotServiceID != sampleServiceID {
-			t.Errorf("%s: serviceId = %q, want %q", tt.name, gotServiceID, sampleServiceID)
+		if gotServiceID != tc.Expected.ServiceID {
+			t.Fatalf("serviceId = %q, want %q", gotServiceID, tc.Expected.ServiceID)
 		}
-		if gotAuth != "Bearer t" {
-			t.Errorf("%s: Authorization = %q, want %q", tt.name, gotAuth, "Bearer t")
+		if gotAuth != tc.Expected.Authorization {
+			t.Fatalf("Authorization = %q, want %q", gotAuth, tc.Expected.Authorization)
 		}
-		srv.Close()
-	}
+	})
+}
+
+type decisionNon200File struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		ServiceID string `json:"serviceId"`
+		Status    int    `json:"status"`
+		Body      string `json:"body"`
+	} `json:"inputs"`
+	Expected struct {
+		ErrorContains string `json:"errorContains"`
+	} `json:"expected"`
 }
 
 // TestQueryDecisionReplicasNon200 checks that a non-OK reply is an error even
 // when its body happens to be a well-formed decision payload.
 func TestQueryDecisionReplicasNon200(t *testing.T) {
-	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-		w.WriteHeader(http.StatusServiceUnavailable)
-		_, _ = w.Write([]byte(decisionBody(`{"serviceId":"svc","replicas":{"active":3}}`)))
-	}))
-	defer srv.Close()
+	cases.Run(t, "testdata/helpers/decision-non-200", func(t *testing.T, tc decisionNon200File) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.WriteHeader(tc.Inputs.Status)
+			_, _ = w.Write([]byte(tc.Inputs.Body))
+		}))
+		defer srv.Close()
 
-	_, err := queryDecisionReplicas(srv.URL, "", "svc", "", nil)
-	if err == nil || !strings.Contains(err.Error(), "status 503") {
-		t.Errorf("error = %v, want it to mention status 503", err)
-	}
+		_, err := queryDecisionReplicas(srv.URL, "", tc.Inputs.ServiceID, "", nil)
+		if err == nil || !strings.Contains(err.Error(), tc.Expected.ErrorContains) {
+			t.Errorf("error = %v, want it to contain %q", err, tc.Expected.ErrorContains)
+		}
+	})
+}
+
+type decisionMisconfigFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		ServerAddress string `json:"serverAddress"`
+		ServiceID     string `json:"serviceId"`
+	} `json:"inputs"`
+	Expected struct {
+		ErrorContains string `json:"errorContains"`
+	} `json:"expected"`
 }
 
 // TestQueryDecisionReplicasMisconfigured covers the two settings that make a
 // request impossible to build. They are caught before the round trip so the
 // error names the field rather than the URL.
 func TestQueryDecisionReplicasMisconfigured(t *testing.T) {
-	if _, err := queryDecisionReplicas("", "", "svc", "", nil); err == nil ||
-		!strings.Contains(err.Error(), "serverAddress") {
-		t.Errorf("empty serverAddress: error = %v, want it to name serverAddress", err)
-	}
-	if _, err := queryDecisionReplicas("http://example.invalid", "", "", "", nil); err == nil ||
-		!strings.Contains(err.Error(), "serviceId") {
-		t.Errorf("empty serviceId: error = %v, want it to name serviceId", err)
-	}
+	cases.Run(t, "testdata/helpers/decision-misconfigured", func(t *testing.T, tc decisionMisconfigFile) {
+		_, err := queryDecisionReplicas(tc.Inputs.ServerAddress, "", tc.Inputs.ServiceID, "", nil)
+		if err == nil || !strings.Contains(err.Error(), tc.Expected.ErrorContains) {
+			t.Errorf("error = %v, want it to contain %q", err, tc.Expected.ErrorContains)
+		}
+	})
+}
+
+type recommendCustomFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Active               string `json:"active"`
+		MinReplicas          int32  `json:"minReplicas"`
+		MaxReplicas          int32  `json:"maxReplicas"`
+		ReadyReplicas        int64  `json:"readyReplicas"`
+		SpecReplicas         int64  `json:"specReplicas"`
+		ServiceID            string `json:"serviceId"`
+		Namespace            string `json:"namespace"`
+		APIVersion           string `json:"apiVersion"`
+		PinAPIVersion        bool   `json:"pinAPIVersion"`
+		AbsentCustomProvider bool   `json:"absentCustomProvider"`
+	} `json:"inputs"`
+	Expected struct {
+		Replicas int32 `json:"replicas"`
+	} `json:"expected"`
 }
 
 // TestRecommendReplicasCustomProvider pins the semantic that separates the two
@@ -266,75 +234,63 @@ func TestQueryDecisionReplicasMisconfigured(t *testing.T) {
 // Prometheus sample is. The CRD bounds still apply on top of it — the server
 // recommends, this operator keeps the rail.
 func TestRecommendReplicasCustomProvider(t *testing.T) {
-	tests := []struct {
-		name               string
-		active             string
-		min, max           int32
-		readyReplicas      int64
-		want               int32
-		unknownAPIVersion  bool
-		absentCustomConfig bool
-	}{
-		{
-			name: "the decision is used as-is, not scaled by readyReplicas",
-			// 8 ready replicas would turn a per-replica metric into a much larger
-			// number; a decision of 15 is already the answer.
-			active: "15", min: 1, max: 20, readyReplicas: 8, want: 15,
-		},
-		{
-			name:   "a decision above maxReplicas is clamped",
-			active: "15", min: 1, max: 5, readyReplicas: 8, want: 5,
-		},
-		{
-			name:   "a decision of zero clamps up to minReplicas",
-			active: "0", min: 2, max: 20, readyReplicas: 8, want: 2,
-		},
-		{
-			name:   "an unreadable decision holds the fleet where it is",
-			active: "15", min: 1, max: 20, readyReplicas: 8, want: 8,
-			unknownAPIVersion: true,
-		},
-		{
-			name:   "a missing customProvider holds the fleet where it is",
-			active: "15", min: 1, max: 20, readyReplicas: 8, want: 8,
-			absentCustomConfig: true,
-		},
-	}
-
-	for _, tt := range tests {
-		apiVersion := decisionsAPIVersion
-		if tt.unknownAPIVersion {
-			apiVersion = "some.other.api/v1"
+	cases.Run(t, "testdata/helpers/recommend-custom", func(t *testing.T, tc recommendCustomFile) {
+		in := tc.Inputs
+		if in.PinAPIVersion && in.APIVersion != decisionsAPIVersion {
+			t.Fatalf("apiVersion %q, want %q", in.APIVersion, decisionsAPIVersion)
 		}
 		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-			_, _ = w.Write(fmt.Appendf(nil,
-				`{"apiVersion":%q,"decisions":[{"namespace":%q,"serviceId":"svc","replicas":{"active":%s}}]}`,
-				apiVersion, decisionNamespace, tt.active))
+			_, _ = fmt.Fprintf(w,
+				`{"apiVersion":%q,"decisions":[{"namespace":%q,"serviceId":%q,"replicas":{"active":%s}}]}`,
+				in.APIVersion, in.Namespace, in.ServiceID, in.Active)
 		}))
+		defer srv.Close()
 
 		scaler := &autoscalingv1alpha1.LLMScaler{
 			Spec: autoscalingv1alpha1.LLMScalerSpec{
 				MetricProvider: autoscalingv1alpha1.MetricProviderCustom,
 				ServerAddress:  srv.URL,
-				CustomProvider: &autoscalingv1alpha1.CustomProviderSpec{ServiceID: "svc"},
-				MinReplicas:    tt.min,
-				MaxReplicas:    tt.max,
+				CustomProvider: &autoscalingv1alpha1.CustomProviderSpec{ServiceID: in.ServiceID},
+				MinReplicas:    in.MinReplicas,
+				MaxReplicas:    in.MaxReplicas,
 			},
 		}
-		if tt.absentCustomConfig {
+		if in.AbsentCustomProvider {
 			scaler.Spec.CustomProvider = nil
 		}
 
 		r := &LLMScalerReconciler{}
 		key := types.NamespacedName{Namespace: "ns", Name: "s"}
-		// specReplicas is what "hold where it is" means, so it is set to the same
-		// figure as readyReplicas to keep the failure cases unambiguous.
-		got := r.recommendReplicas(context.Background(), scaler, key, tt.readyReplicas, tt.readyReplicas)
-		if got != tt.want {
-			t.Errorf("%s: recommendReplicas = %d, want %d", tt.name, got, tt.want)
+		// specReplicas is what "hold where it is" means. The files set it to the
+		// same figure as readyReplicas so a hold is not confused with a clamp.
+		got := r.recommendReplicas(context.Background(), scaler, key, in.ReadyReplicas, in.SpecReplicas)
+		if got != tc.Expected.Replicas {
+			t.Fatalf("recommendReplicas = %d, want %d", got, tc.Expected.Replicas)
 		}
-		srv.Close()
-	}
+	})
+}
+
+type metricInput struct {
+	Name   string `json:"name"`
+	Query  string `json:"query"`
+	Target string `json:"target"`
+}
+
+type recommendPromFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Sample        string        `json:"sample"`
+		ReadyReplicas int64         `json:"readyReplicas"`
+		SpecReplicas  int64         `json:"specReplicas"`
+		MinReplicas   int32         `json:"minReplicas"`
+		MaxReplicas   int32         `json:"maxReplicas"`
+		Metrics       []metricInput `json:"metrics"`
+	} `json:"inputs"`
+	Expected struct {
+		Replicas int32  `json:"replicas"`
+		Path     string `json:"path"`
+	} `json:"expected"`
 }
 
 // TestRecommendReplicasDefaultProviderIsPrometheus checks that an LLMScaler with
@@ -342,31 +298,35 @@ func TestRecommendReplicasCustomProvider(t *testing.T) {
 // Prometheus at admission, but objects reaching the controller from a test or an
 // older CRD may carry it empty.
 func TestRecommendReplicasDefaultProviderIsPrometheus(t *testing.T) {
-	var gotPath string
-	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
-		gotPath = req.URL.Path
-		_, _ = w.Write([]byte(`{"status":"success","data":{"result":[{"metric":{},"value":[0,"1.6"]}]}}`))
-	}))
-	defer srv.Close()
+	cases.Run(t, "testdata/helpers/recommend-default-prometheus", func(t *testing.T, tc recommendPromFile) {
+		var gotPath string
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+			gotPath = req.URL.Path
+			_, _ = fmt.Fprintf(w, `{"status":"success","data":{"result":[{"metric":{},"value":[0,"%s"]}]}}`, tc.Inputs.Sample)
+		}))
+		defer srv.Close()
 
-	scaler := &autoscalingv1alpha1.LLMScaler{
-		Spec: autoscalingv1alpha1.LLMScalerSpec{
-			ServerAddress: srv.URL,
-			MinReplicas:   1,
-			MaxReplicas:   20,
-			Metrics: []autoscalingv1alpha1.MetricSpec{
-				{Name: "kv-cache", Query: "avg(kv)", Target: "0.8"},
+		metrics := make([]autoscalingv1alpha1.MetricSpec, len(tc.Inputs.Metrics))
+		for i, m := range tc.Inputs.Metrics {
+			metrics[i] = autoscalingv1alpha1.MetricSpec{Name: m.Name, Query: m.Query, Target: m.Target}
+		}
+		scaler := &autoscalingv1alpha1.LLMScaler{
+			Spec: autoscalingv1alpha1.LLMScalerSpec{
+				ServerAddress: srv.URL,
+				MinReplicas:   tc.Inputs.MinReplicas,
+				MaxReplicas:   tc.Inputs.MaxReplicas,
+				Metrics:       metrics,
 			},
-		},
-	}
+		}
 
-	r := &LLMScalerReconciler{}
-	key := types.NamespacedName{Namespace: "ns", Name: "s"}
-	// ceil(4 * 1.6/0.8) = 8, i.e. the per-replica ratio the Custom path skips.
-	if got := r.recommendReplicas(context.Background(), scaler, key, 4, 4); got != 8 {
-		t.Errorf("recommendReplicas = %d, want 8", got)
-	}
-	if gotPath != "/api/v1/query" {
-		t.Errorf("path = %q, want the Prometheus query API", gotPath)
-	}
+		r := &LLMScalerReconciler{}
+		key := types.NamespacedName{Namespace: "ns", Name: "s"}
+		got := r.recommendReplicas(context.Background(), scaler, key, tc.Inputs.ReadyReplicas, tc.Inputs.SpecReplicas)
+		if got != tc.Expected.Replicas {
+			t.Errorf("recommendReplicas = %d, want %d", got, tc.Expected.Replicas)
+		}
+		if gotPath != tc.Expected.Path {
+			t.Errorf("path = %q, want %q", gotPath, tc.Expected.Path)
+		}
+	})
 }
diff --git a/internal/controller/autoscaling/guards_cases_test.go b/internal/controller/autoscaling/guards_cases_test.go
deleted file mode 100644
index bf9d5e2..0000000
--- a/internal/controller/autoscaling/guards_cases_test.go
+++ /dev/null
@@ -1,404 +0,0 @@
-/*
-Copyright 2026.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-*/
-
-package controller
-
-import (
-	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
-)
-
-// guardHoldCases run one guard over a WorkloadState and a ScaleIntent.
-// checkTerminating is written onto the package flag for the row and restored.
-// Rollout rows cover each in-progress signal, a finished rollout, a
-// non-Deployment, and a hold in each direction. Settle rows cover direction,
-// a ready count that has not caught up (including 0), and the terminating
-// gate on and off.
-var guardHoldCases = []guardHoldCase{
-	{
-		name:  "rollout: generation not observed holds a scale-up",
-		guard: "rollout",
-		state: WorkloadState{
-			Spec: 5, Total: 5, Updated: 5, Ready: 5,
-			Generation: 4, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:     ScaleIntent{From: 5, To: 7},
-		wantHeld:   true,
-		wantReason: autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantWhy:    msgRolloutInProgress,
-	},
-	{
-		name:  "rollout: updated replicas behind spec holds a scale-down",
-		guard: "rollout",
-		state: WorkloadState{
-			Spec: 5, Total: 4, Updated: 4, Ready: 4,
-			Generation: 3, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:     ScaleIntent{From: 5, To: 3},
-		wantHeld:   true,
-		wantReason: autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantWhy:    msgRolloutInProgress,
-	},
-	{
-		name:  "rollout: old-template pods hold a no-op",
-		guard: "rollout",
-		state: WorkloadState{
-			Spec: 5, Total: 6, Updated: 5, Ready: 5,
-			Generation: 3, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:     ScaleIntent{From: 5, To: 5},
-		wantHeld:   true,
-		wantReason: autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantWhy:    msgRolloutInProgress,
-	},
-	{
-		name:  "rollout: fully rolled out does not hold",
-		guard: "rollout",
-		state: WorkloadState{
-			Spec: 5, Total: 5, Updated: 5, Ready: 5,
-			Generation: 3, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:     ScaleIntent{From: 5, To: 4},
-		wantReason: autoscalingv1alpha1.ReasonRolloutInProgress,
-	},
-	{
-		name:  "rollout: non-Deployment is never held",
-		guard: "rollout",
-		state: WorkloadState{
-			Spec: 5, Total: 6, Updated: 0, Ready: 5,
-			Generation: 4, ObservedGeneration: 1,
-		},
-		intent:     ScaleIntent{From: 5, To: 7},
-		wantReason: autoscalingv1alpha1.ReasonRolloutInProgress,
-	},
-	{
-		name:             "settle: scale-up is not held",
-		guard:            "settle",
-		state:            WorkloadState{Spec: 5, Ready: 2},
-		intent:           ScaleIntent{From: 5, To: 7},
-		checkTerminating: false,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-	{
-		name:             "settle: no-op is not held",
-		guard:            "settle",
-		state:            WorkloadState{Spec: 5, Ready: 2},
-		intent:           ScaleIntent{From: 5, To: 5},
-		checkTerminating: false,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-	{
-		name:             "settle: ready differs from spec",
-		guard:            "settle",
-		state:            WorkloadState{Spec: 5, Ready: 3},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: false,
-		wantHeld:         true,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantWhy:          msgScaleDownNotSettled,
-	},
-	{
-		// A ready count of 0 is the raw missing-field value. Treating it as
-		// spec would skip the guard.
-		name:             "settle: ready count of 0 is not settled",
-		guard:            "settle",
-		state:            WorkloadState{Spec: 5, Ready: 0},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: false,
-		wantHeld:         true,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantWhy:          msgScaleDownNotSettled,
-	},
-	{
-		name:             "settle: ready back at spec",
-		guard:            "settle",
-		state:            WorkloadState{Spec: 5, Ready: 5},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: false,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-	{
-		name:  "settle: terminating pods with the gate off",
-		guard: "settle",
-		state: WorkloadState{
-			Spec: 5, Ready: 5, Terminating: 2, TerminatingReported: true,
-		},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: false,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-	{
-		name:  "settle: terminating pods with the gate on",
-		guard: "settle",
-		state: WorkloadState{
-			Spec: 5, Ready: 5, Terminating: 2, TerminatingReported: true,
-		},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: true,
-		wantHeld:         true,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantWhy:          msgScaleDownNotSettled,
-	},
-	{
-		name:  "settle: terminating count of 0 with the gate on",
-		guard: "settle",
-		state: WorkloadState{
-			Spec: 5, Ready: 5, Terminating: 0, TerminatingReported: true,
-		},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: true,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-	{
-		// The count is ignored when the field was not reported. A non-zero
-		// Terminating here would still be settled; the parser stores 0 in
-		// that case, so this row is what makes the flag observable.
-		name:  "settle: terminating not reported with the gate on",
-		guard: "settle",
-		state: WorkloadState{
-			Spec: 5, Ready: 5, Terminating: 4, TerminatingReported: false,
-		},
-		intent:           ScaleIntent{From: 5, To: 4},
-		checkTerminating: true,
-		wantReason:       autoscalingv1alpha1.ReasonScaleDownSettling,
-	},
-}
-
-// guardOrderCases walk the guards slice. The first hold wins: a rollout is
-// reported ahead of a scale-down that has not settled.
-var guardOrderCases = []guardOrderCase{
-	{
-		name: "rollout wins over an unsettled scale-down",
-		state: WorkloadState{
-			Spec: 5, Total: 5, Updated: 5, Ready: 2,
-			Generation: 4, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:      ScaleIntent{From: 5, To: 3},
-		wantReason:  autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantMessage: msgRolloutInProgress,
-	},
-	{
-		name: "settle holds once the rollout is done",
-		state: WorkloadState{
-			Spec: 5, Total: 5, Updated: 5, Ready: 3,
-			Generation: 3, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent:      ScaleIntent{From: 5, To: 4},
-		wantReason:  autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantMessage: msgScaleDownNotSettled,
-	},
-	{
-		name: "no guard holds a settled scale-down",
-		state: WorkloadState{
-			Spec: 5, Total: 5, Updated: 5, Ready: 5,
-			Generation: 3, ObservedGeneration: 3,
-			TracksRollout: true,
-		},
-		intent: ScaleIntent{From: 5, To: 4},
-	},
-}
-
-// workloadStateCases are readWorkloadState over each kind, then the missing
-// and invalid-field defaults. A nil field is omitted. The four full-status
-// rows share one set of numbers so the only per-kind difference is
-// TracksRollout.
-var workloadStateCases = []workloadStateCase{
-	{
-		name: "Deployment with full status",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "StatefulSet with full status",
-		kind: "StatefulSet",
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-		},
-	},
-	{
-		name: "LeaderWorkerSet with full status",
-		kind: "LeaderWorkerSet",
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-		},
-	},
-	{
-		name: "generic kind with full status",
-		kind: "ReplicaSet",
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-		},
-	},
-	{
-		name: "missing spec.replicas reads as 1",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			ready: 3, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 1, Total: 6, Updated: 2, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "spec.replicas of 0 is kept",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: 0, ready: 0, total: 0, updated: 0,
-			generation: 2, observed: 2,
-		},
-		want: WorkloadState{
-			Spec: 0, Generation: 2, ObservedGeneration: 2,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "missing readyReplicas reads as 0",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: 4, total: 6, updated: 2, terminating: 1,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 0,
-			Terminating: 1, TerminatingReported: true,
-			Generation: 9, ObservedGeneration: 8,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "missing terminatingReplicas is not reported",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, total: 6, updated: 2,
-			generation: 9, observed: 8,
-		},
-		want: WorkloadState{
-			Spec: 4, Total: 6, Updated: 2, Ready: 3,
-			Generation: 9, ObservedGeneration: 8,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "missing generation, observedGeneration, updated, and total read as 0",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: 4, ready: 3, terminating: 1,
-		},
-		want: WorkloadState{
-			Spec: 4, Ready: 3,
-			Terminating: 1, TerminatingReported: true,
-			TracksRollout: true,
-		},
-	},
-	{
-		name: "invalid-typed fields behave like missing",
-		kind: deploymentKind,
-		fields: workloadObjectFields{
-			spec: "nope", ready: "nope", total: "nope", updated: "nope",
-			terminating: "nope", generation: "nope", observed: "nope",
-		},
-		want: WorkloadState{Spec: 1, TracksRollout: true},
-	},
-}
-
-// targetGuardCases are the guard result the controller turns into a
-// condition, starting from a target object rather than a WorkloadState.
-// readyReplicas is the raw count: 0 does not count as settled.
-// gen/observed/updated/spec/ready build the object. status.replicas stays
-// at 5 inside the builder. The runner parses that object and evaluates guards.
-var targetGuardCases = []targetGuardCase{
-	{
-		name: "rollout holds a scale-up",
-		gen:  4, observed: 3, updated: 5,
-		kind: deploymentKind, spec: 5, ready: 5, target: 7,
-		wantReason:  autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantMessage: msgRolloutInProgress,
-	},
-	{
-		name: "rollout wins over an unsettled scale-down",
-		gen:  4, observed: 3, updated: 5,
-		kind: deploymentKind, spec: 5, ready: 2, target: 3,
-		wantReason:  autoscalingv1alpha1.ReasonRolloutInProgress,
-		wantMessage: msgRolloutInProgress,
-	},
-	{
-		name: "unsettled scale-down",
-		gen:  3, observed: 3, updated: 5, // fully rolled out
-		kind: deploymentKind, spec: 5, ready: 3, target: 4,
-		wantReason:  autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantMessage: msgScaleDownNotSettled,
-	},
-	{
-		// readyReplicas 0 is also what a missing field reads as. Treating it
-		// as spec would skip the guard.
-		name: "ready count of 0 is not settled",
-		gen:  3, observed: 3, updated: 5,
-		kind: deploymentKind, spec: 5, ready: 0, target: 4,
-		wantReason:  autoscalingv1alpha1.ReasonScaleDownSettling,
-		wantMessage: msgScaleDownNotSettled,
-	},
-	{
-		name: "settled scale-down may proceed",
-		gen:  3, observed: 3, updated: 5,
-		kind: deploymentKind, spec: 5, ready: 5, target: 4,
-	},
-	{
-		name: "scale-up is not gated on settle",
-		gen:  3, observed: 3, updated: 5,
-		kind: deploymentKind, spec: 5, ready: 2, target: 6,
-	},
-	{
-		name: "non-Deployment ignores the rollout guard",
-		gen:  4, observed: 1, updated: 0,
-		kind: "StatefulSet", spec: 5, ready: 5, target: 7,
-	},
-}
diff --git a/internal/controller/autoscaling/guards_test.go b/internal/controller/autoscaling/guards_test.go
index 0316ca6..13919d0 100644
--- a/internal/controller/autoscaling/guards_test.go
+++ b/internal/controller/autoscaling/guards_test.go
@@ -17,39 +17,68 @@ limitations under the License.
 package controller
 
 import (
+	"math"
 	"testing"
 
 	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-type guardHoldCase struct {
-	name             string
-	guard            string
-	state            WorkloadState
-	intent           ScaleIntent
-	checkTerminating bool
-	wantHeld         bool
-	wantReason       string
-	wantWhy          string
+// stateInput is the WorkloadState a guard case feeds Hold or evaluateGuards.
+// Zero fields are left at the Go zero, which is what an omitted YAML key means.
+type stateInput struct {
+	Spec                int32 `json:"spec"`
+	Total               int32 `json:"total"`
+	Updated             int32 `json:"updated"`
+	Ready               int32 `json:"ready"`
+	Terminating         int32 `json:"terminating"`
+	TerminatingReported bool  `json:"terminatingReported"`
+	Generation          int64 `json:"generation"`
+	ObservedGeneration  int64 `json:"observedGeneration"`
+	TracksRollout       bool  `json:"tracksRollout"`
+}
+
+func (s stateInput) state() WorkloadState { return WorkloadState(s) }
+
+type intentInput struct {
+	From int32 `json:"from"`
+	To   int32 `json:"to"`
+}
+
+func (in intentInput) intent() ScaleIntent { return ScaleIntent(in) }
+
+type guardHoldFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Guard            string      `json:"guard"`
+		State            stateInput  `json:"state"`
+		Intent           intentInput `json:"intent"`
+		CheckTerminating bool        `json:"checkTerminating"`
+	} `json:"inputs"`
+	Expected struct {
+		Held   bool   `json:"held"`
+		Reason string `json:"reason"`
+		Why    string `json:"why"`
+	} `json:"expected"`
 }
 
 func TestGuardHold(t *testing.T) {
-	for _, tc := range guardHoldCases {
-		t.Run(tc.name, func(t *testing.T) {
-			setTerminatingGate(t, tc.checkTerminating)
-			g, ok := guardByName(tc.guard)
-			if !ok {
-				t.Fatalf("unknown guard %q", tc.guard)
-			}
-			if g.Reason() != tc.wantReason {
-				t.Fatalf("Reason() = %q, want %q", g.Reason(), tc.wantReason)
-			}
-			held, why := g.Hold(tc.state, tc.intent)
-			if held != tc.wantHeld || why != tc.wantWhy {
-				t.Fatalf("Hold() = (%v, %q), want (%v, %q)", held, why, tc.wantHeld, tc.wantWhy)
-			}
-		})
-	}
+	cases.Run(t, "testdata/guards/hold", func(t *testing.T, tc guardHoldFile) {
+		setTerminatingGate(t, tc.Inputs.CheckTerminating)
+		g, ok := guardByName(tc.Inputs.Guard)
+		if !ok {
+			t.Fatalf("unknown guard %q", tc.Inputs.Guard)
+		}
+		if g.Reason() != tc.Expected.Reason {
+			t.Fatalf("Reason() = %q, want %q", g.Reason(), tc.Expected.Reason)
+		}
+		held, why := g.Hold(tc.Inputs.State.state(), tc.Inputs.Intent.intent())
+		if held != tc.Expected.Held || why != tc.Expected.Why {
+			t.Fatalf("Hold() = (%v, %q), want (%v, %q)", held, why, tc.Expected.Held, tc.Expected.Why)
+		}
+	})
 }
 
 func guardByName(name string) (Guard, bool) {
@@ -63,26 +92,29 @@ func guardByName(name string) (Guard, bool) {
 	}
 }
 
-type guardOrderCase struct {
-	name        string
-	state       WorkloadState
-	intent      ScaleIntent
-	wantReason  string
-	wantMessage string
+type guardOrderFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		State  stateInput  `json:"state"`
+		Intent intentInput `json:"intent"`
+	} `json:"inputs"`
+	Expected struct {
+		Reason  string `json:"reason"`
+		Message string `json:"message"`
+	} `json:"expected"`
 }
 
 // TestGuardOrder evaluates the guards slice. The flag is forced off so a
 // row does not depend on whichever case ran before it.
 func TestGuardOrder(t *testing.T) {
-	for _, tc := range guardOrderCases {
-		t.Run(tc.name, func(t *testing.T) {
-			setTerminatingGate(t, false)
-			got := evaluateGuards(tc.state, tc.intent)
-			if got.reason != tc.wantReason || got.message != tc.wantMessage {
-				t.Fatalf("got {%q %q}, want {%q %q}", got.reason, got.message, tc.wantReason, tc.wantMessage)
-			}
-		})
-	}
+	cases.Run(t, "testdata/guards/order", func(t *testing.T, tc guardOrderFile) {
+		setTerminatingGate(t, false)
+		got := evaluateGuards(tc.Inputs.State.state(), tc.Inputs.Intent.intent())
+		if got.reason != tc.Expected.Reason || got.message != tc.Expected.Message {
+			t.Fatalf("got {%q %q}, want {%q %q}", got.reason, got.message, tc.Expected.Reason, tc.Expected.Message)
+		}
+	})
 }
 
 func setTerminatingGate(t *testing.T, on bool) {
@@ -92,125 +124,99 @@ func setTerminatingGate(t *testing.T, on bool) {
 	t.Cleanup(func() { checkTerminatingReplicas = prev })
 }
 
-// workloadObjectFields is one target object as data. A nil field is omitted.
-// An int, int32, or int64 is stored as int64, which is what NestedInt64 reads.
-// Any other type, such as a string standing in for an invalid field, is stored
-// as itself.
-type workloadObjectFields struct {
-	spec        any
-	ready       any
-	total       any
-	updated     any
-	terminating any
-	generation  any
-	observed    any
+type workloadStateFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Kind   string                     `json:"kind"`
+		Object *unstructured.Unstructured `json:"object"`
+	} `json:"inputs"`
+	Expected stateInput `json:"expected"`
 }
 
-type workloadStateCase struct {
-	name   string
-	kind   string
-	fields workloadObjectFields
-	want   WorkloadState
+func TestReadWorkloadState(t *testing.T) {
+	cases.Run(t, "testdata/guards/workload-state", func(t *testing.T, tc workloadStateFile) {
+		normalizeUnstructured(tc.Inputs.Object)
+		got := readWorkloadState(tc.Inputs.Object, tc.Inputs.Kind)
+		if got != tc.Expected.state() {
+			t.Fatalf("got %+v, want %+v", got, tc.Expected.state())
+		}
+	})
 }
 
-func TestReadWorkloadState(t *testing.T) {
-	for _, tc := range workloadStateCases {
-		t.Run(tc.name, func(t *testing.T) {
-			got := readWorkloadState(workloadObject(t, tc.fields), tc.kind)
-			if got != tc.want {
-				t.Fatalf("got %+v, want %+v", got, tc.want)
-			}
-		})
-	}
+// targetGuardFile is one object-level guard result. The object is a
+// Deployment-shaped document; kind selects the parse. status.replicas stays
+// at 5 in every file. readyReplicas 0 is the raw count and is not settled.
+type targetGuardFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Kind   string                     `json:"kind"`
+		Object *unstructured.Unstructured `json:"object"`
+		Target int32                      `json:"target"`
+	} `json:"inputs"`
+	Expected struct {
+		Spec               int32  `json:"spec"`
+		Ready              int32  `json:"ready"`
+		Total              int32  `json:"total"`
+		Updated            int32  `json:"updated"`
+		Generation         int64  `json:"generation"`
+		ObservedGeneration int64  `json:"observedGeneration"`
+		Reason             string `json:"reason"`
+		Message            string `json:"message"`
+	} `json:"expected"`
 }
 
-func workloadObject(t *testing.T, f workloadObjectFields) *unstructured.Unstructured {
-	t.Helper()
-	obj := map[string]any{}
-	set := func(v any, fields ...string) {
-		t.Helper()
-		stored, ok := nestedJSONValue(v)
-		if !ok {
-			return
+// TestGuardsOnTarget parses each file's object and evaluates the guards.
+func TestGuardsOnTarget(t *testing.T) {
+	cases.Run(t, "testdata/guards/target", func(t *testing.T, tc targetGuardFile) {
+		normalizeUnstructured(tc.Inputs.Object)
+		state := readWorkloadState(tc.Inputs.Object, tc.Inputs.Kind)
+		exp := tc.Expected
+		if state.Spec != exp.Spec || state.Ready != exp.Ready {
+			t.Fatalf("parsed spec/ready = %d/%d, want %d/%d", state.Spec, state.Ready, exp.Spec, exp.Ready)
 		}
-		if err := unstructured.SetNestedField(obj, stored, fields...); err != nil {
-			t.Fatalf("set %v: %v", fields, err)
+		if state.Total != exp.Total || state.Updated != exp.Updated ||
+			state.Generation != exp.Generation || state.ObservedGeneration != exp.ObservedGeneration {
+			t.Fatalf("parsed rollout fields = total %d updated %d gen %d observed %d",
+				state.Total, state.Updated, state.Generation, state.ObservedGeneration)
 		}
-	}
-	set(f.spec, "spec", "replicas")
-	set(f.ready, "status", "readyReplicas")
-	set(f.total, "status", "replicas")
-	set(f.updated, "status", "updatedReplicas")
-	set(f.terminating, "status", "terminatingReplicas")
-	set(f.generation, "metadata", "generation")
-	set(f.observed, "status", "observedGeneration")
-	return &unstructured.Unstructured{Object: obj}
+		got := evaluateGuards(state, ScaleIntent{From: state.Spec, To: tc.Inputs.Target})
+		if got.reason != exp.Reason || got.message != exp.Message {
+			t.Fatalf("got {%q %q}, want {%q %q}", got.reason, got.message, exp.Reason, exp.Message)
+		}
+	})
 }
 
-// nestedJSONValue converts a case-table number to the int64 NestedInt64
-// accepts. nil means the field is omitted. A non-number is returned unchanged
-// so an invalid type can be written on purpose.
-func nestedJSONValue(v any) (any, bool) {
-	switch n := v.(type) {
-	case nil:
-		return nil, false
-	case int:
-		return int64(n), true
-	case int32:
-		return int64(n), true
-	case int64:
-		return n, true
-	default:
-		return v, true
+// normalizeUnstructured turns whole-number float64 values into int64.
+// NestedInt64 only accepts int64. A string standing in for an invalid field
+// is left alone so it still reads as missing.
+func normalizeUnstructured(obj *unstructured.Unstructured) {
+	if obj == nil || obj.Object == nil {
+		return
 	}
+	obj.Object = normalizeJSON(obj.Object).(map[string]any)
 }
 
-// targetGuardCase is one object-level guard result: the fields build a
-// Deployment-shaped object, kind selects the parse.
-type targetGuardCase struct {
-	name        string
-	gen         int64
-	observed    int64
-	updated     int64
-	kind        string
-	spec        int64
-	ready       int64
-	target      int32
-	wantReason  string
-	wantMessage string
-}
-
-// deploymentObject builds the target a targetGuardCase describes.
-// status.replicas stays at 5; the cases vary generation, updatedReplicas,
-// spec.replicas, and readyReplicas.
-func deploymentObject(gen, observed, updated, spec, ready int64) *unstructured.Unstructured {
-	return &unstructured.Unstructured{Object: map[string]any{
-		"metadata": map[string]any{"generation": gen},
-		"spec":     map[string]any{"replicas": spec},
-		"status": map[string]any{
-			"observedGeneration": observed,
-			"updatedReplicas":    updated,
-			"replicas":           int64(5),
-			"readyReplicas":      ready,
-		},
-	}}
-}
-
-// TestGuardsOnTarget parses each row's object and evaluates the guards.
-func TestGuardsOnTarget(t *testing.T) {
-	for _, tc := range targetGuardCases {
-		t.Run(tc.name, func(t *testing.T) {
-			state := readWorkloadState(deploymentObject(tc.gen, tc.observed, tc.updated, tc.spec, tc.ready), tc.kind)
-			if state.Spec != int32(tc.spec) || state.Ready != int32(tc.ready) {
-				t.Fatalf("parsed spec/ready = %d/%d, want %d/%d", state.Spec, state.Ready, tc.spec, tc.ready)
-			}
-			if state.Total != 5 || state.Updated != int32(tc.updated) || state.Generation != tc.gen || state.ObservedGeneration != tc.observed {
-				t.Fatalf("parsed rollout fields = total %d updated %d gen %d observed %d", state.Total, state.Updated, state.Generation, state.ObservedGeneration)
-			}
-			got := evaluateGuards(state, ScaleIntent{From: state.Spec, To: tc.target})
-			if got.reason != tc.wantReason || got.message != tc.wantMessage {
-				t.Fatalf("got {%q %q}, want {%q %q}", got.reason, got.message, tc.wantReason, tc.wantMessage)
-			}
-		})
+func normalizeJSON(v any) any {
+	switch n := v.(type) {
+	case float64:
+		if !math.IsNaN(n) && !math.IsInf(n, 0) && n == math.Trunc(n) &&
+			n >= math.MinInt64 && n <= math.MaxInt64 {
+			return int64(n)
+		}
+		return n
+	case map[string]any:
+		for k, child := range n {
+			n[k] = normalizeJSON(child)
+		}
+		return n
+	case []any:
+		for i := range n {
+			n[i] = normalizeJSON(n[i])
+		}
+		return n
+	default:
+		return v
 	}
 }
diff --git a/internal/controller/autoscaling/llmscaler_casefile_test.go b/internal/controller/autoscaling/llmscaler_casefile_test.go
new file mode 100644
index 0000000..ffcccc8
--- /dev/null
+++ b/internal/controller/autoscaling/llmscaler_casefile_test.go
@@ -0,0 +1,399 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package controller
+
+import (
+	"encoding/json"
+	"fmt"
+	"time"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+
+	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+)
+
+// sinceCheck is how one step reads status.deferredSince.
+// Omitted means do not check. JSON null means the field was cleared.
+// "seed" is the instant written by seedDeferredSince. Any other string is a
+// Go duration offset from deferralStart, including "0s".
+type sinceCheck struct {
+	kind   sinceKind
+	offset time.Duration
+}
+
+func (s *sinceCheck) UnmarshalJSON(b []byte) error {
+	if string(b) == "null" {
+		s.kind = sinceNone
+		return nil
+	}
+	var raw string
+	if err := json.Unmarshal(b, &raw); err != nil {
+		return fmt.Errorf("deferredSince: %w", err)
+	}
+	if raw == "seed" {
+		s.kind = sinceSeeded
+		return nil
+	}
+	d, err := time.ParseDuration(raw)
+	if err != nil {
+		return fmt.Errorf("deferredSince %q: %w", raw, err)
+	}
+	s.kind = sinceOffset
+	s.offset = d
+	return nil
+}
+
+func (s sinceCheck) expect() sinceExpect { return sinceExpect(s) }
+
+type specPatchInput struct {
+	MinReplicas                *int32 `json:"minReplicas"`
+	MaxStepReplicas            *int32 `json:"maxStepReplicas"`
+	StabilizationWindowSeconds *int32 `json:"stabilizationWindowSeconds"`
+	DeferTimeoutSeconds        *int32 `json:"deferTimeoutSeconds"`
+	SyncPeriodSeconds          *int32 `json:"syncPeriodSeconds"`
+}
+
+func (p specPatchInput) patch() scalerSpecPatch {
+	return scalerSpecPatch{
+		minReplicas:                p.MinReplicas,
+		maxStepReplicas:            p.MaxStepReplicas,
+		stabilizationWindowSeconds: p.StabilizationWindowSeconds,
+		deferTimeoutSeconds:        p.DeferTimeoutSeconds,
+		syncPeriodSeconds:          p.SyncPeriodSeconds,
+	}
+}
+
+type deployInput struct {
+	Spec        *int32 `json:"spec"`
+	Replicas    int32  `json:"replicas"`
+	Updated     int32  `json:"updated"`
+	Ready       int32  `json:"ready"`
+	Available   int32  `json:"available"`
+	Terminating *int32 `json:"terminating"`
+}
+
+func (d deployInput) state() deployState {
+	return deployState{
+		spec:        d.Spec,
+		replicas:    d.Replicas,
+		updated:     d.Updated,
+		ready:       d.Ready,
+		available:   d.Available,
+		terminating: d.Terminating,
+	}
+}
+
+type stepInput struct {
+	By             string         `json:"by"`
+	At             cases.Duration `json:"at"`
+	Metric         string         `json:"metric"`
+	Deploy         *deployInput   `json:"deploy"`
+	ObserveSpec    *int32         `json:"observeSpec"`
+	Rollout        bool           `json:"rollout"`
+	RestoreStatus  bool           `json:"restoreStatus"`
+	SkipReconcile  bool           `json:"skipReconcile"`
+	SnapshotStatus bool           `json:"snapshotStatus"`
+}
+
+type condInput struct {
+	Status                    string   `json:"status"`
+	Reason                    string   `json:"reason"`
+	MessageContains           []string `json:"messageContains"`
+	ObservedGenerationCurrent bool     `json:"observedGenerationCurrent"`
+}
+
+func (c condInput) expect() condExpect {
+	return condExpect{
+		status:                    metav1.ConditionStatus(c.Status),
+		reason:                    c.Reason,
+		messageContains:           c.MessageContains,
+		observedGenerationCurrent: c.ObservedGenerationCurrent,
+	}
+}
+
+type eventInput struct {
+	Prefix   string   `json:"prefix"`
+	Contains []string `json:"contains"`
+}
+
+func (e eventInput) expect() eventExpect {
+	return eventExpect{prefix: e.Prefix, contains: e.Contains}
+}
+
+type stepExpected struct {
+	Replicas      *int32          `json:"replicas"`
+	Desired       *int32          `json:"desired"`
+	Applied       *int32          `json:"applied"`
+	Current       *int32          `json:"current"`
+	DeferredSince sinceCheck      `json:"deferredSince"`
+	Condition     *condInput      `json:"condition"`
+	RequeueAfter  *cases.Duration `json:"requeueAfter"`
+	Events        []eventInput    `json:"events"`
+	NoMoreEvents  bool            `json:"noMoreEvents"`
+	ForcedWrite   *bool           `json:"forcedWrite"`
+	ForcedWriteAt *cases.Duration `json:"forcedWriteAt"`
+}
+
+func (e stepExpected) expect() stepExpect {
+	out := stepExpect{
+		replicas:      e.Replicas,
+		desired:       e.Desired,
+		applied:       e.Applied,
+		current:       e.Current,
+		deferredSince: e.DeferredSince.expect(),
+		requeueAfter:  durPtr(e.RequeueAfter),
+		noMoreEvents:  e.NoMoreEvents,
+		forcedWrite:   e.ForcedWrite,
+		forcedWriteAt: durPtr(e.ForcedWriteAt),
+	}
+	if e.Condition != nil {
+		cond := e.Condition.expect()
+		out.condition = &cond
+	}
+	if len(e.Events) > 0 {
+		out.events = make([]eventExpect, len(e.Events))
+		for i, ev := range e.Events {
+			out.events[i] = ev.expect()
+		}
+	}
+	return out
+}
+
+func (s stepInput) step(exp stepExpected) scenarioStep {
+	out := scenarioStep{
+		by:             s.By,
+		at:             s.At.Std(),
+		metric:         s.Metric,
+		observeSpec:    s.ObserveSpec,
+		rollout:        s.Rollout,
+		restoreStatus:  s.RestoreStatus,
+		skipReconcile:  s.SkipReconcile,
+		snapshotStatus: s.SnapshotStatus,
+		expect:         exp.expect(),
+	}
+	if s.Deploy != nil {
+		st := s.Deploy.state()
+		out.deploy = &st
+	}
+	return out
+}
+
+func durPtr(d *cases.Duration) *time.Duration {
+	if d == nil {
+		return nil
+	}
+	v := d.Std()
+	return &v
+}
+
+type scalerInputs struct {
+	Prelude                  string          `json:"prelude"`
+	Spec                     specPatchInput  `json:"spec"`
+	Deploy                   *deployInput    `json:"deploy"`
+	Rollout                  bool            `json:"rollout"`
+	Metric                   string          `json:"metric"`
+	SeedDeferredSince        *cases.Duration `json:"seedDeferredSince"`
+	CheckTerminating         *bool           `json:"checkTerminating"`
+	PlainReconciler          bool            `json:"plainReconciler"`
+	ExpectNilRecommendations bool            `json:"expectNilRecommendations"`
+	Steps                    []stepInput     `json:"steps"`
+}
+
+type scalerExpected struct {
+	Steps []stepExpected `json:"steps"`
+}
+
+// scalerCase is one Prometheus-provider reconcile spec.
+// inputs.steps and expected.steps are zipped by index.
+type scalerCase struct {
+	Name        string         `json:"name,omitempty"`
+	Description string         `json:"description,omitempty"`
+	Inputs      scalerInputs   `json:"inputs"`
+	Expected    scalerExpected `json:"expected"`
+}
+
+func (c scalerCase) scenario() scalerScenario {
+	in := c.Inputs
+	if len(in.Steps) != len(c.Expected.Steps) {
+		Fail(fmt.Sprintf("%s: %d input steps and %d expected steps", c.Description, len(in.Steps), len(c.Expected.Steps)))
+	}
+	steps := make([]scenarioStep, len(in.Steps))
+	for i := range in.Steps {
+		steps[i] = in.Steps[i].step(c.Expected.Steps[i])
+	}
+	sc := scalerScenario{
+		prelude:                  in.Prelude,
+		spec:                     in.Spec.patch(),
+		rollout:                  in.Rollout,
+		metric:                   in.Metric,
+		seedDeferredSince:        durPtr(in.SeedDeferredSince),
+		checkTerminating:         in.CheckTerminating,
+		plainReconciler:          in.PlainReconciler,
+		expectNilRecommendations: in.ExpectNilRecommendations,
+		steps:                    steps,
+	}
+	if in.Deploy != nil {
+		st := in.Deploy.state()
+		sc.deploy = &st
+	}
+	return sc
+}
+
+func runScalerCase(c scalerCase) {
+	runScalerScenario(c.scenario())
+}
+
+type customReconcileFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		CreateBy    string       `json:"createBy"`
+		Deploy      *deployInput `json:"deploy"`
+		DeployBy    string       `json:"deployBy"`
+		CloseServer bool         `json:"closeServer"`
+		CloseBy     string       `json:"closeBy"`
+		ReconcileBy string       `json:"reconcileBy"`
+		ReplicasBy  string       `json:"replicasBy"`
+		StatusBy    string       `json:"statusBy"`
+	} `json:"inputs"`
+	Expected struct {
+		Replicas                int32  `json:"replicas"`
+		Desired                 *int32 `json:"desired"`
+		DecisionStatus          string `json:"decisionStatus"`
+		DecisionReason          string `json:"decisionReason"`
+		DecisionMessageNonEmpty bool   `json:"decisionMessageNonEmpty"`
+	} `json:"expected"`
+}
+
+func runCustomReconcileFile(c customReconcileFile) {
+	exp := c.Expected
+	tc := customReconcileCase{
+		createBy:                c.Inputs.CreateBy,
+		deployBy:                c.Inputs.DeployBy,
+		closeServer:             c.Inputs.CloseServer,
+		closeBy:                 c.Inputs.CloseBy,
+		reconcileBy:             c.Inputs.ReconcileBy,
+		replicasBy:              c.Inputs.ReplicasBy,
+		wantReplicas:            exp.Replicas,
+		statusBy:                c.Inputs.StatusBy,
+		wantDesired:             exp.Desired,
+		decisionStatus:          metav1.ConditionStatus(exp.DecisionStatus),
+		decisionReason:          exp.DecisionReason,
+		decisionMessageNonEmpty: exp.DecisionMessageNonEmpty,
+	}
+	if c.Inputs.Deploy != nil {
+		st := c.Inputs.Deploy.state()
+		tc.deploy = &st
+	}
+	runCustomProviderReconcile(tc)
+}
+
+type customAdmissionFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		MetricProvider     string `json:"metricProvider"`
+		OmitCustomProvider bool   `json:"omitCustomProvider"`
+		PinProvider        bool   `json:"pinProvider"`
+	} `json:"inputs"`
+	Expected struct {
+		ErrorContains string `json:"errorContains"`
+	} `json:"expected"`
+}
+
+func runCustomAdmissionFile(c customAdmissionFile) {
+	if c.Inputs.PinProvider && c.Inputs.MetricProvider != autoscalingv1alpha1.MetricProviderPrometheus {
+		Fail(fmt.Sprintf("metricProvider %q, want %q", c.Inputs.MetricProvider, autoscalingv1alpha1.MetricProviderPrometheus))
+	}
+	runCustomProviderAdmission(customAdmissionCase{
+		metricProvider:     c.Inputs.MetricProvider,
+		omitCustomProvider: c.Inputs.OmitCustomProvider,
+		wantErr:            c.Expected.ErrorContains,
+	})
+}
+
+type deferDefaultFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+	} `json:"inputs"`
+	Expected struct {
+		Seconds int32 `json:"seconds"`
+	} `json:"expected"`
+}
+
+func runDeferDefault(tc deferDefaultFile) {
+	if tc.Expected.Seconds != autoscalingv1alpha1.DefaultDeferTimeoutSeconds {
+		Fail(fmt.Sprintf("seconds %d, want DefaultDeferTimeoutSeconds %d",
+			tc.Expected.Seconds, autoscalingv1alpha1.DefaultDeferTimeoutSeconds))
+	}
+	scaler := &autoscalingv1alpha1.LLMScaler{}
+	Expect(k8sClient.Get(promFix.ctx, promFix.scaler, scaler)).To(Succeed())
+	// The fixture leaves the field unset; the CRD default fills it.
+	Expect(scaler.Spec.DeferTimeoutSeconds).NotTo(BeNil())
+	Expect(*scaler.Spec.DeferTimeoutSeconds).To(Equal(tc.Expected.Seconds))
+}
+
+type customDefaultFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		CreateBy  string `json:"createBy"`
+		SwitchBy  string `json:"switchBy"`
+		Query     string `json:"query"`
+		Target    string `json:"target"`
+		ServiceID string `json:"serviceId"`
+	} `json:"inputs"`
+	Expected struct {
+		MetricProvider string `json:"metricProvider"`
+		Path           string `json:"path"`
+	} `json:"expected"`
+}
+
+func runCustomDefault(tc customDefaultFile) {
+	if tc.Expected.MetricProvider != autoscalingv1alpha1.MetricProviderPrometheus {
+		Fail(fmt.Sprintf("metricProvider %q, want %q", tc.Expected.MetricProvider, autoscalingv1alpha1.MetricProviderPrometheus))
+	}
+	if tc.Expected.Path != defaultDecisionsPath {
+		Fail(fmt.Sprintf("path %q, want %q", tc.Expected.Path, defaultDecisionsPath))
+	}
+	if tc.Inputs.ServiceID != sampleServiceID {
+		Fail(fmt.Sprintf("serviceId %q, want %q", tc.Inputs.ServiceID, sampleServiceID))
+	}
+
+	By(tc.Inputs.CreateBy)
+	scaler := customFix.newScaler()
+	scaler.Spec.MetricProvider = ""
+	scaler.Spec.CustomProvider = nil
+	scaler.Spec.Metrics = []autoscalingv1alpha1.MetricSpec{{Query: tc.Inputs.Query, Target: tc.Inputs.Target}}
+	Expect(k8sClient.Create(customFix.ctx, scaler)).To(Succeed())
+
+	stored := &autoscalingv1alpha1.LLMScaler{}
+	Expect(k8sClient.Get(customFix.ctx, customFix.scaler, stored)).To(Succeed())
+	Expect(stored.Spec.MetricProvider).To(Equal(tc.Expected.MetricProvider))
+
+	By(tc.Inputs.SwitchBy)
+	stored.Spec.MetricProvider = autoscalingv1alpha1.MetricProviderCustom
+	stored.Spec.CustomProvider = &autoscalingv1alpha1.CustomProviderSpec{ServiceID: tc.Inputs.ServiceID}
+	Expect(k8sClient.Update(customFix.ctx, stored)).To(Succeed())
+
+	Expect(k8sClient.Get(customFix.ctx, customFix.scaler, stored)).To(Succeed())
+	Expect(stored.Spec.CustomProvider.Path).To(Equal(tc.Expected.Path))
+}
diff --git a/internal/controller/autoscaling/llmscaler_controller_cases_test.go b/internal/controller/autoscaling/llmscaler_controller_cases_test.go
deleted file mode 100644
index f379025..0000000
--- a/internal/controller/autoscaling/llmscaler_controller_cases_test.go
+++ /dev/null
@@ -1,722 +0,0 @@
-/*
-Copyright 2026.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-*/
-
-package controller
-
-import (
-	"time"
-
-	. "github.com/onsi/ginkgo/v2"
-	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
-
-	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
-)
-
-// metricScalingCases are plain metric-driven writes: scale up, down to min, to zero, wake from zero, and the ready-replica ratio fallback.
-var metricScalingCases = []TableEntry{
-	// ceil(1 * 0.85/0.5) = ceil(1.7) = 2. The mock sample is 0.85 and the target is 0.5.
-	Entry("should successfully scale the deployment based on metrics", scalerScenario{
-		plainReconciler: true,
-		steps: []scenarioStep{
-			{by: "running the Reconciler"},
-			{by: "checking if the Deployment replicas increased", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(2))}},
-		},
-	}),
-
-	// desired = ceil(5 * 0/0.5) = 0, clamped up to minReplicas (1).
-	Entry("should scale down to minReplicas when the metric is zero", scalerScenario{
-		plainReconciler: true,
-		steps: []scenarioStep{
-			{by: "setting the Deployment to 5 replicas and the metric to 0 (idle)", skipReconcile: true,
-				metric: "0", // empty queue
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 5, available: 5}},
-			{by: "running the Reconciler"},
-			{by: "checking it scaled down to minReplicas", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-
-	// desired = ceil(2 * 0/0.5) = 0, clamped to minReplicas (0).
-	Entry("should scale down to zero when minReplicas is 0", scalerScenario{
-		plainReconciler: true,
-		prelude:         "lowering the scaler's minReplicas to 0",
-		spec:            scalerSpecPatch{minReplicas: ptr(int32(0))},
-		steps: []scenarioStep{
-			{by: "setting the Deployment to 2 replicas, fully rolled out", skipReconcile: true,
-				metric: "0", // idle -> wants zero
-				deploy: &deployState{spec: ptr(int32(2)), replicas: 2, updated: 2, ready: 2, available: 2}},
-			{by: "running the Reconciler"},
-			{by: "checking it scaled down to zero", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(0))}},
-		},
-	}),
-
-	// desired = ceil(1 * 0.85/0.5) = 2 against the single-replica baseline,
-	// not ceil(0 * 0.85/0.5) = 0.
-	Entry("should wake the fleet from zero when the metric turns positive", scalerScenario{
-		plainReconciler: true,
-		prelude:         "lowering the scaler's minReplicas to 0 and parking the Deployment at 0",
-		spec:            scalerSpecPatch{minReplicas: ptr(int32(0))},
-		// Load arrives while the fleet is at zero. mockMetricValue is already the fixture sample.
-		deploy: &deployState{spec: ptr(int32(0)), replicas: 0, updated: 0, ready: 0, available: 0},
-		steps: []scenarioStep{
-			{by: "running the Reconciler"},
-			{by: "checking the fleet woke up", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(2))}},
-		},
-	}),
-
-	// readyReplicas is 0, so the ratio falls back to spec (3):
-	// ceil(3 * 0.85/0.5) = 6, clamped to maxReplicas 5. Without that
-	// fallback the single-replica baseline would recommend 2.
-	// status.currentReplicas stays 0. ready 0 is omitted and reads as missing,
-	// so available must be 0 too or the apiserver rejects the status.
-	Entry("should keep the ready-replica fallback for the scaling ratio only", scalerScenario{
-		plainReconciler: true,
-		steps: []scenarioStep{
-			{deploy: &deployState{spec: ptr(int32(3)), replicas: 3, updated: 3, ready: 0, available: 0},
-				expect: stepExpect{
-					replicas:      ptr(int32(5)),
-					current:       ptr(int32(0)),
-					desired:       ptr(int32(5)),
-					applied:       ptr(int32(5)),
-					deferredSince: sinceNil(),
-					condition: &condExpect{
-						status: metav1.ConditionFalse,
-						reason: autoscalingv1alpha1.ReasonNotDeferred,
-					},
-				}},
-		},
-	}),
-}
-
-// rolloutGuardCases (G8) defer scaling while the Deployment is mid-rollout.
-var rolloutGuardCases = []TableEntry{
-	// The metric would compute desired=2, but the rollout guard defers
-	// scaling, so replicas stay at 1.
-	Entry("should defer scaling while the Deployment is rolling out", scalerScenario{
-		plainReconciler: true,
-		steps: []scenarioStep{
-			{by: "marking the Deployment as mid-rollout", rollout: true, skipReconcile: true},
-			{by: "running the Reconciler"},
-			{by: "checking the Deployment was NOT scaled", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-}
-
-// settleGuardCases (G7) defer scale-down until the fleet is back at spec, and do not gate scale-up.
-var settleGuardCases = []TableEntry{
-	Entry("should defer a scale-down until readyReplicas is back at specReplicas", scalerScenario{
-		plainReconciler: true,
-		metric:          "0", // empty queue -> wants minReplicas
-		steps: []scenarioStep{
-			{by: "setting the Deployment to 5 replicas with only 3 ready",
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 3, available: 3}},
-			{by: "checking the scale-down was deferred", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(5))}},
-			{by: "letting the fleet converge",
-				deploy: &deployState{replicas: 5, updated: 5, ready: 5, available: 5}},
-			{by: "checking the scale-down now proceeds", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-
-	// The fallback path. readyReplicas excludes terminating pods, so the
-	// fleet reads as settled the moment the previous step's pods are marked
-	// for deletion — the descent is then paced by syncPeriodSeconds, not by
-	// the drain.
-	Entry("should ignore still-draining pods while checkTerminatingReplicas is off", scalerScenario{
-		plainReconciler:  true,
-		checkTerminating: ptr(false),
-		metric:           "0",
-		steps: []scenarioStep{
-			{by: "setting 5 ready with 2 still draining",
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 5, available: 5, terminating: ptr(int32(2))}},
-			{by: "checking the scale-down was not held back by the drain", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-
-	Entry("should also wait for the drain when checkTerminatingReplicas is on", scalerScenario{
-		plainReconciler:  true,
-		checkTerminating: ptr(true),
-		metric:           "0",
-		steps: []scenarioStep{
-			{by: "setting 5 ready with 2 still draining",
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 5, available: 5, terminating: ptr(int32(2))}},
-			{by: "checking it is deferred; readyReplicas alone would have said settled", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(5))}},
-			{by: "letting the drain finish",
-				deploy: &deployState{replicas: 5, updated: 5, ready: 5, available: 5, terminating: ptr(int32(0))}},
-			{by: "checking the scale-down now proceeds", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-
-	// ceil(2 * 0.85/0.5) = 4; the settle guard only gates scale-down.
-	Entry("should not gate scale-up on the previous scale-down settling", scalerScenario{
-		plainReconciler: true,
-		steps: []scenarioStep{
-			{by: "setting 3 replicas with 2 ready and 1 still terminating, under load",
-				// Over the 0.5 target -> wants more replicas.
-				metric: mockMetricValue,
-				deploy: &deployState{spec: ptr(int32(3)), replicas: 3, updated: 3, ready: 2, available: 2, terminating: ptr(int32(1))}},
-			{by: "checking capacity was added despite the unsettled fleet", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(4))}},
-		},
-	}),
-}
-
-// stabilizationAndStepCases hold scale-down for the stabilization window and walk down one maxStepReplicas at a time.
-var stabilizationAndStepCases = []TableEntry{
-	Entry("should hold for the stabilization window between scale-down steps", scalerScenario{
-		plainReconciler: true,
-		prelude:         "enabling a 300s window and a 2-replica step cap",
-		spec: scalerSpecPatch{
-			stabilizationWindowSeconds: ptr(int32(300)),
-			maxStepReplicas:            ptr(int32(2)),
-		},
-		metric: "0",
-		steps: []scenarioStep{
-			{by: "starting settled at 5 with an idle metric",
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 5, available: 5}},
-			{by: "checking the first step landed", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(3))}},
-			{by: "settling the fleet at 3 so neither the rollout nor the settle guard fires",
-				observeSpec: ptr(int32(3))},
-			{by: "checking the next step is held; without the pin it would drop to 1", skipReconcile: true,
-				expect: stepExpect{replicas: ptr(int32(3))}},
-		},
-	}),
-
-	Entry("should walk down one step at a time when maxStepReplicas is set", scalerScenario{
-		plainReconciler: true,
-		prelude:         "capping the scale-down step at 2 replicas",
-		spec:            scalerSpecPatch{maxStepReplicas: ptr(int32(2))},
-		steps: []scenarioStep{
-			{by: "setting the Deployment to 5 replicas and the metric to 0 (idle)", skipReconcile: true,
-				metric: "0", // empty queue
-				deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 5, available: 5}},
-			// desired = 1 (clamped to minReplicas), capped to 5 - 2 = 3.
-			{by: "running the Reconciler; raw recommendation is minReplicas but the step is capped",
-				expect: stepExpect{replicas: ptr(int32(3))}},
-			// The cap rate-limits the write; the recommendation is still 1.
-			{by: "checking status reports the uncapped recommendation, not the step", skipReconcile: true,
-				expect: stepExpect{desired: ptr(int32(1))}},
-			// 3 - 2 = 1, which is minReplicas, so it lands there rather than below.
-			{by: "running the Reconciler again; it takes the next step down",
-				deploy: &deployState{spec: ptr(int32(3)), replicas: 3, updated: 3, ready: 3, available: 3},
-				expect: stepExpect{replicas: ptr(int32(1))}},
-		},
-	}),
-}
-
-// deferralTimeoutCases force a held write once the deferral times out, including a restarted reconciler and a timeout of 0.
-var deferralTimeoutCases = []TableEntry{
-	Entry("should force a scale-up once a stuck rollout exceeds the deferral timeout", scalerScenario{
-		spec:    scalerSpecPatch{deferTimeoutSeconds: ptr(int32(60))},
-		rollout: true,
-		steps: []scenarioStep{
-			// Deadline is 60s out and the default sync period is 15s, so the
-			// wait is the sync period. The recommendation is recorded even
-			// though nothing was written.
-			{by: "deferring the first sync and stamping deferredSince", at: 0,
-				expect: stepExpect{
-					requeueAfter:  ptr(15 * time.Second),
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(0),
-					desired:       ptr(int32(2)),
-					applied:       ptr(int32(1)),
-					condition: &condExpect{
-						status:                    metav1.ConditionTrue,
-						reason:                    autoscalingv1alpha1.ReasonRolloutInProgress,
-						observedGenerationCurrent: true,
-						messageContains: []string{
-							"holding 1 replicas (wanted 2)",
-							msgRolloutInProgress,
-							deferralStart.Format(time.RFC3339),
-							"will apply after 60s",
-						},
-					},
-					events: []eventExpect{{
-						prefix:   "Normal " + eventReasonScalingDeferred,
-						contains: []string{"holding 1 replicas (wanted 2)"},
-					}},
-				}},
-			// One second remains until the deadline, under the sync period.
-			{by: "reconciling again before the timeout, without restamping or rewriting",
-				at: 59 * time.Second,
-				expect: stepExpect{
-					requeueAfter:  ptr(time.Second),
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(0),
-					applied:       ptr(int32(1)),
-					noMoreEvents:  true,
-				}},
-			// The rollout guard still matches after the write, so the clock
-			// stays where the deferral started. The forced write itself is not
-			// a hold, so the next sync is a full period.
-			{by: "advancing past the timeout and forcing the held scale-up",
-				at: 60 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(2)),
-					deferredSince: sinceAt(0),
-					desired:       ptr(int32(2)),
-					applied:       ptr(int32(2)),
-					requeueAfter:  ptr(15 * time.Second),
-					condition: &condExpect{
-						status:                    metav1.ConditionFalse,
-						reason:                    autoscalingv1alpha1.ReasonDeferralTimedOut,
-						observedGenerationCurrent: true,
-						messageContains:           []string{"scaling 1 -> 2 anyway"},
-					},
-					events: []eventExpect{{
-						prefix:   "Warning " + eventReasonDeferralTimedOut,
-						contains: []string{msgRolloutInProgress, "1 -> 2"},
-					}},
-					noMoreEvents: true,
-				}},
-		},
-	}),
-
-	// readyReplicas == 0 is omitted from status, so the controller sees the
-	// field as missing. A ready==0 -> spec fallback in the guard would have made
-	// this look settled and scaled immediately; the settle guard now uses
-	// the raw count. The ratio still falls back to spec, which is why an
-	// idle metric recommends minReplicas rather than staying at 5.
-	// available must be 0 too: the apiserver rejects availableReplicas > readyReplicas,
-	// and 0 is omitted, so the controller sees that field as missing as well.
-	Entry("should force a scale-down of an unready fleet once the deferral times out", scalerScenario{
-		spec: scalerSpecPatch{
-			deferTimeoutSeconds: ptr(int32(60)),
-			maxStepReplicas:     ptr(int32(2)),
-		},
-		prelude: "capping the step and parking 5 replicas with none ready",
-		deploy:  &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 0, available: 0},
-		metric:  "0",
-		steps: []scenarioStep{
-			// Uncapped recommendation is minReplicas; the held write is the step.
-			{at: 0,
-				expect: stepExpect{
-					replicas:      ptr(int32(5)),
-					deferredSince: sinceAt(0),
-					desired:       ptr(int32(1)),
-					current:       ptr(int32(0)),
-					applied:       ptr(int32(5)),
-					condition: &condExpect{
-						status: metav1.ConditionTrue,
-						reason: autoscalingv1alpha1.ReasonScaleDownSettling,
-						messageContains: []string{
-							"holding 5 replicas (wanted 3)",
-							"will apply after 60s",
-						},
-					},
-				}},
-			{by: "keeping the same deferral until the timeout", at: 59 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(5)),
-					deferredSince: sinceAt(0),
-				}},
-			// None of the pods are ready, so the settle guard still matches and
-			// the clock is not restarted for the next step.
-			{by: "forcing the capped step, not the uncapped recommendation", at: 60 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(3)),
-					deferredSince: sinceAt(0),
-					desired:       ptr(int32(1)),
-					applied:       ptr(int32(3)),
-					current:       ptr(int32(0)),
-					condition: &condExpect{
-						status:          metav1.ConditionFalse,
-						reason:          autoscalingv1alpha1.ReasonDeferralTimedOut,
-						messageContains: []string{"scaling 5 -> 3 anyway"},
-					},
-					// The entering deferral event, then the timeout warning.
-					events: []eventExpect{
-						{prefix: "Normal " + eventReasonScalingDeferred},
-						{prefix: "Warning " + eventReasonDeferralTimedOut, contains: []string{msgScaleDownNotSettled, "5 -> 3"}},
-					},
-				}},
-		},
-	}),
-
-	// deferredSince lives in status, so a process restart (a new reconciler,
-	// with an empty recommendation map) still enforces the deadline.
-	// The rollout is still in progress, so the deadline carried in status is
-	// what forced the write and it stays put.
-	Entry("should force a deferral that already exceeded the timeout on a fresh reconciler", scalerScenario{
-		spec:                     scalerSpecPatch{deferTimeoutSeconds: ptr(int32(60))},
-		rollout:                  true,
-		seedDeferredSince:        ptr(-2 * time.Hour),
-		expectNilRecommendations: true,
-		steps: []scenarioStep{
-			{at: 0, expect: stepExpect{
-				replicas:      ptr(int32(2)),
-				deferredSince: sinceSeed(),
-				applied:       ptr(int32(2)),
-				condition: &condExpect{
-					status: metav1.ConditionFalse,
-					reason: autoscalingv1alpha1.ReasonDeferralTimedOut,
-				},
-				events: []eventExpect{{
-					prefix:   "Warning " + eventReasonDeferralTimedOut,
-					contains: []string{"1 -> 2"},
-				}},
-			}},
-		},
-	}),
-
-	// Timeout 0 never forces, so the wait stays the sync period even though
-	// the recorded deferral is already a day old.
-	Entry("should never force a deferral when deferTimeoutSeconds is 0", scalerScenario{
-		spec:              scalerSpecPatch{deferTimeoutSeconds: ptr(int32(0))},
-		rollout:           true,
-		seedDeferredSince: ptr(-24 * time.Hour),
-		steps: []scenarioStep{
-			{at: 0, expect: stepExpect{
-				requeueAfter:  ptr(15 * time.Second),
-				replicas:      ptr(int32(1)),
-				deferredSince: sinceSeed(),
-				condition: &condExpect{
-					status: metav1.ConditionTrue,
-					reason: autoscalingv1alpha1.ReasonRolloutInProgress,
-					// Split so this is not a third copy of the controller's limit phrase.
-					messageContains: []string{"never forced (" + "deferTimeoutSeconds is 0)"},
-				},
-				events:       []eventExpect{{prefix: "Normal " + eventReasonScalingDeferred}},
-				noMoreEvents: true,
-			}},
-		},
-	}),
-}
-
-// deferralStatusCases cover the condition, status, and event lifecycle of one deferral.
-var deferralStatusCases = []TableEntry{
-	// 5 replicas, 3 ready, but not yet rolled out: the rollout guard wins
-	// over the settle guard. An idle metric wants to scale down.
-	Entry("should keep one deferral when the guard reason changes", scalerScenario{
-		spec:    scalerSpecPatch{deferTimeoutSeconds: ptr(int32(60))},
-		metric:  "0",
-		deploy:  &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 3, available: 3},
-		rollout: true,
-		steps: []scenarioStep{
-			{at: 0, expect: stepExpect{
-				deferredSince: sinceAt(0),
-				condition:     &condExpect{reason: autoscalingv1alpha1.ReasonRolloutInProgress},
-			}},
-			{by: "finishing the rollout while the fleet is still short of ready",
-				at:     10 * time.Second,
-				deploy: &deployState{replicas: 5, updated: 5, ready: 3, available: 3},
-				expect: stepExpect{
-					replicas:      ptr(int32(5)),
-					deferredSince: sinceAt(0),
-					condition:     &condExpect{reason: autoscalingv1alpha1.ReasonScaleDownSettling},
-					events: []eventExpect{
-						{contains: []string{msgRolloutInProgress}},
-						{contains: []string{msgScaleDownNotSettled}},
-					},
-				}},
-			{by: "not emitting again while the same guard still defers",
-				at: 10 * time.Second,
-				expect: stepExpect{
-					deferredSince: sinceAt(0),
-					noMoreEvents:  true,
-				}},
-		},
-	}),
-
-	// ceil(1 * 0.5/0.5) = 1, which is already spec, and the rollout is done,
-	// so no guard matches. The no-op ends the deferral.
-	Entry("should clear a deferral on a no-op once no guard matches", scalerScenario{
-		spec:    scalerSpecPatch{deferTimeoutSeconds: ptr(int32(60))},
-		rollout: true,
-		steps: []scenarioStep{
-			{at: 0, expect: stepExpect{deferredSince: sinceAt(0)}},
-			{by: "finishing the rollout at the replica count the metric already wants",
-				at:          10 * time.Second,
-				metric:      "0.5",
-				observeSpec: ptr(int32(1)),
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceNil(),
-					applied:       ptr(int32(1)),
-					forcedWrite:   ptr(false),
-					condition: &condExpect{
-						status:          metav1.ConditionFalse,
-						reason:          autoscalingv1alpha1.ReasonNotDeferred,
-						messageContains: []string{"already at the step target"},
-					},
-				}},
-		},
-	}),
-
-	Entry("should keep deferredSince across no-op syncs while a rollout guard still matches", scalerScenario{
-		spec:    scalerSpecPatch{deferTimeoutSeconds: ptr(int32(60))},
-		rollout: true,
-		steps: []scenarioStep{
-			// ceil(1 * 0.5/0.5) = 1, already spec. The guard matches, but there
-			// is nothing to hold, so deferredSince stays nil.
-			{by: "not starting a deferral on a no-op, even though the rollout guard matches",
-				at: 0, metric: "0.5",
-				expect: stepExpect{
-					deferredSince: sinceNil(),
-					replicas:      ptr(int32(1)),
-					condition: &condExpect{
-						status: metav1.ConditionFalse,
-						reason: autoscalingv1alpha1.ReasonNotDeferred,
-					},
-				}},
-			// ceil(1 * 0.85/0.5) = 2.
-			{by: "stamping deferredSince on the first sync that actually wants to write",
-				at: 10 * time.Second, metric: "0.85",
-				expect: stepExpect{
-					deferredSince: sinceAt(10 * time.Second),
-					replicas:      ptr(int32(1)),
-					condition: &condExpect{
-						status: metav1.ConditionTrue,
-						reason: autoscalingv1alpha1.ReasonRolloutInProgress,
-					},
-				}},
-			// The no-ops must not move the clock. The stamp is deferralStart+10s.
-			{by: "alternating a no-op with a held scale-up; the no-ops must not move the clock",
-				at: 20 * time.Second, metric: "0.5",
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(10 * time.Second),
-					condition: &condExpect{
-						status: metav1.ConditionTrue,
-						reason: autoscalingv1alpha1.ReasonRolloutInProgress,
-						messageContains: []string{
-							"nothing to write this sync",
-							"guard is still active",
-							"deferral clock keeps running",
-							deferralStart.Add(10 * time.Second).Format(time.RFC3339),
-							msgRolloutInProgress,
-						},
-					},
-				}},
-			{at: 30 * time.Second, metric: "0.85",
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(10 * time.Second),
-					condition: &condExpect{
-						status: metav1.ConditionTrue,
-						reason: autoscalingv1alpha1.ReasonRolloutInProgress,
-					},
-				}},
-			{at: 40 * time.Second, metric: "0.5",
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(10 * time.Second),
-					condition: &condExpect{
-						status: metav1.ConditionTrue,
-						reason: autoscalingv1alpha1.ReasonRolloutInProgress,
-						messageContains: []string{
-							"nothing to write this sync",
-							"guard is still active",
-							"deferral clock keeps running",
-							deferralStart.Add(10 * time.Second).Format(time.RFC3339),
-							msgRolloutInProgress,
-						},
-					},
-				}},
-			{by: "forcing the held scale-up once the original deadline passes",
-				at: 70 * time.Second, metric: "0.85",
-				expect: stepExpect{
-					replicas:      ptr(int32(2)),
-					deferredSince: sinceAt(10 * time.Second),
-					condition: &condExpect{
-						status:          metav1.ConditionFalse,
-						reason:          autoscalingv1alpha1.ReasonDeferralTimedOut,
-						messageContains: []string{"scaling 1 -> 2 anyway"},
-					},
-				}},
-		},
-	}),
-}
-
-// forcedWritePacingCases pace a second forced step to one per sync period, and clear the pace when the guard lifts.
-var forcedWritePacingCases = []TableEntry{
-	Entry("should force the next stuck scale-down step after one sync period", scalerScenario{
-		spec: scalerSpecPatch{
-			deferTimeoutSeconds: ptr(int32(60)),
-			syncPeriodSeconds:   ptr(int32(15)),
-			maxStepReplicas:     ptr(int32(2)),
-		},
-		prelude: "capping the step and parking 5 replicas with none ready",
-		deploy:  &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 0, available: 0},
-		metric:  "0",
-		steps: []scenarioStep{
-			{at: 0,
-				expect: stepExpect{
-					replicas:      ptr(int32(5)),
-					deferredSince: sinceAt(0),
-					condition:     &condExpect{reason: autoscalingv1alpha1.ReasonScaleDownSettling},
-				}},
-			{by: "forcing the first step once the timeout elapses", at: 60 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(3)),
-					deferredSince: sinceAt(0),
-					forcedWrite:   ptr(true),
-					forcedWriteAt: ptr(60 * time.Second),
-				}},
-			// The spec write bumped generation. Observe it with nothing ready
-			// so the settle guard, not a phantom rollout, is what still matches.
-			{by: "keeping the fleet unready and taking the next step one sync period later, without another timeout",
-				at:          75 * time.Second,
-				observeSpec: ptr(int32(0)),
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(0),
-					applied:       ptr(int32(1)),
-					condition: &condExpect{
-						status: metav1.ConditionFalse,
-						reason: autoscalingv1alpha1.ReasonDeferralTimedOut,
-						messageContains: []string{
-							"scaling 3 -> 1 anyway",
-							msgScaleDownNotSettled,
-						},
-					},
-				}},
-		},
-	}),
-
-	Entry("should not take a second forced step until a sync period has elapsed", scalerScenario{
-		spec: scalerSpecPatch{
-			deferTimeoutSeconds: ptr(int32(60)),
-			syncPeriodSeconds:   ptr(int32(15)),
-			maxStepReplicas:     ptr(int32(2)),
-		},
-		metric: "0",
-		deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 0, available: 0},
-		steps: []scenarioStep{
-			// Snapshot the pre-write status. The next step is the forced write
-			// this By describes.
-			{by: "forcing one scale-down step of an unready fleet", at: 0, snapshotStatus: true},
-			{at: 60 * time.Second, expect: stepExpect{replicas: ptr(int32(3))}},
-			// The target write stuck at 3. The status write that would have
-			// recorded it did not, so a retry still sees the deadline passed.
-			{by: "putting the status back to its pre-write state, as if that update had conflicted",
-				restoreStatus: true, skipReconcile: true},
-			{by: "reconciling again with the clock unchanged, which must not take the next step",
-				at: 60 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(3)),
-					requeueAfter:  ptr(15 * time.Second),
-					deferredSince: sinceAt(0),
-					applied:       ptr(int32(3)),
-					condition: &condExpect{
-						status:          metav1.ConditionTrue,
-						messageContains: []string{"forced write paced to one per sync period"},
-					},
-				}},
-			{by: "still holding one second short of the sync period",
-				at: 74 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(3)),
-					requeueAfter:  ptr(time.Second),
-					deferredSince: sinceAt(0),
-				}},
-			{by: "forcing the next step once the sync period has elapsed",
-				at: 75 * time.Second,
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceAt(0),
-					applied:       ptr(int32(1)),
-				}},
-		},
-	}),
-
-	Entry("should clear deferredSince and the forced-write pace when no guard matches", scalerScenario{
-		spec: scalerSpecPatch{
-			deferTimeoutSeconds: ptr(int32(60)),
-			maxStepReplicas:     ptr(int32(2)),
-		},
-		metric: "0",
-		deploy: &deployState{spec: ptr(int32(5)), replicas: 5, updated: 5, ready: 0, available: 0},
-		steps: []scenarioStep{
-			{at: 0},
-			{at: 60 * time.Second, expect: stepExpect{
-				replicas:      ptr(int32(3)),
-				forcedWrite:   ptr(true),
-				deferredSince: sinceAt(0),
-			}},
-			// Settled and fully observed: the next scale-down is an ordinary
-			// write, not a forced one, even though the pace has not elapsed.
-			{by: "letting the fleet become ready at the replica count just written",
-				at:          60 * time.Second,
-				observeSpec: ptr(int32(3)),
-				expect: stepExpect{
-					replicas:      ptr(int32(1)),
-					deferredSince: sinceNil(),
-					forcedWrite:   ptr(false),
-					condition: &condExpect{
-						status:          metav1.ConditionFalse,
-						reason:          autoscalingv1alpha1.ReasonNotDeferred,
-						messageContains: []string{"scaled 3 -> 1"},
-					},
-				}},
-		},
-	}),
-}
-
-// customProviderReconcileCases scale from a decision server, and hold replicas when that server cannot be reached.
-var customProviderReconcileCases = []TableEntry{
-	// spec.metrics is required only under the Prometheus provider, so a
-	// Custom scaler must be accepted without it. active 4 is an absolute
-	// count, so 1 ready replica scales straight to 4 rather than through
-	// the per-replica HPA ratio.
-	Entry("should scale the Deployment to the count the decision server reports", customReconcileCase{
-		createBy:       "creating an LLMScaler with no metrics at all",
-		reconcileBy:    "running the Reconciler",
-		replicasBy:     "checking the Deployment landed on the decision, not on a ratio of it",
-		wantReplicas:   4,
-		statusBy:       "checking the status reports the same recommendation",
-		wantDesired:    ptr(int32(4)),
-		decisionStatus: metav1.ConditionTrue,
-	}),
-
-	// Above minReplicas on purpose: at 1 replica "held" and "collapsed to
-	// minReplicas" are the same number, so the test would pass either way.
-	Entry("should hold replicas when the decision server is unreachable", customReconcileCase{
-		deploy:                  &deployState{spec: ptr(int32(3)), replicas: 3, updated: 3, ready: 3, available: 3},
-		deployBy:                "putting the Deployment at 3 replicas, above minReplicas",
-		closeServer:             true,
-		closeBy:                 "taking the decision server down",
-		reconcileBy:             "running the Reconciler",
-		replicasBy:              "checking the Deployment was left where it was, not collapsed to minReplicas",
-		wantReplicas:            3,
-		statusBy:                "checking the failure is reported on the status",
-		decisionStatus:          metav1.ConditionFalse,
-		decisionReason:          autoscalingv1alpha1.ReasonDecisionUnavailable,
-		decisionMessageNonEmpty: true,
-	}),
-}
-
-// customProviderAdmissionCases are the CRD union rules for the metric provider.
-var customProviderAdmissionCases = []TableEntry{
-	Entry("should reject a Custom scaler with no customProvider", customAdmissionCase{
-		omitCustomProvider: true,
-		wantErr:            "spec.customProvider is required",
-	}),
-	Entry("should reject a Prometheus scaler with no metrics", customAdmissionCase{
-		metricProvider:     autoscalingv1alpha1.MetricProviderPrometheus,
-		omitCustomProvider: true,
-		wantErr:            "spec.metrics must be non-empty",
-	}),
-}
diff --git a/internal/controller/autoscaling/llmscaler_controller_test.go b/internal/controller/autoscaling/llmscaler_controller_test.go
index 212a83a..f87395f 100644
--- a/internal/controller/autoscaling/llmscaler_controller_test.go
+++ b/internal/controller/autoscaling/llmscaler_controller_test.go
@@ -18,15 +18,14 @@ package controller
 
 import (
 	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
 
-	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-// The reconcile specs are tables of scalerScenario values. The runner and the
-// fixtures live in llmscaler_scenario_test.go; the rows live in
-// llmscaler_controller_cases_test.go. Helper tests that do not need envtest
-// are plain testing funcs in llmscaler_helpers_test.go.
+// The reconcile specs are one YAML file each under testdata/controller.
+// The runner and the fixtures live in llmscaler_scenario_test.go.
+// Helper tests that do not need envtest are plain testing funcs in
+// llmscaler_helpers_test.go.
 var _ = Describe("LLMScaler Controller", func() {
 	Context("When reconciling a resource", func() {
 		BeforeEach(func() {
@@ -37,21 +36,25 @@ var _ = Describe("LLMScaler Controller", func() {
 			promFix = nil
 		})
 
-		DescribeTable("metric scaling", runScalerScenario, metricScalingCases)
-		DescribeTable("rollout guard (G8)", runScalerScenario, rolloutGuardCases)
-		DescribeTable("settle guard (G7)", runScalerScenario, settleGuardCases)
-		DescribeTable("stabilization and max step", runScalerScenario, stabilizationAndStepCases)
-		DescribeTable("deferral timeout", runScalerScenario, deferralTimeoutCases)
-		DescribeTable("deferral status", runScalerScenario, deferralStatusCases)
-		DescribeTable("forced-write pacing", runScalerScenario, forcedWritePacingCases)
+		DescribeTable("metric scaling", runScalerCase,
+			cases.Entries("testdata/controller/metric-scaling", scalerDesc))
+		DescribeTable("rollout guard (G8)", runScalerCase,
+			cases.Entries("testdata/controller/rollout-guard", scalerDesc))
+		DescribeTable("settle guard (G7)", runScalerCase,
+			cases.Entries("testdata/controller/settle-guard", scalerDesc))
+		DescribeTable("stabilization and max step", runScalerCase,
+			cases.Entries("testdata/controller/stabilization-and-step", scalerDesc))
+		DescribeTable("deferral timeout", runScalerCase,
+			cases.Entries("testdata/controller/deferral-timeout", scalerDesc))
+		DescribeTable("deferral status", runScalerCase,
+			cases.Entries("testdata/controller/deferral-status", scalerDesc))
+		DescribeTable("forced-write pacing", runScalerCase,
+			cases.Entries("testdata/controller/forced-write-pacing", scalerDesc))
 
-		It("should default deferTimeoutSeconds to 600", func() {
-			scaler := &autoscalingv1alpha1.LLMScaler{}
-			Expect(k8sClient.Get(promFix.ctx, promFix.scaler, scaler)).To(Succeed())
-			// The fixture leaves the field unset; the CRD default fills it.
-			Expect(scaler.Spec.DeferTimeoutSeconds).NotTo(BeNil())
-			Expect(*scaler.Spec.DeferTimeoutSeconds).To(Equal(autoscalingv1alpha1.DefaultDeferTimeoutSeconds))
-		})
+		for _, f := range cases.MustLoad[deferDefaultFile]("testdata/controller/crd-default") {
+			tc := f.Case
+			It(tc.Description, func() { runDeferDefault(tc) })
+		}
 	})
 
 	// These run against the envtest apiserver so the CRD's own union rules are
@@ -65,28 +68,20 @@ var _ = Describe("LLMScaler Controller", func() {
 			customFix = nil
 		})
 
-		DescribeTable("decision server", runCustomProviderReconcile, customProviderReconcileCases)
-		DescribeTable("admission", runCustomProviderAdmission, customProviderAdmissionCases)
+		DescribeTable("decision server", runCustomReconcileFile,
+			cases.Entries("testdata/controller/custom-reconcile", customReconcileDesc))
+		DescribeTable("admission", runCustomAdmissionFile,
+			cases.Entries("testdata/controller/custom-admission", customAdmissionDesc))
 
-		It("should default metricProvider to Prometheus and customProvider.path to /decisions", func() {
-			By("creating a scaler that leaves both unset")
-			scaler := customFix.newScaler()
-			scaler.Spec.MetricProvider = ""
-			scaler.Spec.CustomProvider = nil
-			scaler.Spec.Metrics = []autoscalingv1alpha1.MetricSpec{{Query: "avg(x)", Target: "1"}}
-			Expect(k8sClient.Create(customFix.ctx, scaler)).To(Succeed())
+		for _, f := range cases.MustLoad[customDefaultFile]("testdata/controller/custom-defaults") {
+			tc := f.Case
+			It(tc.Description, func() { runCustomDefault(tc) })
+		}
+	})
+})
 
-			stored := &autoscalingv1alpha1.LLMScaler{}
-			Expect(k8sClient.Get(customFix.ctx, customFix.scaler, stored)).To(Succeed())
-			Expect(stored.Spec.MetricProvider).To(Equal(autoscalingv1alpha1.MetricProviderPrometheus))
+func scalerDesc(c scalerCase) string { return c.Description }
 
-			By("switching it to the Custom provider without naming a path")
-			stored.Spec.MetricProvider = autoscalingv1alpha1.MetricProviderCustom
-			stored.Spec.CustomProvider = &autoscalingv1alpha1.CustomProviderSpec{ServiceID: sampleServiceID}
-			Expect(k8sClient.Update(customFix.ctx, stored)).To(Succeed())
+func customReconcileDesc(c customReconcileFile) string { return c.Description }
 
-			Expect(k8sClient.Get(customFix.ctx, customFix.scaler, stored)).To(Succeed())
-			Expect(stored.Spec.CustomProvider.Path).To(Equal(defaultDecisionsPath))
-		})
-	})
-})
+func customAdmissionDesc(c customAdmissionFile) string { return c.Description }
diff --git a/internal/controller/autoscaling/llmscaler_helpers_cases_test.go b/internal/controller/autoscaling/llmscaler_helpers_cases_test.go
deleted file mode 100644
index 5aeb195..0000000
--- a/internal/controller/autoscaling/llmscaler_helpers_cases_test.go
+++ /dev/null
@@ -1,199 +0,0 @@
-/*
-Copyright 2026.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-*/
-
-package controller
-
-import (
-	"time"
-
-	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
-)
-
-// deferTimeoutCases cover a nil spec (CRD default / in-memory object) and an
-// explicit 0, which must stay distinguishable so "never force" is opt-in.
-var deferTimeoutCases = []deferTimeoutCase{
-	{name: "nil spec", want: time.Duration(autoscalingv1alpha1.DefaultDeferTimeoutSeconds) * time.Second},
-	{name: "explicit 0", deferTimeout: ptr(int32(0)), want: 0},
-	{name: "explicit 30", deferTimeout: ptr(int32(30)), want: 30 * time.Second},
-}
-
-// pastDeferTimeoutCases pin the deadline: equal to the timeout forces, short of
-// it does not, and 0 never does. Elapsed is added to the runner's fixed base.
-var pastDeferTimeoutCases = []pastDeferTimeoutCase{
-	{name: "nil deferredSince", nilSince: true, timeout: time.Minute, elapsed: time.Hour, want: false},
-	{name: "timeout 0", timeout: 0, elapsed: 24 * time.Hour, want: false},
-	{name: "59s of a 60s timeout", timeout: time.Minute, elapsed: 59 * time.Second, want: false},
-	{name: "exactly the timeout", timeout: time.Minute, elapsed: time.Minute, want: true},
-	{name: "past the timeout", timeout: time.Minute, elapsed: time.Minute + time.Second, want: true},
-}
-
-// deferralRequeueAfterCases is how long a deferred sync waits. The runner passes
-// a 15s sync period, the controller default; rows whose want is 15s are the ones
-// that stay on that period.
-var deferralRequeueAfterCases = []deferralRequeueAfterCase{
-	{name: "deadline sooner than sync period", holding: true, timeout: time.Minute, until: 5 * time.Second, want: 5 * time.Second},
-	{name: "deadline later than sync period", holding: true, timeout: time.Minute, until: 30 * time.Second, want: 15 * time.Second},
-	{name: "deadline already passed", holding: true, timeout: time.Minute, until: -time.Second, want: time.Second},
-	{name: "deadline due this instant", holding: true, timeout: time.Minute, until: 0, want: time.Second},
-	{name: "deadline under the 1s floor", holding: true, timeout: time.Minute, until: 200 * time.Millisecond, want: time.Second},
-	{name: "timeout 0 stays on sync period", holding: true, timeout: 0, until: 5 * time.Second, want: 15 * time.Second},
-	{name: "not holding stays on sync period", holding: false, timeout: time.Minute, until: 5 * time.Second, want: 15 * time.Second},
-	{name: "pacing sooner than sync period", holding: true, timeout: time.Minute, until: -time.Second, pacing: 4 * time.Second, want: 4 * time.Second},
-	{name: "pacing later than sync period", holding: true, timeout: time.Minute, until: -time.Second, pacing: 30 * time.Second, want: 15 * time.Second},
-	{name: "pacing under the 1s floor", holding: true, timeout: time.Minute, until: -time.Second, pacing: 100 * time.Millisecond, want: time.Second},
-}
-
-// forcedWritePaceCases cover the per-process gap between forced writes.
-// wantRemaining nil leaves the remaining duration unchecked.
-var forcedWritePaceCases = []forcedWritePaceCase{
-	{name: "no record", recorded: false, sync: 15 * time.Second, at: 0, wantAllowed: true, wantRemaining: ptr(time.Duration(0))},
-	{name: "zero sync period", recorded: true, sync: 0, at: time.Hour, wantAllowed: true, wantRemaining: ptr(time.Duration(0))},
-	{name: "exactly one sync period", recorded: true, sync: 15 * time.Second, at: 15 * time.Second, wantAllowed: true},
-	{name: "just recorded", recorded: true, sync: 15 * time.Second, at: 0, wantAllowed: false},
-	{name: "14s into a 15s period", recorded: true, sync: 15 * time.Second, at: 14 * time.Second, wantAllowed: false, wantRemaining: ptr(time.Second)},
-}
-
-// stabilizationSequenceCases run an ordered list of recommend/hold ops on one
-// fresh reconciler per case. at is an offset from the runner's fixed base,
-// not the wall clock. TestStabilizeDesired and TestHoldAfterScaleDown are both
-// rows here, plus the window-0 reconciler.
-var stabilizationSequenceCases = []stabilizationCase{
-	{
-		// First recommendation sticks, a dip inside the window holds the peak,
-		// the peak expires after the window, and a scale-up is immediate.
-		name:   "peak holds inside the window then expires",
-		window: 60 * time.Second,
-		ops: []stabilizationOp{
-			{at: 0, recommend: ptr(int32(5)), want: ptr(int32(5))},
-			// Metric dipped to 1, but the peak of 5 is still within the window.
-			{at: 10 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(5))},
-			// After the window the peak has expired.
-			{at: 61 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(1))},
-			// Scale-up is immediate regardless of history.
-			{at: 61 * time.Second, recommend: ptr(int32(9)), want: ptr(int32(9))},
-		},
-	},
-	{
-		// Pinning the written count after a scale-down paces the next one a
-		// full window later, rather than letting the descent resolve off one
-		// collapsed reading.
-		name:   "pin after a scale-down paces the next step",
-		window: 300 * time.Second,
-		ops: []stabilizationOp{
-			// Fleet of 8, metric collapses; the peak is still in the window.
-			{at: 0, recommend: ptr(int32(8)), want: ptr(int32(8))},
-			// A window later the peak has expired and the low recommendation
-			// wins. The controller then writes 6 (an 8 -> 1 recommendation
-			// capped to a 2-replica step) and pins it.
-			{at: 301 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(1))},
-			{at: 301 * time.Second, holdAt: ptr(int32(6))},
-			// Within the window the fleet is held at 6 even though the metric
-			// still reads idle. Before the pin these would each have shrunk it.
-			{at: 311 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(6))},
-			{at: 400 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(6))},
-			{at: 600 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(6))},
-			// Once the pin ages out, the next step is allowed.
-			{at: 602 * time.Second, recommend: ptr(int32(1)), want: ptr(int32(1))},
-			// Scale-up is never blocked by the pin.
-			{at: 700 * time.Second, holdAt: ptr(int32(6))},
-			{at: 710 * time.Second, recommend: ptr(int32(9)), want: ptr(int32(9))},
-		},
-	},
-	{
-		// With stabilization off there is nothing to pace with, so the pin is a no-op.
-		name:   "window 0 makes the pin a no-op",
-		window: 0,
-		ops: []stabilizationOp{
-			{at: 0, holdAt: ptr(int32(6))},
-			{at: 0, recommend: ptr(int32(1)), want: ptr(int32(1))},
-		},
-	},
-}
-
-// parseTargetValueCases cover the target parser. It is the denominator of the
-// scaling ratio and must never yield zero, a negative, or a non-finite value.
-var parseTargetValueCases = []parseTargetValueCase{
-	{name: "plain integer", in: "5", want: 5},
-	{name: "ratio", in: "0.8", want: 0.8},
-	{name: "surrounding whitespace", in: "  2.5 ", want: 2.5},
-	{name: "percent suffix is rejected, not stripped", in: "80%", wantErr: true},
-	{name: "zero would make the ratio infinite", in: "0", wantErr: true},
-	{name: "negative would invert scaling", in: "-1", wantErr: true},
-	{name: "NaN parses as a float but is not usable", in: "NaN", wantErr: true},
-	{name: "Inf parses as a float but is not usable", in: "+Inf", wantErr: true},
-	{name: "not a number", in: "abc", wantErr: true},
-	{name: "empty", in: "", wantErr: true},
-}
-
-// queryPrometheusScalarCases resolve an instant-query vector: exactly one series
-// is a value, and anything else is an error the caller skips the metric on.
-var queryPrometheusScalarCases = []queryPrometheusScalarCase{
-	{
-		name:   "single series resolves",
-		result: `[{"metric":{},"value":[0,"1.5"]}]`,
-		want:   1.5,
-	},
-	{
-		name:            "empty vector is not zero",
-		result:          `[]`,
-		wantErrContains: []string{"returned no data"},
-	},
-	{
-		name: "an under-aggregated query is rejected, naming the series",
-		result: `[{"metric":{"__name__":"m","model":"qwen","backend":"10.0.0.1:8000"},"value":[0,"1"]},
-		          {"metric":{"__name__":"m","model":"qwen","backend":"10.0.0.2:8000"},"value":[0,"9"]}]`,
-		wantErrContains: []string{"returned 2 series"},
-	},
-	{
-		// The multi-series error has to point at the label to aggregate away,
-		// so it quotes both series with their labels sorted.
-		name: "multi-series error names both label sets",
-		result: `[
-			{"metric":{"model":"qwen","backend":"10.0.0.1:8000"},"value":[0,"1"]},
-			{"metric":{"model":"qwen","backend":"10.0.0.2:8000"},"value":[0,"9"]}]`,
-		wantErrContains: []string{
-			`{backend="10.0.0.1:8000",model="qwen"}`,
-			`{backend="10.0.0.2:8000",model="qwen"}`,
-			"avg(...)",
-		},
-	},
-}
-
-// computeDesiredFromMetricsCases cover a non-finite sample (not a measurement)
-// and the scale-from-zero baseline. A NaN or Inf sample must not set haveMetric:
-// histogram_quantile over an empty window returns NaN, and treating that as zero
-// would read as "scale to minReplicas". With no ready replicas a positive signal
-// is evaluated against a single-replica baseline, so the fleet wakes; a zero
-// signal still recommends 0.
-var computeDesiredFromMetricsCases = []computeDesiredCase{
-	{name: "NaN", sample: "NaN", ready: 4, metricName: "guardrail", query: "histogram_quantile(0.95, x)", target: "2"},
-	{name: "+Inf", sample: "+Inf", ready: 4, metricName: "guardrail", query: "histogram_quantile(0.95, x)", target: "2"},
-	{name: "-Inf", sample: "-Inf", ready: 4, metricName: "guardrail", query: "histogram_quantile(0.95, x)", target: "2"},
-	// ceil(1 * 0.85/0.5) = 2, not ceil(0 * 0.85/0.5) = 0.
-	{name: "positive signal wakes the fleet", sample: mockMetricValue, ready: 0, minReplicas: 0, maxReplicas: 5, metricName: "queue", query: "sum(queue_depth)", target: "0.5", wantHave: true, wantDesired: 2},
-	// A zero signal still recommends 0.
-	{name: "zero signal recommends zero", sample: "0", ready: 0, minReplicas: 0, maxReplicas: 5, metricName: "queue", query: "sum(queue_depth)", target: "0.5", wantHave: true, wantDesired: 0},
-}
-
-// scaleGuardCases are the scale-down rate limiter. Scale-up and a no-op pass through.
-var scaleGuardCases = []scaleGuardCase{
-	{name: "unlimited when maxStep is 0", current: 10, desired: 1, maxStep: 0, want: 1},
-	{name: "caps a burst scale-down", current: 10, desired: 1, maxStep: 2, want: 8},
-	{name: "passes through a step within the cap", current: 10, desired: 9, maxStep: 2, want: 9},
-	{name: "passes through a step exactly at the cap", current: 10, desired: 8, maxStep: 2, want: 8},
-	{name: "leaves scale-up untouched", current: 2, desired: 10, maxStep: 1, want: 10},
-	{name: "leaves a no-op untouched", current: 5, desired: 5, maxStep: 1, want: 5},
-}
diff --git a/internal/controller/autoscaling/llmscaler_helpers_test.go b/internal/controller/autoscaling/llmscaler_helpers_test.go
index 6f46897..cf8f681 100644
--- a/internal/controller/autoscaling/llmscaler_helpers_test.go
+++ b/internal/controller/autoscaling/llmscaler_helpers_test.go
@@ -29,6 +29,7 @@ import (
 	"k8s.io/apimachinery/pkg/types"
 
 	autoscalingv1alpha1 "github.com/modelsphere/llm-operator/api/autoscaling/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
 // pastDeferBase is the deferredSince instant for TestPastDeferTimeout.
@@ -47,288 +48,365 @@ var stabilizationKey = types.NamespacedName{Namespace: "ns", Name: "s"}
 // deferralRequeueSync is the sync period passed to every deferralRequeueAfter row.
 const deferralRequeueSync = 15 * time.Second
 
-type deferTimeoutCase struct {
-	name         string
-	deferTimeout *int32
-	want         time.Duration
+type deferTimeoutFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		DeferTimeoutSeconds *int32 `json:"deferTimeoutSeconds"`
+	} `json:"inputs"`
+	Expected struct {
+		Duration      cases.Duration `json:"duration"`
+		EqualsDefault bool           `json:"equalsDefault"`
+	} `json:"expected"`
 }
 
 func TestDeferTimeout(t *testing.T) {
-	for _, tc := range deferTimeoutCases {
-		t.Run(tc.name, func(t *testing.T) {
-			scaler := &autoscalingv1alpha1.LLMScaler{}
-			scaler.Spec.DeferTimeoutSeconds = tc.deferTimeout
-			got := deferTimeout(scaler)
-			if got != tc.want {
-				t.Fatalf("got %s, want %s", got, tc.want)
+	cases.Run(t, "testdata/helpers/defer-timeout", func(t *testing.T, tc deferTimeoutFile) {
+		scaler := &autoscalingv1alpha1.LLMScaler{}
+		scaler.Spec.DeferTimeoutSeconds = tc.Inputs.DeferTimeoutSeconds
+		got := deferTimeout(scaler)
+		if got != tc.Expected.Duration.Std() {
+			t.Fatalf("got %s, want %s", got, tc.Expected.Duration.Std())
+		}
+		if tc.Expected.EqualsDefault {
+			want := time.Duration(autoscalingv1alpha1.DefaultDeferTimeoutSeconds) * time.Second
+			if got != want {
+				t.Fatalf("got %s, want default %s", got, want)
 			}
-		})
-	}
+		}
+	})
 }
 
-type pastDeferTimeoutCase struct {
-	name     string
-	nilSince bool
-	timeout  time.Duration
-	elapsed  time.Duration
-	want     bool
+type pastDeferFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		NilSince bool           `json:"nilSince"`
+		Timeout  cases.Duration `json:"timeout"`
+		Elapsed  cases.Duration `json:"elapsed"`
+	} `json:"inputs"`
+	Expected struct {
+		Past bool `json:"past"`
+	} `json:"expected"`
 }
 
 func TestPastDeferTimeout(t *testing.T) {
-	for _, tc := range pastDeferTimeoutCases {
-		t.Run(tc.name, func(t *testing.T) {
-			var since *metav1.Time
-			if !tc.nilSince {
-				stamped := metav1.NewTime(pastDeferBase)
-				since = &stamped
-			}
-			got := pastDeferTimeout(since, tc.timeout, pastDeferBase.Add(tc.elapsed))
-			if got != tc.want {
-				t.Fatalf("got %v, want %v", got, tc.want)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/past-defer-timeout", func(t *testing.T, tc pastDeferFile) {
+		var since *metav1.Time
+		if !tc.Inputs.NilSince {
+			stamped := metav1.NewTime(pastDeferBase)
+			since = &stamped
+		}
+		now := pastDeferBase.Add(tc.Inputs.Elapsed.Std())
+		got := pastDeferTimeout(since, tc.Inputs.Timeout.Std(), now)
+		if got != tc.Expected.Past {
+			t.Fatalf("got %v, want %v", got, tc.Expected.Past)
+		}
+	})
 }
 
-type deferralRequeueAfterCase struct {
-	name    string
-	holding bool
-	timeout time.Duration
-	until   time.Duration
-	pacing  time.Duration
-	want    time.Duration
+type requeueFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Holding bool           `json:"holding"`
+		Timeout cases.Duration `json:"timeout"`
+		Until   cases.Duration `json:"until"`
+		Pacing  cases.Duration `json:"pacing"`
+	} `json:"inputs"`
+	Expected struct {
+		After cases.Duration `json:"after"`
+	} `json:"expected"`
 }
 
 func TestDeferralRequeueAfter(t *testing.T) {
-	for _, tc := range deferralRequeueAfterCases {
-		t.Run(tc.name, func(t *testing.T) {
-			got := deferralRequeueAfter(deferralRequeueSync, tc.holding, tc.timeout, tc.until, tc.pacing)
-			if got != tc.want {
-				t.Fatalf("got %s, want %s", got, tc.want)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/deferral-requeue-after", func(t *testing.T, tc requeueFile) {
+		in := tc.Inputs
+		got := deferralRequeueAfter(deferralRequeueSync, in.Holding, in.Timeout.Std(), in.Until.Std(), in.Pacing.Std())
+		if got != tc.Expected.After.Std() {
+			t.Fatalf("got %s, want %s", got, tc.Expected.After.Std())
+		}
+	})
 }
 
-type forcedWritePaceCase struct {
-	name          string
-	recorded      bool
-	sync          time.Duration
-	at            time.Duration
-	wantAllowed   bool
-	wantRemaining *time.Duration
+type forcedWriteFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Recorded bool           `json:"recorded"`
+		Sync     cases.Duration `json:"sync"`
+		At       cases.Duration `json:"at"`
+	} `json:"inputs"`
+	Expected struct {
+		Allowed   bool            `json:"allowed"`
+		Remaining *cases.Duration `json:"remaining"`
+	} `json:"expected"`
 }
 
 func TestForcedWritePace(t *testing.T) {
-	for _, tc := range forcedWritePaceCases {
-		t.Run(tc.name, func(t *testing.T) {
-			allowed, remaining := forcedWritePace(forcedWriteBase, tc.recorded, tc.sync, forcedWriteBase.Add(tc.at))
-			if allowed != tc.wantAllowed {
-				t.Fatalf("allowed=%v, want %v", allowed, tc.wantAllowed)
-			}
-			if tc.wantRemaining != nil && remaining != *tc.wantRemaining {
-				t.Fatalf("remaining=%s, want %s", remaining, *tc.wantRemaining)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/forced-write-pace", func(t *testing.T, tc forcedWriteFile) {
+		at := forcedWriteBase.Add(tc.Inputs.At.Std())
+		allowed, remaining := forcedWritePace(forcedWriteBase, tc.Inputs.Recorded, tc.Inputs.Sync.Std(), at)
+		if allowed != tc.Expected.Allowed {
+			t.Fatalf("allowed=%v, want %v", allowed, tc.Expected.Allowed)
+		}
+		if tc.Expected.Remaining != nil && remaining != tc.Expected.Remaining.Std() {
+			t.Fatalf("remaining=%s, want %s", remaining, tc.Expected.Remaining.Std())
+		}
+	})
+}
+
+type memoryStep struct {
+	Op string `json:"op"`
+}
+
+type memoryFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Namespace string       `json:"namespace"`
+		Scaler    string       `json:"scaler"`
+		At        string       `json:"at"`
+		Steps     []memoryStep `json:"steps"`
+	} `json:"inputs"`
+	Expected struct{} `json:"expected"`
 }
 
 // TestDeferralMemoryClearedOnEndAndDelete checks the two ways the in-memory
 // pace record goes away: the deferral ending, and the scaler being deleted.
 func TestDeferralMemoryClearedOnEndAndDelete(t *testing.T) {
-	r := &LLMScalerReconciler{}
-	key := types.NamespacedName{Namespace: "ns", Name: "s"}
-	now := time.Date(2026, 3, 1, 0, 0, 0, 0, time.UTC)
-	r.recordForcedWrite(key, now)
-	scaler := &autoscalingv1alpha1.LLMScaler{}
-	stamped := metav1.NewTime(now)
-	scaler.Status.DeferredSince = &stamped
-
-	r.endDeferral(key, scaler)
-	if scaler.Status.DeferredSince != nil {
-		t.Fatal("endDeferral left deferredSince set")
-	}
-	if _, ok := r.lastForcedWriteAt(key); ok {
-		t.Fatal("endDeferral left the forced-write record")
-	}
-
-	r.recordForcedWrite(key, now)
-	r.forgetRecommendations(key)
-	if _, ok := r.lastForcedWriteAt(key); ok {
-		t.Fatal("forgetRecommendations left the forced-write record")
-	}
+	cases.Run(t, "testdata/helpers/deferral-memory", func(t *testing.T, tc memoryFile) {
+		r := &LLMScalerReconciler{}
+		key := types.NamespacedName{Namespace: tc.Inputs.Namespace, Name: tc.Inputs.Scaler}
+		now, err := time.Parse(time.RFC3339, tc.Inputs.At)
+		if err != nil {
+			t.Fatalf("at: %v", err)
+		}
+		scaler := &autoscalingv1alpha1.LLMScaler{}
+		for _, step := range tc.Inputs.Steps {
+			switch step.Op {
+			case "record":
+				r.recordForcedWrite(key, now)
+			case "setSince":
+				stamped := metav1.NewTime(now)
+				scaler.Status.DeferredSince = &stamped
+			case "end":
+				r.endDeferral(key, scaler)
+			case "forget":
+				r.forgetRecommendations(key)
+			case "expectSinceCleared":
+				if scaler.Status.DeferredSince != nil {
+					t.Fatal("endDeferral left deferredSince set")
+				}
+			case "expectRecordCleared":
+				if _, ok := r.lastForcedWriteAt(key); ok {
+					t.Fatal("forced-write record still set")
+				}
+			default:
+				t.Fatalf("unknown step %q", step.Op)
+			}
+		}
+	})
+}
+
+type reasonFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+	} `json:"inputs"`
+	Expected struct {
+		Reason string `json:"reason"`
+	} `json:"expected"`
 }
 
 // TestReasonDeferralTimedOut pins the published condition reason, which users
 // match on in alerts and kubectl output.
 func TestReasonDeferralTimedOut(t *testing.T) {
-	if autoscalingv1alpha1.ReasonDeferralTimedOut != "DeferralTimedOut" {
-		t.Fatalf("ReasonDeferralTimedOut = %q, want DeferralTimedOut", autoscalingv1alpha1.ReasonDeferralTimedOut)
-	}
+	cases.Run(t, "testdata/helpers/reason-deferral-timed-out", func(t *testing.T, tc reasonFile) {
+		if autoscalingv1alpha1.ReasonDeferralTimedOut != tc.Expected.Reason {
+			t.Fatalf("ReasonDeferralTimedOut = %q, want %q", autoscalingv1alpha1.ReasonDeferralTimedOut, tc.Expected.Reason)
+		}
+	})
 }
 
-type stabilizationOp struct {
-	at        time.Duration
-	recommend *int32
-	holdAt    *int32
-	want      *int32
+type stabilizationOpFile struct {
+	At        cases.Duration `json:"at"`
+	Recommend *int32         `json:"recommend"`
+	HoldAt    *int32         `json:"holdAt"`
+	Want      *int32         `json:"want"`
 }
 
-type stabilizationCase struct {
-	name   string
-	window time.Duration
-	ops    []stabilizationOp
+type stabilizationFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Window cases.Duration        `json:"window"`
+		Ops    []stabilizationOpFile `json:"ops"`
+	} `json:"inputs"`
+	Expected struct{} `json:"expected"`
 }
 
 // TestStabilizeDesired runs each stabilization sequence on its own reconciler.
-// TestHoldAfterScaleDown's steps, including the window-0 reconciler, are rows
-// of the same table.
+// The window-0 pin is a row of the same table.
 func TestStabilizeDesired(t *testing.T) {
-	for _, tc := range stabilizationSequenceCases {
-		t.Run(tc.name, func(t *testing.T) {
-			r := &LLMScalerReconciler{}
-			for _, op := range tc.ops {
-				now := stabilizationBase.Add(op.at)
-				switch {
-				case op.holdAt != nil && op.recommend != nil:
-					t.Fatalf("op at %s sets both holdAt and recommend", op.at)
-				case op.holdAt != nil:
-					r.holdAfterScaleDown(stabilizationKey, *op.holdAt, tc.window, now)
-				case op.recommend != nil:
-					if op.want == nil {
-						t.Fatalf("recommend at %s has no want", op.at)
-					}
-					got := r.stabilizeDesired(stabilizationKey, *op.recommend, tc.window, now)
-					if got != *op.want {
-						t.Fatalf("at %s: got %d, want %d", op.at, got, *op.want)
-					}
-				default:
-					t.Fatalf("op at %s has neither holdAt nor recommend", op.at)
+	cases.Run(t, "testdata/helpers/stabilization", func(t *testing.T, tc stabilizationFile) {
+		r := &LLMScalerReconciler{}
+		window := tc.Inputs.Window.Std()
+		for _, op := range tc.Inputs.Ops {
+			now := stabilizationBase.Add(op.At.Std())
+			switch {
+			case op.HoldAt != nil && op.Recommend != nil:
+				t.Fatalf("op at %s sets both holdAt and recommend", op.At.Std())
+			case op.HoldAt != nil:
+				r.holdAfterScaleDown(stabilizationKey, *op.HoldAt, window, now)
+			case op.Recommend != nil:
+				if op.Want == nil {
+					t.Fatalf("recommend at %s has no want", op.At.Std())
+				}
+				got := r.stabilizeDesired(stabilizationKey, *op.Recommend, window, now)
+				if got != *op.Want {
+					t.Fatalf("at %s: got %d, want %d", op.At.Std(), got, *op.Want)
 				}
+			default:
+				t.Fatalf("op at %s has neither holdAt nor recommend", op.At.Std())
 			}
-		})
-	}
+		}
+	})
 }
 
-type parseTargetValueCase struct {
-	name    string
-	in      string
-	want    float64
-	wantErr bool
+type parseTargetFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Value string `json:"value"`
+	} `json:"inputs"`
+	Expected struct {
+		Value float64 `json:"value"`
+		Error bool    `json:"error"`
+	} `json:"expected"`
 }
 
 func TestParseTargetValue(t *testing.T) {
-	for _, tc := range parseTargetValueCases {
-		t.Run(tc.name, func(t *testing.T) {
-			got, err := parseTargetValue(tc.in)
-			if (err != nil) != tc.wantErr {
-				t.Fatalf("parseTargetValue(%q) error = %v, wantErr %v", tc.in, err, tc.wantErr)
-			}
-			if err == nil && got != tc.want {
-				t.Fatalf("parseTargetValue(%q) = %v, want %v", tc.in, got, tc.want)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/parse-target-value", func(t *testing.T, tc parseTargetFile) {
+		got, err := parseTargetValue(tc.Inputs.Value)
+		if (err != nil) != tc.Expected.Error {
+			t.Fatalf("parseTargetValue(%q) error = %v, wantErr %v", tc.Inputs.Value, err, tc.Expected.Error)
+		}
+		if err == nil && got != tc.Expected.Value {
+			t.Fatalf("parseTargetValue(%q) = %v, want %v", tc.Inputs.Value, got, tc.Expected.Value)
+		}
+	})
 }
 
-type queryPrometheusScalarCase struct {
-	name            string
-	result          string
-	want            float64
-	wantErrContains []string
+type promScalarFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Result string `json:"result"`
+	} `json:"inputs"`
+	Expected struct {
+		Value         float64  `json:"value"`
+		ErrorContains []string `json:"errorContains"`
+	} `json:"expected"`
 }
 
 func TestQueryPrometheusScalar(t *testing.T) {
-	for _, tc := range queryPrometheusScalarCases {
-		t.Run(tc.name, func(t *testing.T) {
-			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-				_, _ = fmt.Fprintf(w, `{"status":"success","data":{"result":%s}}`, tc.result)
-			}))
-			defer srv.Close()
-
-			got, err := queryPrometheusScalar(srv.URL, "some_query", nil)
-			if len(tc.wantErrContains) == 0 {
-				if err != nil {
-					t.Fatalf("unexpected error: %v", err)
-				}
-				if got != tc.want {
-					t.Fatalf("got %v, want %v", got, tc.want)
-				}
-				return
+	cases.Run(t, "testdata/helpers/query-prometheus-scalar", func(t *testing.T, tc promScalarFile) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			_, _ = fmt.Fprintf(w, `{"status":"success","data":{"result":%s}}`, tc.Inputs.Result)
+		}))
+		defer srv.Close()
+
+		got, err := queryPrometheusScalar(srv.URL, "some_query", nil)
+		if len(tc.Expected.ErrorContains) == 0 {
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
 			}
-			if err == nil {
-				t.Fatalf("expected error containing %q, got value %v", tc.wantErrContains, got)
+			if got != tc.Expected.Value {
+				t.Fatalf("got %v, want %v", got, tc.Expected.Value)
 			}
-			for _, want := range tc.wantErrContains {
-				if !strings.Contains(err.Error(), want) {
-					t.Errorf("error = %q, want it to contain %q", err, want)
-				}
+			return
+		}
+		if err == nil {
+			t.Fatalf("expected error containing %q, got value %v", tc.Expected.ErrorContains, got)
+		}
+		for _, want := range tc.Expected.ErrorContains {
+			if !strings.Contains(err.Error(), want) {
+				t.Errorf("error = %q, want it to contain %q", err, want)
 			}
-		})
-	}
+		}
+	})
 }
 
-type computeDesiredCase struct {
-	name        string
-	sample      string
-	ready       int64
-	minReplicas int32
-	maxReplicas int32
-	metricName  string
-	query       string
-	target      string
-	wantHave    bool
-	wantDesired int32
+type computeDesiredFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Sample      string `json:"sample"`
+		Ready       int64  `json:"ready"`
+		MinReplicas int32  `json:"minReplicas"`
+		MaxReplicas int32  `json:"maxReplicas"`
+		MetricName  string `json:"metricName"`
+		Query       string `json:"query"`
+		Target      string `json:"target"`
+	} `json:"inputs"`
+	Expected struct {
+		Have    bool  `json:"have"`
+		Desired int32 `json:"desired"`
+	} `json:"expected"`
 }
 
 // TestComputeDesiredFromMetrics covers non-finite samples and scale-from-zero.
 func TestComputeDesiredFromMetrics(t *testing.T) {
-	for _, tc := range computeDesiredFromMetricsCases {
-		t.Run(tc.name, func(t *testing.T) {
-			srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-				_, _ = fmt.Fprintf(w,
-					`{"status":"success","data":{"result":[{"metric":{},"value":[0,"%s"]}]}}`, tc.sample)
-			}))
-			defer srv.Close()
-
-			scaler := &autoscalingv1alpha1.LLMScaler{
-				Spec: autoscalingv1alpha1.LLMScalerSpec{
-					ServerAddress: srv.URL,
-					MinReplicas:   tc.minReplicas,
-					MaxReplicas:   tc.maxReplicas,
-					Metrics: []autoscalingv1alpha1.MetricSpec{{
-						Name:   tc.metricName,
-						Query:  tc.query,
-						Target: tc.target,
-					}},
-				},
-			}
-			desired, haveMetric := (&LLMScalerReconciler{}).computeDesiredFromMetrics(context.Background(), scaler, tc.ready)
-			if haveMetric != tc.wantHave {
-				t.Errorf("haveMetric = %v, want %v", haveMetric, tc.wantHave)
-			}
-			if desired != tc.wantDesired {
-				t.Errorf("desired = %d, want %d", desired, tc.wantDesired)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/compute-desired", func(t *testing.T, tc computeDesiredFile) {
+		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			_, _ = fmt.Fprintf(w,
+				`{"status":"success","data":{"result":[{"metric":{},"value":[0,"%s"]}]}}`, tc.Inputs.Sample)
+		}))
+		defer srv.Close()
+
+		scaler := &autoscalingv1alpha1.LLMScaler{
+			Spec: autoscalingv1alpha1.LLMScalerSpec{
+				ServerAddress: srv.URL,
+				MinReplicas:   tc.Inputs.MinReplicas,
+				MaxReplicas:   tc.Inputs.MaxReplicas,
+				Metrics: []autoscalingv1alpha1.MetricSpec{{
+					Name:   tc.Inputs.MetricName,
+					Query:  tc.Inputs.Query,
+					Target: tc.Inputs.Target,
+				}},
+			},
+		}
+		desired, haveMetric := (&LLMScalerReconciler{}).computeDesiredFromMetrics(
+			context.Background(), scaler, tc.Inputs.Ready)
+		if haveMetric != tc.Expected.Have {
+			t.Errorf("haveMetric = %v, want %v", haveMetric, tc.Expected.Have)
+		}
+		if desired != tc.Expected.Desired {
+			t.Errorf("desired = %d, want %d", desired, tc.Expected.Desired)
+		}
+	})
 }
 
-type scaleGuardCase struct {
-	name    string
-	current int32
-	desired int32
-	maxStep int32
-	want    int32
+type scaleGuardFile struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Current int32 `json:"current"`
+		Desired int32 `json:"desired"`
+		MaxStep int32 `json:"maxStep"`
+	} `json:"inputs"`
+	Expected struct {
+		Want int32 `json:"want"`
+	} `json:"expected"`
 }
 
 func TestScaleGuard(t *testing.T) {
-	for _, tc := range scaleGuardCases {
-		t.Run(tc.name, func(t *testing.T) {
-			got := scaleGuard(tc.current, tc.desired, tc.maxStep)
-			if got != tc.want {
-				t.Fatalf("scaleGuard(%d, %d, %d) = %d, want %d", tc.current, tc.desired, tc.maxStep, got, tc.want)
-			}
-		})
-	}
+	cases.Run(t, "testdata/helpers/scale-guard", func(t *testing.T, tc scaleGuardFile) {
+		in := tc.Inputs
+		got := scaleGuard(in.Current, in.Desired, in.MaxStep)
+		if got != tc.Expected.Want {
+			t.Fatalf("scaleGuard(%d, %d, %d) = %d, want %d", in.Current, in.Desired, in.MaxStep, got, tc.Expected.Want)
+		}
+	})
 }
diff --git a/internal/controller/autoscaling/llmscaler_scenario_test.go b/internal/controller/autoscaling/llmscaler_scenario_test.go
index 8521afa..554c257 100644
--- a/internal/controller/autoscaling/llmscaler_scenario_test.go
+++ b/internal/controller/autoscaling/llmscaler_scenario_test.go
@@ -93,14 +93,6 @@ type sinceExpect struct {
 	offset time.Duration
 }
 
-func sinceAt(offset time.Duration) sinceExpect {
-	return sinceExpect{kind: sinceOffset, offset: offset}
-}
-
-func sinceNil() sinceExpect { return sinceExpect{kind: sinceNone} }
-
-func sinceSeed() sinceExpect { return sinceExpect{kind: sinceSeeded} }
-
 // scalerSpecPatch is applied to the fixture LLMScaler before the steps.
 // A nil field is left at the fixture value. DeferTimeoutSeconds is a pointer
 // on the API, so a pointer to 0 is "never force" and nil means "don't patch".
diff --git a/internal/controller/autoscaling/testdata/controller/crd-default/should-default-defertimeoutseconds-to-600.yaml b/internal/controller/autoscaling/testdata/controller/crd-default/should-default-defertimeoutseconds-to-600.yaml
new file mode 100644
index 0000000..41cf7bb
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/crd-default/should-default-defertimeoutseconds-to-600.yaml
@@ -0,0 +1,6 @@
+# The fixture leaves the field unset; the CRD default fills it.
+"name": "should-default-defertimeoutseconds-to-600"
+"description": "should default deferTimeoutSeconds to 600"
+"inputs": {}
+"expected":
+  "seconds": 600
diff --git a/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-custom-scaler-with-no-customprovider.yaml b/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-custom-scaler-with-no-customprovider.yaml
new file mode 100644
index 0000000..a83874c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-custom-scaler-with-no-customprovider.yaml
@@ -0,0 +1,6 @@
+"name": "should-reject-a-custom-scaler-with-no-customprovider"
+"description": "should reject a Custom scaler with no customProvider"
+"inputs":
+  "omitCustomProvider": true
+"expected":
+  "errorContains": "spec.customProvider is required"
diff --git a/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-prometheus-scaler-with-no-metrics.yaml b/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-prometheus-scaler-with-no-metrics.yaml
new file mode 100644
index 0000000..2dff606
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/custom-admission/should-reject-a-prometheus-scaler-with-no-metrics.yaml
@@ -0,0 +1,8 @@
+"name": "should-reject-a-prometheus-scaler-with-no-metrics"
+"description": "should reject a Prometheus scaler with no metrics"
+"inputs":
+  "metricProvider": "Prometheus"
+  "omitCustomProvider": true
+  "pinProvider": true
+"expected":
+  "errorContains": "spec.metrics must be non-empty"
diff --git a/internal/controller/autoscaling/testdata/controller/custom-defaults/should-default-metricprovider-to-prometheus-and-customprovider-path-to-decisions.yaml b/internal/controller/autoscaling/testdata/controller/custom-defaults/should-default-metricprovider-to-prometheus-and-customprovider-path-to-decisions.yaml
new file mode 100644
index 0000000..2f73a20
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/custom-defaults/should-default-metricprovider-to-prometheus-and-customprovider-path-to-decisions.yaml
@@ -0,0 +1,11 @@
+"name": "should-default-metricprovider-to-prometheus-and-customprovider-path-to-decisions"
+"description": "should default metricProvider to Prometheus and customProvider.path to /decisions"
+"inputs":
+  "createBy": "creating a scaler that leaves both unset"
+  "switchBy": "switching it to the Custom provider without naming a path"
+  "query": "avg(x)"
+  "target": "1"
+  "serviceId": "fallback-modelforge-01"
+"expected":
+  "metricProvider": "Prometheus"
+  "path": "/decisions"
diff --git a/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-hold-replicas-when-the-decision-server-is-unreachable.yaml b/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-hold-replicas-when-the-decision-server-is-unreachable.yaml
new file mode 100644
index 0000000..6537faa
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-hold-replicas-when-the-decision-server-is-unreachable.yaml
@@ -0,0 +1,20 @@
+"name": "should-hold-replicas-when-the-decision-server-is-unreachable"
+"description": "should hold replicas when the decision server is unreachable"
+"inputs":
+  "deploy":
+    "replicas": 3
+    "updated": 3
+    "ready": 3
+    "available": 3
+    "spec": 3
+  "deployBy": "putting the Deployment at 3 replicas, above minReplicas"
+  "closeServer": true
+  "closeBy": "taking the decision server down"
+  "reconcileBy": "running the Reconciler"
+  "replicasBy": "checking the Deployment was left where it was, not collapsed to minReplicas"
+  "statusBy": "checking the failure is reported on the status"
+"expected":
+  "replicas": 3
+  "decisionStatus": "False"
+  "decisionReason": "DecisionUnavailable"
+  "decisionMessageNonEmpty": true
diff --git a/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-scale-the-deployment-to-the-count-the-decision-server-reports.yaml b/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-scale-the-deployment-to-the-count-the-decision-server-reports.yaml
new file mode 100644
index 0000000..734a302
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/custom-reconcile/should-scale-the-deployment-to-the-count-the-decision-server-reports.yaml
@@ -0,0 +1,12 @@
+# active 4 is an absolute count, so 1 ready replica scales straight to 4.
+"name": "should-scale-the-deployment-to-the-count-the-decision-server-reports"
+"description": "should scale the Deployment to the count the decision server reports"
+"inputs":
+  "createBy": "creating an LLMScaler with no metrics at all"
+  "reconcileBy": "running the Reconciler"
+  "replicasBy": "checking the Deployment landed on the decision, not on a ratio of it"
+  "statusBy": "checking the status reports the same recommendation"
+"expected":
+  "replicas": 4
+  "desired": 4
+  "decisionStatus": "True"
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-status/should-clear-a-deferral-on-a-no-op-once-no-guard-matches.yaml b/internal/controller/autoscaling/testdata/controller/deferral-status/should-clear-a-deferral-on-a-no-op-once-no-guard-matches.yaml
new file mode 100644
index 0000000..2a3182d
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-status/should-clear-a-deferral-on-a-no-op-once-no-guard-matches.yaml
@@ -0,0 +1,25 @@
+# ceil(1 * 0.5/0.5) = 1, which is already spec, and the rollout is done, so no guard matches.
+"name": "should-clear-a-deferral-on-a-no-op-once-no-guard-matches"
+"description": "should clear a deferral on a no-op once no guard matches"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+  "rollout": true
+  "steps":
+  - "at": "0s"
+  - "by": "finishing the rollout at the replica count the metric already wants"
+    "at": "10s"
+    "metric": "0.5"
+    "observeSpec": 1
+"expected":
+  "steps":
+  - "deferredSince": "0s"
+  - "replicas": 1
+    "deferredSince": null
+    "applied": 1
+    "forcedWrite": false
+    "condition":
+      "status": "False"
+      "reason": "NotDeferred"
+      "messageContains":
+      - "already at the step target"
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-deferredsince-across-no-op-syncs-while-a-rollout-guard-still-matches.yaml b/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-deferredsince-across-no-op-syncs-while-a-rollout-guard-still-matches.yaml
new file mode 100644
index 0000000..b23401e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-deferredsince-across-no-op-syncs-while-a-rollout-guard-still-matches.yaml
@@ -0,0 +1,64 @@
+"name": "should-keep-deferredsince-across-no-op-syncs-while-a-rollout-guard-still-matches"
+"description": "should keep deferredSince across no-op syncs while a rollout guard still matches"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+  "rollout": true
+  "steps":
+  - "by": "not starting a deferral on a no-op, even though the rollout guard matches"
+    "at": "0s"
+    "metric": "0.5"
+  - "by": "stamping deferredSince on the first sync that actually wants to write"
+    "at": "10s"
+    "metric": "0.85"
+  - "by": "alternating a no-op with a held scale-up; the no-ops must not move the clock"
+    "at": "20s"
+    "metric": "0.5"
+  - "at": "30s"
+    "metric": "0.85"
+  - "at": "40s"
+    "metric": "0.5"
+  - "by": "forcing the held scale-up once the original deadline passes"
+    "at": "70s"
+    "metric": "0.85"
+"expected":
+  "steps":
+  - "deferredSince": null
+    "replicas": 1
+    "condition":
+      "status": "False"
+      "reason": "NotDeferred"
+  - "deferredSince": "10s"
+    "replicas": 1
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+  - "replicas": 1
+    "deferredSince": "10s"
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+      "messageContains": &id001
+      - "nothing to write this sync"
+      - "guard is still active"
+      - "deferral clock keeps running"
+      - "2026-03-01T12:00:10Z"
+      - "Deployment rollout in progress"
+  - "replicas": 1
+    "deferredSince": "10s"
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+  - "replicas": 1
+    "deferredSince": "10s"
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+      "messageContains": *id001
+  - "replicas": 2
+    "deferredSince": "10s"
+    "condition":
+      "status": "False"
+      "reason": "DeferralTimedOut"
+      "messageContains":
+      - "scaling 1 -> 2 anyway"
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-one-deferral-when-the-guard-reason-changes.yaml b/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-one-deferral-when-the-guard-reason-changes.yaml
new file mode 100644
index 0000000..4fce1e1
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-status/should-keep-one-deferral-when-the-guard-reason-changes.yaml
@@ -0,0 +1,40 @@
+"name": "should-keep-one-deferral-when-the-guard-reason-changes"
+"description": "should keep one deferral when the guard reason changes"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+  "metric": "0"
+  "deploy":
+    "replicas": 5
+    "updated": 5
+    "ready": 3
+    "available": 3
+    "spec": 5
+  "rollout": true
+  "steps":
+  - "at": "0s"
+  - "by": "finishing the rollout while the fleet is still short of ready"
+    "at": "10s"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 3
+      "available": 3
+  - "by": "not emitting again while the same guard still defers"
+    "at": "10s"
+"expected":
+  "steps":
+  - "deferredSince": "0s"
+    "condition":
+      "reason": "RolloutInProgress"
+  - "replicas": 5
+    "deferredSince": "0s"
+    "condition":
+      "reason": "ScaleDownSettling"
+    "events":
+    - "contains":
+      - "Deployment rollout in progress"
+    - "contains":
+      - "previous scale-down has not settled"
+  - "deferredSince": "0s"
+    "noMoreEvents": true
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-deferral-that-already-exceeded-the-timeout-on-a-fresh-reconciler.yaml b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-deferral-that-already-exceeded-the-timeout-on-a-fresh-reconciler.yaml
new file mode 100644
index 0000000..273eea2
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-deferral-that-already-exceeded-the-timeout-on-a-fresh-reconciler.yaml
@@ -0,0 +1,23 @@
+# deferredSince lives in status, so a fresh reconciler still enforces the deadline.
+"name": "should-force-a-deferral-that-already-exceeded-the-timeout-on-a-fresh-reconciler"
+"description": "should force a deferral that already exceeded the timeout on a fresh reconciler"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+  "rollout": true
+  "seedDeferredSince": "-2h"
+  "expectNilRecommendations": true
+  "steps":
+  - "at": "0s"
+"expected":
+  "steps":
+  - "replicas": 2
+    "deferredSince": "seed"
+    "applied": 2
+    "condition":
+      "status": "False"
+      "reason": "DeferralTimedOut"
+    "events":
+    - "prefix": "Warning DeferralTimedOut"
+      "contains":
+      - "1 -> 2"
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-down-of-an-unready-fleet-once-the-deferral-times-out.yaml b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-down-of-an-unready-fleet-once-the-deferral-times-out.yaml
new file mode 100644
index 0000000..d108f09
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-down-of-an-unready-fleet-once-the-deferral-times-out.yaml
@@ -0,0 +1,52 @@
+# readyReplicas 0 is omitted from status and reads as missing. available is 0 too, because availableReplicas cannot exceed readyReplicas.
+"name": "should-force-a-scale-down-of-an-unready-fleet-once-the-deferral-times-out"
+"description": "should force a scale-down of an unready fleet once the deferral times out"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+    "maxStepReplicas": 2
+  "prelude": "capping the step and parking 5 replicas with none ready"
+  "deploy":
+    "replicas": 5
+    "updated": 5
+    "ready": 0
+    "available": 0
+    "spec": 5
+  "metric": "0"
+  "steps":
+  - "at": "0s"
+  - "by": "keeping the same deferral until the timeout"
+    "at": "59s"
+  - "by": "forcing the capped step, not the uncapped recommendation"
+    "at": "60s"
+"expected":
+  "steps":
+  - "replicas": 5
+    "deferredSince": "0s"
+    "desired": 1
+    "current": 0
+    "applied": 5
+    "condition":
+      "status": "True"
+      "reason": "ScaleDownSettling"
+      "messageContains":
+      - "holding 5 replicas (wanted 3)"
+      - "will apply after 60s"
+  - "replicas": 5
+    "deferredSince": "0s"
+  - "replicas": 3
+    "deferredSince": "0s"
+    "desired": 1
+    "applied": 3
+    "current": 0
+    "condition":
+      "status": "False"
+      "reason": "DeferralTimedOut"
+      "messageContains":
+      - "scaling 5 -> 3 anyway"
+    "events":
+    - "prefix": "Normal ScalingDeferred"
+    - "prefix": "Warning DeferralTimedOut"
+      "contains":
+      - "previous scale-down has not settled"
+      - "5 -> 3"
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-up-once-a-stuck-rollout-exceeds-the-deferral-timeout.yaml b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-up-once-a-stuck-rollout-exceeds-the-deferral-timeout.yaml
new file mode 100644
index 0000000..a97d880
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-force-a-scale-up-once-a-stuck-rollout-exceeds-the-deferral-timeout.yaml
@@ -0,0 +1,55 @@
+"name": "should-force-a-scale-up-once-a-stuck-rollout-exceeds-the-deferral-timeout"
+"description": "should force a scale-up once a stuck rollout exceeds the deferral timeout"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+  "rollout": true
+  "steps":
+  - "by": "deferring the first sync and stamping deferredSince"
+    "at": "0s"
+  - "by": "reconciling again before the timeout, without restamping or rewriting"
+    "at": "59s"
+  - "by": "advancing past the timeout and forcing the held scale-up"
+    "at": "60s"
+"expected":
+  "steps":
+  - "requeueAfter": "15s"
+    "replicas": 1
+    "deferredSince": "0s"
+    "desired": 2
+    "applied": 1
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+      "messageContains":
+      - "holding 1 replicas (wanted 2)"
+      - "Deployment rollout in progress"
+      - "2026-03-01T12:00:00Z"
+      - "will apply after 60s"
+      "observedGenerationCurrent": true
+    "events":
+    - "prefix": "Normal ScalingDeferred"
+      "contains":
+      - "holding 1 replicas (wanted 2)"
+  - "requeueAfter": "1s"
+    "replicas": 1
+    "deferredSince": "0s"
+    "applied": 1
+    "noMoreEvents": true
+  - "replicas": 2
+    "deferredSince": "0s"
+    "desired": 2
+    "applied": 2
+    "requeueAfter": "15s"
+    "condition":
+      "status": "False"
+      "reason": "DeferralTimedOut"
+      "messageContains":
+      - "scaling 1 -> 2 anyway"
+      "observedGenerationCurrent": true
+    "events":
+    - "prefix": "Warning DeferralTimedOut"
+      "contains":
+      - "Deployment rollout in progress"
+      - "1 -> 2"
+    "noMoreEvents": true
diff --git a/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-never-force-a-deferral-when-defertimeoutseconds-is-0.yaml b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-never-force-a-deferral-when-defertimeoutseconds-is-0.yaml
new file mode 100644
index 0000000..6cfde4a
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/deferral-timeout/should-never-force-a-deferral-when-defertimeoutseconds-is-0.yaml
@@ -0,0 +1,22 @@
+"name": "should-never-force-a-deferral-when-defertimeoutseconds-is-0"
+"description": "should never force a deferral when deferTimeoutSeconds is 0"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 0
+  "rollout": true
+  "seedDeferredSince": "-24h"
+  "steps":
+  - "at": "0s"
+"expected":
+  "steps":
+  - "requeueAfter": "15s"
+    "replicas": 1
+    "deferredSince": "seed"
+    "condition":
+      "status": "True"
+      "reason": "RolloutInProgress"
+      "messageContains":
+      - "never forced (deferTimeoutSeconds is 0)"
+    "events":
+    - "prefix": "Normal ScalingDeferred"
+    "noMoreEvents": true
diff --git a/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-clear-deferredsince-and-the-forced-write-pace-when-no-guard-matches.yaml b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-clear-deferredsince-and-the-forced-write-pace-when-no-guard-matches.yaml
new file mode 100644
index 0000000..e69bfe5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-clear-deferredsince-and-the-forced-write-pace-when-no-guard-matches.yaml
@@ -0,0 +1,33 @@
+"name": "should-clear-deferredsince-and-the-forced-write-pace-when-no-guard-matches"
+"description": "should clear deferredSince and the forced-write pace when no guard matches"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+    "maxStepReplicas": 2
+  "metric": "0"
+  "deploy":
+    "replicas": 5
+    "updated": 5
+    "ready": 0
+    "available": 0
+    "spec": 5
+  "steps":
+  - "at": "0s"
+  - "at": "60s"
+  - "by": "letting the fleet become ready at the replica count just written"
+    "at": "60s"
+    "observeSpec": 3
+"expected":
+  "steps":
+  - {}
+  - "replicas": 3
+    "forcedWrite": true
+    "deferredSince": "0s"
+  - "replicas": 1
+    "deferredSince": null
+    "forcedWrite": false
+    "condition":
+      "status": "False"
+      "reason": "NotDeferred"
+      "messageContains":
+      - "scaled 3 -> 1"
diff --git a/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-force-the-next-stuck-scale-down-step-after-one-sync-period.yaml b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-force-the-next-stuck-scale-down-step-after-one-sync-period.yaml
new file mode 100644
index 0000000..5a52f47
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-force-the-next-stuck-scale-down-step-after-one-sync-period.yaml
@@ -0,0 +1,41 @@
+"name": "should-force-the-next-stuck-scale-down-step-after-one-sync-period"
+"description": "should force the next stuck scale-down step after one sync period"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+    "syncPeriodSeconds": 15
+    "maxStepReplicas": 2
+  "prelude": "capping the step and parking 5 replicas with none ready"
+  "deploy":
+    "replicas": 5
+    "updated": 5
+    "ready": 0
+    "available": 0
+    "spec": 5
+  "metric": "0"
+  "steps":
+  - "at": "0s"
+  - "by": "forcing the first step once the timeout elapses"
+    "at": "60s"
+  - "by": "keeping the fleet unready and taking the next step one sync period later, without another timeout"
+    "at": "75s"
+    "observeSpec": 0
+"expected":
+  "steps":
+  - "replicas": 5
+    "deferredSince": "0s"
+    "condition":
+      "reason": "ScaleDownSettling"
+  - "replicas": 3
+    "deferredSince": "0s"
+    "forcedWrite": true
+    "forcedWriteAt": "60s"
+  - "replicas": 1
+    "deferredSince": "0s"
+    "applied": 1
+    "condition":
+      "status": "False"
+      "reason": "DeferralTimedOut"
+      "messageContains":
+      - "scaling 3 -> 1 anyway"
+      - "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-not-take-a-second-forced-step-until-a-sync-period-has-elapsed.yaml b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-not-take-a-second-forced-step-until-a-sync-period-has-elapsed.yaml
new file mode 100644
index 0000000..6e348a5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/forced-write-pacing/should-not-take-a-second-forced-step-until-a-sync-period-has-elapsed.yaml
@@ -0,0 +1,47 @@
+"name": "should-not-take-a-second-forced-step-until-a-sync-period-has-elapsed"
+"description": "should not take a second forced step until a sync period has elapsed"
+"inputs":
+  "spec":
+    "deferTimeoutSeconds": 60
+    "syncPeriodSeconds": 15
+    "maxStepReplicas": 2
+  "metric": "0"
+  "deploy":
+    "replicas": 5
+    "updated": 5
+    "ready": 0
+    "available": 0
+    "spec": 5
+  "steps":
+  - "by": "forcing one scale-down step of an unready fleet"
+    "at": "0s"
+    "snapshotStatus": true
+  - "at": "60s"
+  - "by": "putting the status back to its pre-write state, as if that update had conflicted"
+    "restoreStatus": true
+    "skipReconcile": true
+  - "by": "reconciling again with the clock unchanged, which must not take the next step"
+    "at": "60s"
+  - "by": "still holding one second short of the sync period"
+    "at": "74s"
+  - "by": "forcing the next step once the sync period has elapsed"
+    "at": "75s"
+"expected":
+  "steps":
+  - {}
+  - "replicas": 3
+  - {}
+  - "replicas": 3
+    "requeueAfter": "15s"
+    "deferredSince": "0s"
+    "applied": 3
+    "condition":
+      "status": "True"
+      "messageContains":
+      - "forced write paced to one per sync period"
+  - "replicas": 3
+    "requeueAfter": "1s"
+    "deferredSince": "0s"
+  - "replicas": 1
+    "deferredSince": "0s"
+    "applied": 1
diff --git a/internal/controller/autoscaling/testdata/controller/metric-scaling/should-keep-the-ready-replica-fallback-for-the-scaling-ratio-only.yaml b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-keep-the-ready-replica-fallback-for-the-scaling-ratio-only.yaml
new file mode 100644
index 0000000..c182d5a
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-keep-the-ready-replica-fallback-for-the-scaling-ratio-only.yaml
@@ -0,0 +1,22 @@
+# readyReplicas is 0, so the ratio falls back to spec (3): ceil(3 * 0.85/0.5) = 6, clamped to maxReplicas 5. status.currentReplicas stays 0.
+"name": "should-keep-the-ready-replica-fallback-for-the-scaling-ratio-only"
+"description": "should keep the ready-replica fallback for the scaling ratio only"
+"inputs":
+  "plainReconciler": true
+  "steps":
+  - "deploy":
+      "replicas": 3
+      "updated": 3
+      "ready": 0
+      "available": 0
+      "spec": 3
+"expected":
+  "steps":
+  - "replicas": 5
+    "current": 0
+    "desired": 5
+    "applied": 5
+    "deferredSince": null
+    "condition":
+      "status": "False"
+      "reason": "NotDeferred"
diff --git a/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-minreplicas-when-the-metric-is-zero.yaml b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-minreplicas-when-the-metric-is-zero.yaml
new file mode 100644
index 0000000..2713d3e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-minreplicas-when-the-metric-is-zero.yaml
@@ -0,0 +1,23 @@
+# desired = ceil(5 * 0/0.5) = 0, clamped up to minReplicas (1).
+"name": "should-scale-down-to-minreplicas-when-the-metric-is-zero"
+"description": "should scale down to minReplicas when the metric is zero"
+"inputs":
+  "plainReconciler": true
+  "steps":
+  - "by": "setting the Deployment to 5 replicas and the metric to 0 (idle)"
+    "skipReconcile": true
+    "metric": "0"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "spec": 5
+  - "by": "running the Reconciler"
+  - "by": "checking it scaled down to minReplicas"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - {}
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-zero-when-minreplicas-is-0.yaml b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-zero-when-minreplicas-is-0.yaml
new file mode 100644
index 0000000..a12d78c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-scale-down-to-zero-when-minreplicas-is-0.yaml
@@ -0,0 +1,26 @@
+# desired = ceil(2 * 0/0.5) = 0, clamped to minReplicas (0).
+"name": "should-scale-down-to-zero-when-minreplicas-is-0"
+"description": "should scale down to zero when minReplicas is 0"
+"inputs":
+  "plainReconciler": true
+  "prelude": "lowering the scaler's minReplicas to 0"
+  "spec":
+    "minReplicas": 0
+  "steps":
+  - "by": "setting the Deployment to 2 replicas, fully rolled out"
+    "skipReconcile": true
+    "metric": "0"
+    "deploy":
+      "replicas": 2
+      "updated": 2
+      "ready": 2
+      "available": 2
+      "spec": 2
+  - "by": "running the Reconciler"
+  - "by": "checking it scaled down to zero"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - {}
+  - "replicas": 0
diff --git a/internal/controller/autoscaling/testdata/controller/metric-scaling/should-successfully-scale-the-deployment-based-on-metrics.yaml b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-successfully-scale-the-deployment-based-on-metrics.yaml
new file mode 100644
index 0000000..f8e2fc6
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-successfully-scale-the-deployment-based-on-metrics.yaml
@@ -0,0 +1,13 @@
+# ceil(1 * 0.85/0.5) = ceil(1.7) = 2. The mock sample is 0.85 and the target is 0.5.
+"name": "should-successfully-scale-the-deployment-based-on-metrics"
+"description": "should successfully scale the deployment based on metrics"
+"inputs":
+  "plainReconciler": true
+  "steps":
+  - "by": "running the Reconciler"
+  - "by": "checking if the Deployment replicas increased"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 2
diff --git a/internal/controller/autoscaling/testdata/controller/metric-scaling/should-wake-the-fleet-from-zero-when-the-metric-turns-positive.yaml b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-wake-the-fleet-from-zero-when-the-metric-turns-positive.yaml
new file mode 100644
index 0000000..63bc655
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/metric-scaling/should-wake-the-fleet-from-zero-when-the-metric-turns-positive.yaml
@@ -0,0 +1,22 @@
+# desired = ceil(1 * 0.85/0.5) = 2 against the single-replica baseline, not ceil(0 * 0.85/0.5) = 0.
+"name": "should-wake-the-fleet-from-zero-when-the-metric-turns-positive"
+"description": "should wake the fleet from zero when the metric turns positive"
+"inputs":
+  "plainReconciler": true
+  "prelude": "lowering the scaler's minReplicas to 0 and parking the Deployment at 0"
+  "spec":
+    "minReplicas": 0
+  "deploy":
+    "replicas": 0
+    "updated": 0
+    "ready": 0
+    "available": 0
+    "spec": 0
+  "steps":
+  - "by": "running the Reconciler"
+  - "by": "checking the fleet woke up"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 2
diff --git a/internal/controller/autoscaling/testdata/controller/rollout-guard/should-defer-scaling-while-the-deployment-is-rolling-out.yaml b/internal/controller/autoscaling/testdata/controller/rollout-guard/should-defer-scaling-while-the-deployment-is-rolling-out.yaml
new file mode 100644
index 0000000..b6e13d2
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/rollout-guard/should-defer-scaling-while-the-deployment-is-rolling-out.yaml
@@ -0,0 +1,17 @@
+# The metric would compute desired=2, but the rollout guard defers scaling, so replicas stay at 1.
+"name": "should-defer-scaling-while-the-deployment-is-rolling-out"
+"description": "should defer scaling while the Deployment is rolling out"
+"inputs":
+  "plainReconciler": true
+  "steps":
+  - "by": "marking the Deployment as mid-rollout"
+    "rollout": true
+    "skipReconcile": true
+  - "by": "running the Reconciler"
+  - "by": "checking the Deployment was NOT scaled"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - {}
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/controller/settle-guard/should-also-wait-for-the-drain-when-checkterminatingreplicas-is-on.yaml b/internal/controller/autoscaling/testdata/controller/settle-guard/should-also-wait-for-the-drain-when-checkterminatingreplicas-is-on.yaml
new file mode 100644
index 0000000..a8d6dc7
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/settle-guard/should-also-wait-for-the-drain-when-checkterminatingreplicas-is-on.yaml
@@ -0,0 +1,32 @@
+"name": "should-also-wait-for-the-drain-when-checkterminatingreplicas-is-on"
+"description": "should also wait for the drain when checkTerminatingReplicas is on"
+"inputs":
+  "plainReconciler": true
+  "checkTerminating": true
+  "metric": "0"
+  "steps":
+  - "by": "setting 5 ready with 2 still draining"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "spec": 5
+      "terminating": 2
+  - "by": "checking it is deferred; readyReplicas alone would have said settled"
+    "skipReconcile": true
+  - "by": "letting the drain finish"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "terminating": 0
+  - "by": "checking the scale-down now proceeds"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 5
+  - {}
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/controller/settle-guard/should-defer-a-scale-down-until-readyreplicas-is-back-at-specreplicas.yaml b/internal/controller/autoscaling/testdata/controller/settle-guard/should-defer-a-scale-down-until-readyreplicas-is-back-at-specreplicas.yaml
new file mode 100644
index 0000000..c4c06d9
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/settle-guard/should-defer-a-scale-down-until-readyreplicas-is-back-at-specreplicas.yaml
@@ -0,0 +1,29 @@
+"name": "should-defer-a-scale-down-until-readyreplicas-is-back-at-specreplicas"
+"description": "should defer a scale-down until readyReplicas is back at specReplicas"
+"inputs":
+  "plainReconciler": true
+  "metric": "0"
+  "steps":
+  - "by": "setting the Deployment to 5 replicas with only 3 ready"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 3
+      "available": 3
+      "spec": 5
+  - "by": "checking the scale-down was deferred"
+    "skipReconcile": true
+  - "by": "letting the fleet converge"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+  - "by": "checking the scale-down now proceeds"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 5
+  - {}
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/controller/settle-guard/should-ignore-still-draining-pods-while-checkterminatingreplicas-is-off.yaml b/internal/controller/autoscaling/testdata/controller/settle-guard/should-ignore-still-draining-pods-while-checkterminatingreplicas-is-off.yaml
new file mode 100644
index 0000000..23c2572
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/settle-guard/should-ignore-still-draining-pods-while-checkterminatingreplicas-is-off.yaml
@@ -0,0 +1,22 @@
+# readyReplicas excludes terminating pods, so the fleet reads as settled the moment the previous step's pods are marked for deletion.
+"name": "should-ignore-still-draining-pods-while-checkterminatingreplicas-is-off"
+"description": "should ignore still-draining pods while checkTerminatingReplicas is off"
+"inputs":
+  "plainReconciler": true
+  "checkTerminating": false
+  "metric": "0"
+  "steps":
+  - "by": "setting 5 ready with 2 still draining"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "spec": 5
+      "terminating": 2
+  - "by": "checking the scale-down was not held back by the drain"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/controller/settle-guard/should-not-gate-scale-up-on-the-previous-scale-down-settling.yaml b/internal/controller/autoscaling/testdata/controller/settle-guard/should-not-gate-scale-up-on-the-previous-scale-down-settling.yaml
new file mode 100644
index 0000000..8b68273
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/settle-guard/should-not-gate-scale-up-on-the-previous-scale-down-settling.yaml
@@ -0,0 +1,21 @@
+# ceil(2 * 0.85/0.5) = 4; the settle guard only gates scale-down.
+"name": "should-not-gate-scale-up-on-the-previous-scale-down-settling"
+"description": "should not gate scale-up on the previous scale-down settling"
+"inputs":
+  "plainReconciler": true
+  "steps":
+  - "by": "setting 3 replicas with 2 ready and 1 still terminating, under load"
+    "metric": "0.85"
+    "deploy":
+      "replicas": 3
+      "updated": 3
+      "ready": 2
+      "available": 2
+      "spec": 3
+      "terminating": 1
+  - "by": "checking capacity was added despite the unsettled fleet"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 4
diff --git a/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-hold-for-the-stabilization-window-between-scale-down-steps.yaml b/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-hold-for-the-stabilization-window-between-scale-down-steps.yaml
new file mode 100644
index 0000000..0c78180
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-hold-for-the-stabilization-window-between-scale-down-steps.yaml
@@ -0,0 +1,29 @@
+"name": "should-hold-for-the-stabilization-window-between-scale-down-steps"
+"description": "should hold for the stabilization window between scale-down steps"
+"inputs":
+  "plainReconciler": true
+  "prelude": "enabling a 300s window and a 2-replica step cap"
+  "spec":
+    "stabilizationWindowSeconds": 300
+    "maxStepReplicas": 2
+  "metric": "0"
+  "steps":
+  - "by": "starting settled at 5 with an idle metric"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "spec": 5
+  - "by": "checking the first step landed"
+    "skipReconcile": true
+  - "by": "settling the fleet at 3 so neither the rollout nor the settle guard fires"
+    "observeSpec": 3
+  - "by": "checking the next step is held; without the pin it would drop to 1"
+    "skipReconcile": true
+"expected":
+  "steps":
+  - {}
+  - "replicas": 3
+  - {}
+  - "replicas": 3
diff --git a/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-walk-down-one-step-at-a-time-when-maxstepreplicas-is-set.yaml b/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-walk-down-one-step-at-a-time-when-maxstepreplicas-is-set.yaml
new file mode 100644
index 0000000..3bfcab2
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/controller/stabilization-and-step/should-walk-down-one-step-at-a-time-when-maxstepreplicas-is-set.yaml
@@ -0,0 +1,33 @@
+"name": "should-walk-down-one-step-at-a-time-when-maxstepreplicas-is-set"
+"description": "should walk down one step at a time when maxStepReplicas is set"
+"inputs":
+  "plainReconciler": true
+  "prelude": "capping the scale-down step at 2 replicas"
+  "spec":
+    "maxStepReplicas": 2
+  "steps":
+  - "by": "setting the Deployment to 5 replicas and the metric to 0 (idle)"
+    "skipReconcile": true
+    "metric": "0"
+    "deploy":
+      "replicas": 5
+      "updated": 5
+      "ready": 5
+      "available": 5
+      "spec": 5
+  - "by": "running the Reconciler; raw recommendation is minReplicas but the step is capped"
+  - "by": "checking status reports the uncapped recommendation, not the step"
+    "skipReconcile": true
+  - "by": "running the Reconciler again; it takes the next step down"
+    "deploy":
+      "replicas": 3
+      "updated": 3
+      "ready": 3
+      "available": 3
+      "spec": 3
+"expected":
+  "steps":
+  - {}
+  - "replicas": 3
+  - "desired": 1
+  - "replicas": 1
diff --git a/internal/controller/autoscaling/testdata/guards/hold/rollout-fully-rolled-out-does-not-hold.yaml b/internal/controller/autoscaling/testdata/guards/hold/rollout-fully-rolled-out-does-not-hold.yaml
new file mode 100644
index 0000000..e45635c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/rollout-fully-rolled-out-does-not-hold.yaml
@@ -0,0 +1,18 @@
+"name": "rollout-fully-rolled-out-does-not-hold"
+"description": "rollout: fully rolled out does not hold"
+"inputs":
+  "guard": "rollout"
+  "state":
+    "spec": 5
+    "total": 5
+    "updated": 5
+    "ready": 5
+    "generation": 3
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "held": false
+  "reason": "RolloutInProgress"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/rollout-generation-not-observed-holds-a-scale-up.yaml b/internal/controller/autoscaling/testdata/guards/hold/rollout-generation-not-observed-holds-a-scale-up.yaml
new file mode 100644
index 0000000..a6ef99e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/rollout-generation-not-observed-holds-a-scale-up.yaml
@@ -0,0 +1,19 @@
+"name": "rollout-generation-not-observed-holds-a-scale-up"
+"description": "rollout: generation not observed holds a scale-up"
+"inputs":
+  "guard": "rollout"
+  "state":
+    "spec": 5
+    "total": 5
+    "updated": 5
+    "ready": 5
+    "generation": 4
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 7
+"expected":
+  "held": true
+  "reason": "RolloutInProgress"
+  "why": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/rollout-non-deployment-is-never-held.yaml b/internal/controller/autoscaling/testdata/guards/hold/rollout-non-deployment-is-never-held.yaml
new file mode 100644
index 0000000..2e899fc
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/rollout-non-deployment-is-never-held.yaml
@@ -0,0 +1,17 @@
+"name": "rollout-non-deployment-is-never-held"
+"description": "rollout: non-Deployment is never held"
+"inputs":
+  "guard": "rollout"
+  "state":
+    "spec": 5
+    "total": 6
+    "updated": 0
+    "ready": 5
+    "generation": 4
+    "observedGeneration": 1
+  "intent":
+    "from": 5
+    "to": 7
+"expected":
+  "held": false
+  "reason": "RolloutInProgress"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/rollout-old-template-pods-hold-a-no-op.yaml b/internal/controller/autoscaling/testdata/guards/hold/rollout-old-template-pods-hold-a-no-op.yaml
new file mode 100644
index 0000000..e1517f2
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/rollout-old-template-pods-hold-a-no-op.yaml
@@ -0,0 +1,19 @@
+"name": "rollout-old-template-pods-hold-a-no-op"
+"description": "rollout: old-template pods hold a no-op"
+"inputs":
+  "guard": "rollout"
+  "state":
+    "spec": 5
+    "total": 6
+    "updated": 5
+    "ready": 5
+    "generation": 3
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 5
+"expected":
+  "held": true
+  "reason": "RolloutInProgress"
+  "why": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/rollout-updated-replicas-behind-spec-holds-a-scale-down.yaml b/internal/controller/autoscaling/testdata/guards/hold/rollout-updated-replicas-behind-spec-holds-a-scale-down.yaml
new file mode 100644
index 0000000..21f415c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/rollout-updated-replicas-behind-spec-holds-a-scale-down.yaml
@@ -0,0 +1,19 @@
+"name": "rollout-updated-replicas-behind-spec-holds-a-scale-down"
+"description": "rollout: updated replicas behind spec holds a scale-down"
+"inputs":
+  "guard": "rollout"
+  "state":
+    "spec": 5
+    "total": 4
+    "updated": 4
+    "ready": 4
+    "generation": 3
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 3
+"expected":
+  "held": true
+  "reason": "RolloutInProgress"
+  "why": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-no-op-is-not-held.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-no-op-is-not-held.yaml
new file mode 100644
index 0000000..17c8008
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-no-op-is-not-held.yaml
@@ -0,0 +1,13 @@
+"name": "settle-no-op-is-not-held"
+"description": "settle: no-op is not held"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 2
+  "intent":
+    "from": 5
+    "to": 5
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-ready-back-at-spec.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-back-at-spec.yaml
new file mode 100644
index 0000000..b103b60
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-back-at-spec.yaml
@@ -0,0 +1,13 @@
+"name": "settle-ready-back-at-spec"
+"description": "settle: ready back at spec"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 5
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-ready-count-of-0-is-not-settled.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-count-of-0-is-not-settled.yaml
new file mode 100644
index 0000000..3dab245
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-count-of-0-is-not-settled.yaml
@@ -0,0 +1,15 @@
+# A ready count of 0 is the raw missing-field value. Treating it as spec would skip the guard.
+"name": "settle-ready-count-of-0-is-not-settled"
+"description": "settle: ready count of 0 is not settled"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 0
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "held": true
+  "reason": "ScaleDownSettling"
+  "why": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-ready-differs-from-spec.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-differs-from-spec.yaml
new file mode 100644
index 0000000..4167edc
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-ready-differs-from-spec.yaml
@@ -0,0 +1,14 @@
+"name": "settle-ready-differs-from-spec"
+"description": "settle: ready differs from spec"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 3
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "held": true
+  "reason": "ScaleDownSettling"
+  "why": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-scale-up-is-not-held.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-scale-up-is-not-held.yaml
new file mode 100644
index 0000000..39d5508
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-scale-up-is-not-held.yaml
@@ -0,0 +1,13 @@
+"name": "settle-scale-up-is-not-held"
+"description": "settle: scale-up is not held"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 2
+  "intent":
+    "from": 5
+    "to": 7
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-count-of-0-with-the-gate-on.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-count-of-0-with-the-gate-on.yaml
new file mode 100644
index 0000000..903eac8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-count-of-0-with-the-gate-on.yaml
@@ -0,0 +1,16 @@
+"name": "settle-terminating-count-of-0-with-the-gate-on"
+"description": "settle: terminating count of 0 with the gate on"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 5
+    "terminating": 0
+    "terminatingReported": true
+  "intent":
+    "from": 5
+    "to": 4
+  "checkTerminating": true
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-not-reported-with-the-gate-on.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-not-reported-with-the-gate-on.yaml
new file mode 100644
index 0000000..ac3ef3f
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-not-reported-with-the-gate-on.yaml
@@ -0,0 +1,17 @@
+# The count is ignored when the field was not reported. A non-zero Terminating here would still be settled; the parser stores 0 in that case, so this row is what makes the flag observable.
+"name": "settle-terminating-not-reported-with-the-gate-on"
+"description": "settle: terminating not reported with the gate on"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 5
+    "terminating": 4
+    "terminatingReported": false
+  "intent":
+    "from": 5
+    "to": 4
+  "checkTerminating": true
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-off.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-off.yaml
new file mode 100644
index 0000000..3238c7e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-off.yaml
@@ -0,0 +1,15 @@
+"name": "settle-terminating-pods-with-the-gate-off"
+"description": "settle: terminating pods with the gate off"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 5
+    "terminating": 2
+    "terminatingReported": true
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "held": false
+  "reason": "ScaleDownSettling"
diff --git a/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-on.yaml b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-on.yaml
new file mode 100644
index 0000000..49b1bb8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/hold/settle-terminating-pods-with-the-gate-on.yaml
@@ -0,0 +1,17 @@
+"name": "settle-terminating-pods-with-the-gate-on"
+"description": "settle: terminating pods with the gate on"
+"inputs":
+  "guard": "settle"
+  "state":
+    "spec": 5
+    "ready": 5
+    "terminating": 2
+    "terminatingReported": true
+  "intent":
+    "from": 5
+    "to": 4
+  "checkTerminating": true
+"expected":
+  "held": true
+  "reason": "ScaleDownSettling"
+  "why": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/order/no-guard-holds-a-settled-scale-down.yaml b/internal/controller/autoscaling/testdata/guards/order/no-guard-holds-a-settled-scale-down.yaml
new file mode 100644
index 0000000..1410df5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/order/no-guard-holds-a-settled-scale-down.yaml
@@ -0,0 +1,15 @@
+"name": "no-guard-holds-a-settled-scale-down"
+"description": "no guard holds a settled scale-down"
+"inputs":
+  "state":
+    "spec": 5
+    "total": 5
+    "updated": 5
+    "ready": 5
+    "generation": 3
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 4
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/guards/order/rollout-wins-over-an-unsettled-scale-down.yaml b/internal/controller/autoscaling/testdata/guards/order/rollout-wins-over-an-unsettled-scale-down.yaml
new file mode 100644
index 0000000..17c79d9
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/order/rollout-wins-over-an-unsettled-scale-down.yaml
@@ -0,0 +1,17 @@
+"name": "rollout-wins-over-an-unsettled-scale-down"
+"description": "rollout wins over an unsettled scale-down"
+"inputs":
+  "state":
+    "spec": 5
+    "total": 5
+    "updated": 5
+    "ready": 2
+    "generation": 4
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 3
+"expected":
+  "reason": "RolloutInProgress"
+  "message": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/order/settle-holds-once-the-rollout-is-done.yaml b/internal/controller/autoscaling/testdata/guards/order/settle-holds-once-the-rollout-is-done.yaml
new file mode 100644
index 0000000..2f84b7e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/order/settle-holds-once-the-rollout-is-done.yaml
@@ -0,0 +1,17 @@
+"name": "settle-holds-once-the-rollout-is-done"
+"description": "settle holds once the rollout is done"
+"inputs":
+  "state":
+    "spec": 5
+    "total": 5
+    "updated": 5
+    "ready": 3
+    "generation": 3
+    "observedGeneration": 3
+    "tracksRollout": true
+  "intent":
+    "from": 5
+    "to": 4
+"expected":
+  "reason": "ScaleDownSettling"
+  "message": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/target/non-deployment-ignores-the-rollout-guard.yaml b/internal/controller/autoscaling/testdata/guards/target/non-deployment-ignores-the-rollout-guard.yaml
new file mode 100644
index 0000000..5d8db59
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/non-deployment-ignores-the-rollout-guard.yaml
@@ -0,0 +1,24 @@
+"name": "non-deployment-ignores-the-rollout-guard"
+"description": "non-Deployment ignores the rollout guard"
+"inputs":
+  "kind": "StatefulSet"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "StatefulSet"
+    "metadata":
+      "generation": 4
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 1
+      "updatedReplicas": 0
+      "replicas": 5
+      "readyReplicas": 5
+  "target": 7
+"expected":
+  "spec": 5
+  "ready": 5
+  "total": 5
+  "updated": 0
+  "generation": 4
+  "observedGeneration": 1
diff --git a/internal/controller/autoscaling/testdata/guards/target/ready-count-of-0-is-not-settled.yaml b/internal/controller/autoscaling/testdata/guards/target/ready-count-of-0-is-not-settled.yaml
new file mode 100644
index 0000000..0b397f6
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/ready-count-of-0-is-not-settled.yaml
@@ -0,0 +1,27 @@
+# readyReplicas 0 is also what a missing field reads as. Treating it as spec would skip the guard.
+"name": "ready-count-of-0-is-not-settled"
+"description": "ready count of 0 is not settled"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 3
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 0
+  "target": 4
+"expected":
+  "spec": 5
+  "ready": 0
+  "total": 5
+  "updated": 5
+  "generation": 3
+  "observedGeneration": 3
+  "reason": "ScaleDownSettling"
+  "message": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/target/rollout-holds-a-scale-up.yaml b/internal/controller/autoscaling/testdata/guards/target/rollout-holds-a-scale-up.yaml
new file mode 100644
index 0000000..a90dc43
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/rollout-holds-a-scale-up.yaml
@@ -0,0 +1,26 @@
+"name": "rollout-holds-a-scale-up"
+"description": "rollout holds a scale-up"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 4
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 5
+  "target": 7
+"expected":
+  "spec": 5
+  "ready": 5
+  "total": 5
+  "updated": 5
+  "generation": 4
+  "observedGeneration": 3
+  "reason": "RolloutInProgress"
+  "message": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/target/rollout-wins-over-an-unsettled-scale-down.yaml b/internal/controller/autoscaling/testdata/guards/target/rollout-wins-over-an-unsettled-scale-down.yaml
new file mode 100644
index 0000000..d5b13b5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/rollout-wins-over-an-unsettled-scale-down.yaml
@@ -0,0 +1,26 @@
+"name": "rollout-wins-over-an-unsettled-scale-down"
+"description": "rollout wins over an unsettled scale-down"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 4
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 2
+  "target": 3
+"expected":
+  "spec": 5
+  "ready": 2
+  "total": 5
+  "updated": 5
+  "generation": 4
+  "observedGeneration": 3
+  "reason": "RolloutInProgress"
+  "message": "Deployment rollout in progress"
diff --git a/internal/controller/autoscaling/testdata/guards/target/scale-up-is-not-gated-on-settle.yaml b/internal/controller/autoscaling/testdata/guards/target/scale-up-is-not-gated-on-settle.yaml
new file mode 100644
index 0000000..2e8152e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/scale-up-is-not-gated-on-settle.yaml
@@ -0,0 +1,24 @@
+"name": "scale-up-is-not-gated-on-settle"
+"description": "scale-up is not gated on settle"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 3
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 2
+  "target": 6
+"expected":
+  "spec": 5
+  "ready": 2
+  "total": 5
+  "updated": 5
+  "generation": 3
+  "observedGeneration": 3
diff --git a/internal/controller/autoscaling/testdata/guards/target/settled-scale-down-may-proceed.yaml b/internal/controller/autoscaling/testdata/guards/target/settled-scale-down-may-proceed.yaml
new file mode 100644
index 0000000..1ecb526
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/settled-scale-down-may-proceed.yaml
@@ -0,0 +1,24 @@
+"name": "settled-scale-down-may-proceed"
+"description": "settled scale-down may proceed"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 3
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 5
+  "target": 4
+"expected":
+  "spec": 5
+  "ready": 5
+  "total": 5
+  "updated": 5
+  "generation": 3
+  "observedGeneration": 3
diff --git a/internal/controller/autoscaling/testdata/guards/target/unsettled-scale-down.yaml b/internal/controller/autoscaling/testdata/guards/target/unsettled-scale-down.yaml
new file mode 100644
index 0000000..e4b4f25
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/target/unsettled-scale-down.yaml
@@ -0,0 +1,27 @@
+# Fully rolled out. readyReplicas is the raw count.
+"name": "unsettled-scale-down"
+"description": "unsettled scale-down"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 3
+    "spec":
+      "replicas": 5
+    "status":
+      "observedGeneration": 3
+      "updatedReplicas": 5
+      "replicas": 5
+      "readyReplicas": 3
+  "target": 4
+"expected":
+  "spec": 5
+  "ready": 3
+  "total": 5
+  "updated": 5
+  "generation": 3
+  "observedGeneration": 3
+  "reason": "ScaleDownSettling"
+  "message": "previous scale-down has not settled"
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/deployment-with-full-status.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/deployment-with-full-status.yaml
new file mode 100644
index 0000000..3d5cc10
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/deployment-with-full-status.yaml
@@ -0,0 +1,27 @@
+"name": "deployment-with-full-status"
+"description": "Deployment with full status"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/generic-kind-with-full-status.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/generic-kind-with-full-status.yaml
new file mode 100644
index 0000000..0e7e41a
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/generic-kind-with-full-status.yaml
@@ -0,0 +1,26 @@
+"name": "generic-kind-with-full-status"
+"description": "generic kind with full status"
+"inputs":
+  "kind": "ReplicaSet"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "ReplicaSet"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/invalid-typed-fields-behave-like-missing.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/invalid-typed-fields-behave-like-missing.yaml
new file mode 100644
index 0000000..821d145
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/invalid-typed-fields-behave-like-missing.yaml
@@ -0,0 +1,20 @@
+"name": "invalid-typed-fields-behave-like-missing"
+"description": "invalid-typed fields behave like missing"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": "nope"
+    "spec":
+      "replicas": "nope"
+    "status":
+      "readyReplicas": "nope"
+      "replicas": "nope"
+      "updatedReplicas": "nope"
+      "terminatingReplicas": "nope"
+      "observedGeneration": "nope"
+"expected":
+  "spec": 1
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/leaderworkerset-with-full-status.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/leaderworkerset-with-full-status.yaml
new file mode 100644
index 0000000..b728ca1
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/leaderworkerset-with-full-status.yaml
@@ -0,0 +1,26 @@
+"name": "leaderworkerset-with-full-status"
+"description": "LeaderWorkerSet with full status"
+"inputs":
+  "kind": "LeaderWorkerSet"
+  "object":
+    "apiVersion": "leaderworkerset.x-k8s.io/v1"
+    "kind": "LeaderWorkerSet"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/missing-generation-observedgeneration-updated-and-total-read-as-0.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/missing-generation-observedgeneration-updated-and-total-read-as-0.yaml
new file mode 100644
index 0000000..cc9c211
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/missing-generation-observedgeneration-updated-and-total-read-as-0.yaml
@@ -0,0 +1,18 @@
+"name": "missing-generation-observedgeneration-updated-and-total-read-as-0"
+"description": "missing generation, observedGeneration, updated, and total read as 0"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "terminatingReplicas": 1
+"expected":
+  "spec": 4
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/missing-readyreplicas-reads-as-0.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/missing-readyreplicas-reads-as-0.yaml
new file mode 100644
index 0000000..c40d6ea
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/missing-readyreplicas-reads-as-0.yaml
@@ -0,0 +1,26 @@
+"name": "missing-readyreplicas-reads-as-0"
+"description": "missing readyReplicas reads as 0"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 0
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/missing-spec-replicas-reads-as-1.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/missing-spec-replicas-reads-as-1.yaml
new file mode 100644
index 0000000..338ffca
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/missing-spec-replicas-reads-as-1.yaml
@@ -0,0 +1,25 @@
+"name": "missing-spec-replicas-reads-as-1"
+"description": "missing spec.replicas reads as 1"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 9
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 1
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/missing-terminatingreplicas-is-not-reported.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/missing-terminatingreplicas-is-not-reported.yaml
new file mode 100644
index 0000000..4782f37
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/missing-terminatingreplicas-is-not-reported.yaml
@@ -0,0 +1,24 @@
+"name": "missing-terminatingreplicas-is-not-reported"
+"description": "missing terminatingReplicas is not reported"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "generation": 9
+  "observedGeneration": 8
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/spec-replicas-of-0-is-kept.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/spec-replicas-of-0-is-kept.yaml
new file mode 100644
index 0000000..8477632
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/spec-replicas-of-0-is-kept.yaml
@@ -0,0 +1,21 @@
+"name": "spec-replicas-of-0-is-kept"
+"description": "spec.replicas of 0 is kept"
+"inputs":
+  "kind": "Deployment"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "Deployment"
+    "metadata":
+      "generation": 2
+    "spec":
+      "replicas": 0
+    "status":
+      "readyReplicas": 0
+      "replicas": 0
+      "updatedReplicas": 0
+      "observedGeneration": 2
+"expected":
+  "spec": 0
+  "generation": 2
+  "observedGeneration": 2
+  "tracksRollout": true
diff --git a/internal/controller/autoscaling/testdata/guards/workload-state/statefulset-with-full-status.yaml b/internal/controller/autoscaling/testdata/guards/workload-state/statefulset-with-full-status.yaml
new file mode 100644
index 0000000..2ba4e30
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/guards/workload-state/statefulset-with-full-status.yaml
@@ -0,0 +1,26 @@
+"name": "statefulset-with-full-status"
+"description": "StatefulSet with full status"
+"inputs":
+  "kind": "StatefulSet"
+  "object":
+    "apiVersion": "apps/v1"
+    "kind": "StatefulSet"
+    "metadata":
+      "generation": 9
+    "spec":
+      "replicas": 4
+    "status":
+      "readyReplicas": 3
+      "replicas": 6
+      "updatedReplicas": 2
+      "terminatingReplicas": 1
+      "observedGeneration": 8
+"expected":
+  "spec": 4
+  "total": 6
+  "updated": 2
+  "ready": 3
+  "terminating": 1
+  "terminatingReported": true
+  "generation": 9
+  "observedGeneration": 8
diff --git a/internal/controller/autoscaling/testdata/helpers/compute-desired/minus-inf.yaml b/internal/controller/autoscaling/testdata/helpers/compute-desired/minus-inf.yaml
new file mode 100644
index 0000000..8423bed
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/compute-desired/minus-inf.yaml
@@ -0,0 +1,9 @@
+"name": "minus-inf"
+"description": "-Inf"
+"inputs":
+  "sample": "-Inf"
+  "ready": 4
+  "metricName": "guardrail"
+  "query": "histogram_quantile(0.95, x)"
+  "target": "2"
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/compute-desired/nan.yaml b/internal/controller/autoscaling/testdata/helpers/compute-desired/nan.yaml
new file mode 100644
index 0000000..1d50b3f
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/compute-desired/nan.yaml
@@ -0,0 +1,9 @@
+"name": "nan"
+"description": "NaN"
+"inputs":
+  "sample": "NaN"
+  "ready": 4
+  "metricName": "guardrail"
+  "query": "histogram_quantile(0.95, x)"
+  "target": "2"
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/compute-desired/plus-inf.yaml b/internal/controller/autoscaling/testdata/helpers/compute-desired/plus-inf.yaml
new file mode 100644
index 0000000..5882327
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/compute-desired/plus-inf.yaml
@@ -0,0 +1,9 @@
+"name": "plus-inf"
+"description": "+Inf"
+"inputs":
+  "sample": "+Inf"
+  "ready": 4
+  "metricName": "guardrail"
+  "query": "histogram_quantile(0.95, x)"
+  "target": "2"
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/compute-desired/positive-signal-wakes-the-fleet.yaml b/internal/controller/autoscaling/testdata/helpers/compute-desired/positive-signal-wakes-the-fleet.yaml
new file mode 100644
index 0000000..acd1ecf
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/compute-desired/positive-signal-wakes-the-fleet.yaml
@@ -0,0 +1,14 @@
+# ceil(1 * 0.85/0.5) = 2, not ceil(0 * 0.85/0.5) = 0.
+"name": "positive-signal-wakes-the-fleet"
+"description": "positive signal wakes the fleet"
+"inputs":
+  "sample": "0.85"
+  "ready": 0
+  "minReplicas": 0
+  "maxReplicas": 5
+  "metricName": "queue"
+  "query": "sum(queue_depth)"
+  "target": "0.5"
+"expected":
+  "have": true
+  "desired": 2
diff --git a/internal/controller/autoscaling/testdata/helpers/compute-desired/zero-signal-recommends-zero.yaml b/internal/controller/autoscaling/testdata/helpers/compute-desired/zero-signal-recommends-zero.yaml
new file mode 100644
index 0000000..d77be3c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/compute-desired/zero-signal-recommends-zero.yaml
@@ -0,0 +1,14 @@
+# A zero signal still recommends 0.
+"name": "zero-signal-recommends-zero"
+"description": "zero signal recommends zero"
+"inputs":
+  "sample": "0"
+  "ready": 0
+  "minReplicas": 0
+  "maxReplicas": 5
+  "metricName": "queue"
+  "query": "sum(queue_depth)"
+  "target": "0.5"
+"expected":
+  "have": true
+  "desired": 0
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serveraddress.yaml b/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serveraddress.yaml
new file mode 100644
index 0000000..f700327
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serveraddress.yaml
@@ -0,0 +1,7 @@
+"name": "empty-serveraddress"
+"description": "empty serverAddress"
+"inputs":
+  "serverAddress": ""
+  "serviceId": "svc"
+"expected":
+  "errorContains": "serverAddress"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serviceid.yaml b/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serviceid.yaml
new file mode 100644
index 0000000..46befec
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-misconfigured/empty-serviceid.yaml
@@ -0,0 +1,7 @@
+"name": "empty-serviceid"
+"description": "empty serviceId"
+"inputs":
+  "serverAddress": "http://example.invalid"
+  "serviceId": ""
+"expected":
+  "errorContains": "serviceId"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-non-200/a-non-ok-reply-is-an-error-even-when-the-body-is-a-decision.yaml b/internal/controller/autoscaling/testdata/helpers/decision-non-200/a-non-ok-reply-is-an-error-even-when-the-body-is-a-decision.yaml
new file mode 100644
index 0000000..9f12d80
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-non-200/a-non-ok-reply-is-an-error-even-when-the-body-is-a-decision.yaml
@@ -0,0 +1,8 @@
+"name": "a-non-ok-reply-is-an-error-even-when-the-body-is-a-decision"
+"description": "a non-OK reply is an error even when the body is a decision"
+"inputs":
+  "serviceId": "svc"
+  "status": 503
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":3}}]}"
+"expected":
+  "errorContains": "status 503"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
new file mode 100644
index 0000000..c60fdbe
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
@@ -0,0 +1,7 @@
+"name": "a-body-that-is-not-json-is-an-error"
+"description": "a body that is not JSON is an error"
+"inputs":
+  "serviceId": "svc"
+  "body": "<html>502 Bad Gateway</html>"
+"expected":
+  "errorContains": "failed to decode"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
new file mode 100644
index 0000000..f570b8e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
@@ -0,0 +1,7 @@
+"name": "a-count-past-int32-is-refused-rather-than-wrapped"
+"description": "a count past int32 is refused rather than wrapped"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":4294967296}}]}"
+"expected":
+  "errorContains": "out of range"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
new file mode 100644
index 0000000..944223e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
@@ -0,0 +1,8 @@
+# A decision for someone else must not drive this scaler just because it came back.
+"name": "a-decision-for-another-service-does-not-match"
+"description": "a decision for another service does not match"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"other\",\"replicas\":{\"active\":9}}]}"
+"expected":
+  "errorContains": "no decision for serviceId \"svc\""
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
new file mode 100644
index 0000000..a03b318
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
@@ -0,0 +1,7 @@
+"name": "a-decision-with-no-active-count-is-not-a-zero"
+"description": "a decision with no active count is not a zero"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{}}]}"
+"expected":
+  "errorContains": "carries no replicas.active"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
new file mode 100644
index 0000000..a4bb415
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
@@ -0,0 +1,7 @@
+"name": "a-matching-decision-resolves-to-its-active-count"
+"description": "a matching decision resolves to its active count"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":15}}]}"
+"expected":
+  "replicas": 15
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
new file mode 100644
index 0000000..18664d7
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
@@ -0,0 +1,7 @@
+"name": "a-negative-count-is-refused"
+"description": "a negative count is refused"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":-1}}]}"
+"expected":
+  "errorContains": "out of range"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
new file mode 100644
index 0000000..ec878c0
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
@@ -0,0 +1,7 @@
+"name": "a-payload-with-no-apiversion-is-refused"
+"description": "a payload with no apiVersion is refused"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":3}}]}"
+"expected":
+  "errorContains": "unsupported decision schema"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
new file mode 100644
index 0000000..de280d4
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
@@ -0,0 +1,7 @@
+"name": "an-empty-decision-list-is-not-a-zero"
+"description": "an empty decision list is not a zero"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[]}"
+"expected":
+  "errorContains": "no decision for serviceId \"svc\""
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
new file mode 100644
index 0000000..49c66da
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
@@ -0,0 +1,7 @@
+"name": "an-unknown-schema-is-refused-not-parsed-hopefully"
+"description": "an unknown schema is refused, not parsed hopefully"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v2\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":3}}]}"
+"expected":
+  "errorContains": "unsupported decision schema"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
new file mode 100644
index 0000000..bbacfff
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
@@ -0,0 +1,7 @@
+"name": "duplicates-are-an-error-not-a-pick-first"
+"description": "duplicates are an error, not a pick-first"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}},{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":7}}]}"
+"expected":
+  "errorContains": "2 decisions match"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
new file mode 100644
index 0000000..51f4233
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
@@ -0,0 +1,8 @@
+"name": "the-namespace-filter-is-reported-when-nothing-matches"
+"description": "the namespace filter is reported when nothing matches"
+"inputs":
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}}]}"
+"expected":
+  "errorContains": "in namespace \"modelforge\""
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
new file mode 100644
index 0000000..8837b9c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
@@ -0,0 +1,8 @@
+"name": "the-namespace-filter-picks-between-same-named-services"
+"description": "the namespace filter picks between same-named services"
+"inputs":
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}},{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":7}}]}"
+"expected":
+  "replicas": 7
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
new file mode 100644
index 0000000..60ac9ac
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
@@ -0,0 +1,7 @@
+"name": "zero-is-a-decision-not-a-missing-one"
+"description": "zero is a decision, not a missing one"
+"inputs":
+  "serviceId": "svc"
+  "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":0}}]}"
+"expected":
+  "replicas": 0
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-request/a-configured-path-is-used.yaml b/internal/controller/autoscaling/testdata/helpers/decision-request/a-configured-path-is-used.yaml
new file mode 100644
index 0000000..a1e8db1
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-request/a-configured-path-is-used.yaml
@@ -0,0 +1,12 @@
+"name": "a-configured-path-is-used"
+"description": "a configured path is used"
+"inputs":
+  "serviceId": "fallback-modelforge-01"
+  "path": "/v2/decisions"
+  "headers":
+    "Authorization": "Bearer t"
+"expected":
+  "replicas": 4
+  "path": "/v2/decisions"
+  "serviceId": "fallback-modelforge-01"
+  "authorization": "Bearer t"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-request/a-path-without-a-leading-slash-still-forms-one.yaml b/internal/controller/autoscaling/testdata/helpers/decision-request/a-path-without-a-leading-slash-still-forms-one.yaml
new file mode 100644
index 0000000..55e3bb5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-request/a-path-without-a-leading-slash-still-forms-one.yaml
@@ -0,0 +1,13 @@
+"name": "a-path-without-a-leading-slash-still-forms-one"
+"description": "a path without a leading slash still forms one"
+"inputs":
+  "serviceId": "fallback-modelforge-01"
+  "path": "decisions"
+  "headers":
+    "Authorization": "Bearer t"
+"expected":
+  "replicas": 4
+  "path": "/decisions"
+  "serviceId": "fallback-modelforge-01"
+  "authorization": "Bearer t"
+  "pinDefaultPath": true
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-request/a-trailing-slash-on-serveraddress-is-not-doubled.yaml b/internal/controller/autoscaling/testdata/helpers/decision-request/a-trailing-slash-on-serveraddress-is-not-doubled.yaml
new file mode 100644
index 0000000..ed99eb0
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-request/a-trailing-slash-on-serveraddress-is-not-doubled.yaml
@@ -0,0 +1,14 @@
+"name": "a-trailing-slash-on-serveraddress-is-not-doubled"
+"description": "a trailing slash on serverAddress is not doubled"
+"inputs":
+  "serviceId": "fallback-modelforge-01"
+  "path": "/decisions"
+  "trailingSlash": true
+  "headers":
+    "Authorization": "Bearer t"
+"expected":
+  "replicas": 4
+  "path": "/decisions"
+  "serviceId": "fallback-modelforge-01"
+  "authorization": "Bearer t"
+  "pinDefaultPath": true
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-request/an-empty-path-falls-back-to-the-default.yaml b/internal/controller/autoscaling/testdata/helpers/decision-request/an-empty-path-falls-back-to-the-default.yaml
new file mode 100644
index 0000000..560f815
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-request/an-empty-path-falls-back-to-the-default.yaml
@@ -0,0 +1,12 @@
+"name": "an-empty-path-falls-back-to-the-default"
+"description": "an empty path falls back to the default"
+"inputs":
+  "serviceId": "fallback-modelforge-01"
+  "headers":
+    "Authorization": "Bearer t"
+"expected":
+  "replicas": 4
+  "path": "/decisions"
+  "serviceId": "fallback-modelforge-01"
+  "authorization": "Bearer t"
+  "pinDefaultPath": true
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml b/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
new file mode 100644
index 0000000..9ec7fe8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
@@ -0,0 +1,21 @@
+# Wire format from the decision server, not a fixture built by the table.
+"name": "a-verbatim-decision-server-payload-resolves-to-15"
+"description": "a verbatim decision-server payload resolves to 15"
+"inputs":
+  "serviceId": "fallback-modelforge-01"
+  "body": |-
+    {
+      "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1",
+      "decisions": [
+        {
+          "namespace": "modelforge",
+          "serviceId": "fallback-modelforge-01",
+          "replicas": {
+            "active": 15
+          }
+        }
+      ]
+    }
+  "pinAPIVersion": true
+"expected":
+  "replicas": 15
diff --git a/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-0.yaml b/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-0.yaml
new file mode 100644
index 0000000..2499429
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-0.yaml
@@ -0,0 +1,6 @@
+"name": "explicit-0"
+"description": "explicit 0"
+"inputs":
+  "deferTimeoutSeconds": 0
+"expected":
+  "duration": "0s"
diff --git a/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-30.yaml b/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-30.yaml
new file mode 100644
index 0000000..bed9777
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/defer-timeout/explicit-30.yaml
@@ -0,0 +1,6 @@
+"name": "explicit-30"
+"description": "explicit 30"
+"inputs":
+  "deferTimeoutSeconds": 30
+"expected":
+  "duration": "30s"
diff --git a/internal/controller/autoscaling/testdata/helpers/defer-timeout/nil-spec.yaml b/internal/controller/autoscaling/testdata/helpers/defer-timeout/nil-spec.yaml
new file mode 100644
index 0000000..a5ed99c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/defer-timeout/nil-spec.yaml
@@ -0,0 +1,6 @@
+"name": "nil-spec"
+"description": "nil spec"
+"inputs": {}
+"expected":
+  "duration": "600s"
+  "equalsDefault": true
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-memory/enddeferral-and-forgetrecommendations-clear-the-forced-write-record.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-memory/enddeferral-and-forgetrecommendations-clear-the-forced-write-record.yaml
new file mode 100644
index 0000000..739f16f
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-memory/enddeferral-and-forgetrecommendations-clear-the-forced-write-record.yaml
@@ -0,0 +1,16 @@
+"name": "enddeferral-and-forgetrecommendations-clear-the-forced-write-record"
+"description": "endDeferral and forgetRecommendations clear the forced-write record"
+"inputs":
+  "namespace": "ns"
+  "scaler": "s"
+  "at": "2026-03-01T00:00:00Z"
+  "steps":
+  - "op": "record"
+  - "op": "setSince"
+  - "op": "end"
+  - "op": "expectSinceCleared"
+  - "op": "expectRecordCleared"
+  - "op": "record"
+  - "op": "forget"
+  - "op": "expectRecordCleared"
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-already-passed.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-already-passed.yaml
new file mode 100644
index 0000000..85d0778
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-already-passed.yaml
@@ -0,0 +1,8 @@
+"name": "deadline-already-passed"
+"description": "deadline already passed"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "-1s"
+"expected":
+  "after": "1s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-due-this-instant.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-due-this-instant.yaml
new file mode 100644
index 0000000..25e3bda
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-due-this-instant.yaml
@@ -0,0 +1,8 @@
+"name": "deadline-due-this-instant"
+"description": "deadline due this instant"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "0s"
+"expected":
+  "after": "1s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-later-than-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-later-than-sync-period.yaml
new file mode 100644
index 0000000..1a90ff3
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-later-than-sync-period.yaml
@@ -0,0 +1,8 @@
+"name": "deadline-later-than-sync-period"
+"description": "deadline later than sync period"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "30s"
+"expected":
+  "after": "15s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-sooner-than-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-sooner-than-sync-period.yaml
new file mode 100644
index 0000000..6dda718
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-sooner-than-sync-period.yaml
@@ -0,0 +1,8 @@
+"name": "deadline-sooner-than-sync-period"
+"description": "deadline sooner than sync period"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "5s"
+"expected":
+  "after": "5s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-under-the-1s-floor.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-under-the-1s-floor.yaml
new file mode 100644
index 0000000..e76800b
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/deadline-under-the-1s-floor.yaml
@@ -0,0 +1,8 @@
+"name": "deadline-under-the-1s-floor"
+"description": "deadline under the 1s floor"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "200ms"
+"expected":
+  "after": "1s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/not-holding-stays-on-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/not-holding-stays-on-sync-period.yaml
new file mode 100644
index 0000000..1f065b6
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/not-holding-stays-on-sync-period.yaml
@@ -0,0 +1,8 @@
+"name": "not-holding-stays-on-sync-period"
+"description": "not holding stays on sync period"
+"inputs":
+  "holding": false
+  "timeout": "1m"
+  "until": "5s"
+"expected":
+  "after": "15s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-later-than-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-later-than-sync-period.yaml
new file mode 100644
index 0000000..f8d1353
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-later-than-sync-period.yaml
@@ -0,0 +1,9 @@
+"name": "pacing-later-than-sync-period"
+"description": "pacing later than sync period"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "-1s"
+  "pacing": "30s"
+"expected":
+  "after": "15s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-sooner-than-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-sooner-than-sync-period.yaml
new file mode 100644
index 0000000..e43f746
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-sooner-than-sync-period.yaml
@@ -0,0 +1,9 @@
+"name": "pacing-sooner-than-sync-period"
+"description": "pacing sooner than sync period"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "-1s"
+  "pacing": "4s"
+"expected":
+  "after": "4s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-under-the-1s-floor.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-under-the-1s-floor.yaml
new file mode 100644
index 0000000..a777abc
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/pacing-under-the-1s-floor.yaml
@@ -0,0 +1,9 @@
+"name": "pacing-under-the-1s-floor"
+"description": "pacing under the 1s floor"
+"inputs":
+  "holding": true
+  "timeout": "1m"
+  "until": "-1s"
+  "pacing": "100ms"
+"expected":
+  "after": "1s"
diff --git a/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/timeout-0-stays-on-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/timeout-0-stays-on-sync-period.yaml
new file mode 100644
index 0000000..3c171bb
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/deferral-requeue-after/timeout-0-stays-on-sync-period.yaml
@@ -0,0 +1,8 @@
+"name": "timeout-0-stays-on-sync-period"
+"description": "timeout 0 stays on sync period"
+"inputs":
+  "holding": true
+  "timeout": "0s"
+  "until": "5s"
+"expected":
+  "after": "15s"
diff --git a/internal/controller/autoscaling/testdata/helpers/forced-write-pace/14s-into-a-15s-period.yaml b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/14s-into-a-15s-period.yaml
new file mode 100644
index 0000000..6825126
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/14s-into-a-15s-period.yaml
@@ -0,0 +1,9 @@
+"name": "14s-into-a-15s-period"
+"description": "14s into a 15s period"
+"inputs":
+  "recorded": true
+  "sync": "15s"
+  "at": "14s"
+"expected":
+  "allowed": false
+  "remaining": "1s"
diff --git a/internal/controller/autoscaling/testdata/helpers/forced-write-pace/exactly-one-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/exactly-one-sync-period.yaml
new file mode 100644
index 0000000..69cbb8b
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/exactly-one-sync-period.yaml
@@ -0,0 +1,8 @@
+"name": "exactly-one-sync-period"
+"description": "exactly one sync period"
+"inputs":
+  "recorded": true
+  "sync": "15s"
+  "at": "15s"
+"expected":
+  "allowed": true
diff --git a/internal/controller/autoscaling/testdata/helpers/forced-write-pace/just-recorded.yaml b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/just-recorded.yaml
new file mode 100644
index 0000000..57ed154
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/just-recorded.yaml
@@ -0,0 +1,8 @@
+"name": "just-recorded"
+"description": "just recorded"
+"inputs":
+  "recorded": true
+  "sync": "15s"
+  "at": "0s"
+"expected":
+  "allowed": false
diff --git a/internal/controller/autoscaling/testdata/helpers/forced-write-pace/no-record.yaml b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/no-record.yaml
new file mode 100644
index 0000000..e4af893
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/no-record.yaml
@@ -0,0 +1,9 @@
+"name": "no-record"
+"description": "no record"
+"inputs":
+  "recorded": false
+  "sync": "15s"
+  "at": "0s"
+"expected":
+  "allowed": true
+  "remaining": "0s"
diff --git a/internal/controller/autoscaling/testdata/helpers/forced-write-pace/zero-sync-period.yaml b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/zero-sync-period.yaml
new file mode 100644
index 0000000..4bfe8d3
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/forced-write-pace/zero-sync-period.yaml
@@ -0,0 +1,9 @@
+"name": "zero-sync-period"
+"description": "zero sync period"
+"inputs":
+  "recorded": true
+  "sync": "0s"
+  "at": "1h"
+"expected":
+  "allowed": true
+  "remaining": "0s"
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/empty.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/empty.yaml
new file mode 100644
index 0000000..ac8035d
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/empty.yaml
@@ -0,0 +1,6 @@
+"name": "empty"
+"description": "empty"
+"inputs":
+  "value": ""
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/inf-parses-as-a-float-but-is-not-usable.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/inf-parses-as-a-float-but-is-not-usable.yaml
new file mode 100644
index 0000000..c14be6e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/inf-parses-as-a-float-but-is-not-usable.yaml
@@ -0,0 +1,6 @@
+"name": "inf-parses-as-a-float-but-is-not-usable"
+"description": "Inf parses as a float but is not usable"
+"inputs":
+  "value": "+Inf"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/nan-parses-as-a-float-but-is-not-usable.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/nan-parses-as-a-float-but-is-not-usable.yaml
new file mode 100644
index 0000000..664966d
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/nan-parses-as-a-float-but-is-not-usable.yaml
@@ -0,0 +1,6 @@
+"name": "nan-parses-as-a-float-but-is-not-usable"
+"description": "NaN parses as a float but is not usable"
+"inputs":
+  "value": "NaN"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/negative-would-invert-scaling.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/negative-would-invert-scaling.yaml
new file mode 100644
index 0000000..e2084dc
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/negative-would-invert-scaling.yaml
@@ -0,0 +1,6 @@
+"name": "negative-would-invert-scaling"
+"description": "negative would invert scaling"
+"inputs":
+  "value": "-1"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/not-a-number.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/not-a-number.yaml
new file mode 100644
index 0000000..83b3203
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/not-a-number.yaml
@@ -0,0 +1,6 @@
+"name": "not-a-number"
+"description": "not a number"
+"inputs":
+  "value": "abc"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/percent-suffix-is-rejected-not-stripped.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/percent-suffix-is-rejected-not-stripped.yaml
new file mode 100644
index 0000000..fc811f3
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/percent-suffix-is-rejected-not-stripped.yaml
@@ -0,0 +1,6 @@
+"name": "percent-suffix-is-rejected-not-stripped"
+"description": "percent suffix is rejected, not stripped"
+"inputs":
+  "value": "80%"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/plain-integer.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/plain-integer.yaml
new file mode 100644
index 0000000..d4a875f
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/plain-integer.yaml
@@ -0,0 +1,6 @@
+"name": "plain-integer"
+"description": "plain integer"
+"inputs":
+  "value": "5"
+"expected":
+  "value": 5
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/ratio.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/ratio.yaml
new file mode 100644
index 0000000..9999cf6
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/ratio.yaml
@@ -0,0 +1,6 @@
+"name": "ratio"
+"description": "ratio"
+"inputs":
+  "value": "0.8"
+"expected":
+  "value": 0.8
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/surrounding-whitespace.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/surrounding-whitespace.yaml
new file mode 100644
index 0000000..168c2bb
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/surrounding-whitespace.yaml
@@ -0,0 +1,6 @@
+"name": "surrounding-whitespace"
+"description": "surrounding whitespace"
+"inputs":
+  "value": "  2.5 "
+"expected":
+  "value": 2.5
diff --git a/internal/controller/autoscaling/testdata/helpers/parse-target-value/zero-would-make-the-ratio-infinite.yaml b/internal/controller/autoscaling/testdata/helpers/parse-target-value/zero-would-make-the-ratio-infinite.yaml
new file mode 100644
index 0000000..5291ffe
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/parse-target-value/zero-would-make-the-ratio-infinite.yaml
@@ -0,0 +1,6 @@
+"name": "zero-would-make-the-ratio-infinite"
+"description": "zero would make the ratio infinite"
+"inputs":
+  "value": "0"
+"expected":
+  "error": true
diff --git a/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/59s-of-a-60s-timeout.yaml b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/59s-of-a-60s-timeout.yaml
new file mode 100644
index 0000000..cf07d0c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/59s-of-a-60s-timeout.yaml
@@ -0,0 +1,7 @@
+"name": "59s-of-a-60s-timeout"
+"description": "59s of a 60s timeout"
+"inputs":
+  "timeout": "1m"
+  "elapsed": "59s"
+"expected":
+  "past": false
diff --git a/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/exactly-the-timeout.yaml b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/exactly-the-timeout.yaml
new file mode 100644
index 0000000..0cac796
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/exactly-the-timeout.yaml
@@ -0,0 +1,7 @@
+"name": "exactly-the-timeout"
+"description": "exactly the timeout"
+"inputs":
+  "timeout": "1m"
+  "elapsed": "1m"
+"expected":
+  "past": true
diff --git a/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/nil-deferredsince.yaml b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/nil-deferredsince.yaml
new file mode 100644
index 0000000..4a53360
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/nil-deferredsince.yaml
@@ -0,0 +1,8 @@
+"name": "nil-deferredsince"
+"description": "nil deferredSince"
+"inputs":
+  "nilSince": true
+  "timeout": "1m"
+  "elapsed": "1h"
+"expected":
+  "past": false
diff --git a/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/past-the-timeout.yaml b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/past-the-timeout.yaml
new file mode 100644
index 0000000..57f71e9
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/past-the-timeout.yaml
@@ -0,0 +1,7 @@
+"name": "past-the-timeout"
+"description": "past the timeout"
+"inputs":
+  "timeout": "1m"
+  "elapsed": "1m1s"
+"expected":
+  "past": true
diff --git a/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/timeout-0.yaml b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/timeout-0.yaml
new file mode 100644
index 0000000..b5c583c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/past-defer-timeout/timeout-0.yaml
@@ -0,0 +1,7 @@
+"name": "timeout-0"
+"description": "timeout 0"
+"inputs":
+  "timeout": "0s"
+  "elapsed": "24h"
+"expected":
+  "past": false
diff --git a/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/an-under-aggregated-query-is-rejected-naming-the-series.yaml b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/an-under-aggregated-query-is-rejected-naming-the-series.yaml
new file mode 100644
index 0000000..469d0c9
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/an-under-aggregated-query-is-rejected-naming-the-series.yaml
@@ -0,0 +1,9 @@
+"name": "an-under-aggregated-query-is-rejected-naming-the-series"
+"description": "an under-aggregated query is rejected, naming the series"
+"inputs":
+  "result": |-
+    [{"metric":{"__name__":"m","model":"qwen","backend":"10.0.0.1:8000"},"value":[0,"1"]},
+                      {"metric":{"__name__":"m","model":"qwen","backend":"10.0.0.2:8000"},"value":[0,"9"]}]
+"expected":
+  "errorContains":
+  - "returned 2 series"
diff --git a/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/empty-vector-is-not-zero.yaml b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/empty-vector-is-not-zero.yaml
new file mode 100644
index 0000000..a1adcf8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/empty-vector-is-not-zero.yaml
@@ -0,0 +1,7 @@
+"name": "empty-vector-is-not-zero"
+"description": "empty vector is not zero"
+"inputs":
+  "result": "[]"
+"expected":
+  "errorContains":
+  - "returned no data"
diff --git a/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/multi-series-error-names-both-label-sets.yaml b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/multi-series-error-names-both-label-sets.yaml
new file mode 100644
index 0000000..5a27b8c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/multi-series-error-names-both-label-sets.yaml
@@ -0,0 +1,10 @@
+# The multi-series error quotes both series with their labels sorted.
+"name": "multi-series-error-names-both-label-sets"
+"description": "multi-series error names both label sets"
+"inputs":
+  "result": "[\n\t\t\t{\"metric\":{\"model\":\"qwen\",\"backend\":\"10.0.0.1:8000\"},\"value\":[0,\"1\"]},\n\t\t\t{\"metric\":{\"model\":\"qwen\",\"backend\":\"10.0.0.2:8000\"},\"value\":[0,\"9\"]}]"
+"expected":
+  "errorContains":
+  - "{backend=\"10.0.0.1:8000\",model=\"qwen\"}"
+  - "{backend=\"10.0.0.2:8000\",model=\"qwen\"}"
+  - "avg(...)"
diff --git a/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/single-series-resolves.yaml b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/single-series-resolves.yaml
new file mode 100644
index 0000000..a9c4bb8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/query-prometheus-scalar/single-series-resolves.yaml
@@ -0,0 +1,6 @@
+"name": "single-series-resolves"
+"description": "single series resolves"
+"inputs":
+  "result": "[{\"metric\":{},\"value\":[0,\"1.5\"]}]"
+"expected":
+  "value": 1.5
diff --git a/internal/controller/autoscaling/testdata/helpers/reason-deferral-timed-out/pins-the-published-condition-reason.yaml b/internal/controller/autoscaling/testdata/helpers/reason-deferral-timed-out/pins-the-published-condition-reason.yaml
new file mode 100644
index 0000000..0fd98c4
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/reason-deferral-timed-out/pins-the-published-condition-reason.yaml
@@ -0,0 +1,5 @@
+"name": "pins-the-published-condition-reason"
+"description": "pins the published condition reason"
+"inputs": {}
+"expected":
+  "reason": "DeferralTimedOut"
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-above-maxreplicas-is-clamped.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-above-maxreplicas-is-clamped.yaml
new file mode 100644
index 0000000..34f4b8a
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-above-maxreplicas-is-clamped.yaml
@@ -0,0 +1,14 @@
+"name": "a-decision-above-maxreplicas-is-clamped"
+"description": "a decision above maxReplicas is clamped"
+"inputs":
+  "active": "15"
+  "minReplicas": 1
+  "maxReplicas": 5
+  "readyReplicas": 8
+  "specReplicas": 8
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1"
+  "pinAPIVersion": true
+"expected":
+  "replicas": 5
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-of-zero-clamps-up-to-minreplicas.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-of-zero-clamps-up-to-minreplicas.yaml
new file mode 100644
index 0000000..13800e5
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-decision-of-zero-clamps-up-to-minreplicas.yaml
@@ -0,0 +1,14 @@
+"name": "a-decision-of-zero-clamps-up-to-minreplicas"
+"description": "a decision of zero clamps up to minReplicas"
+"inputs":
+  "active": "0"
+  "minReplicas": 2
+  "maxReplicas": 20
+  "readyReplicas": 8
+  "specReplicas": 8
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1"
+  "pinAPIVersion": true
+"expected":
+  "replicas": 2
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-missing-customprovider-holds-the-fleet-where-it-is.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-missing-customprovider-holds-the-fleet-where-it-is.yaml
new file mode 100644
index 0000000..33ecae8
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-custom/a-missing-customprovider-holds-the-fleet-where-it-is.yaml
@@ -0,0 +1,15 @@
+"name": "a-missing-customprovider-holds-the-fleet-where-it-is"
+"description": "a missing customProvider holds the fleet where it is"
+"inputs":
+  "active": "15"
+  "minReplicas": 1
+  "maxReplicas": 20
+  "readyReplicas": 8
+  "specReplicas": 8
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1"
+  "absentCustomProvider": true
+  "pinAPIVersion": true
+"expected":
+  "replicas": 8
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-custom/an-unreadable-decision-holds-the-fleet-where-it-is.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-custom/an-unreadable-decision-holds-the-fleet-where-it-is.yaml
new file mode 100644
index 0000000..26c0199
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-custom/an-unreadable-decision-holds-the-fleet-where-it-is.yaml
@@ -0,0 +1,13 @@
+"name": "an-unreadable-decision-holds-the-fleet-where-it-is"
+"description": "an unreadable decision holds the fleet where it is"
+"inputs":
+  "active": "15"
+  "minReplicas": 1
+  "maxReplicas": 20
+  "readyReplicas": 8
+  "specReplicas": 8
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "apiVersion": "some.other.api/v1"
+"expected":
+  "replicas": 8
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-custom/the-decision-is-used-as-is-not-scaled-by-readyreplicas.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-custom/the-decision-is-used-as-is-not-scaled-by-readyreplicas.yaml
new file mode 100644
index 0000000..b8bb5a7
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-custom/the-decision-is-used-as-is-not-scaled-by-readyreplicas.yaml
@@ -0,0 +1,15 @@
+# 8 ready replicas would turn a per-replica metric into a much larger number; a decision of 15 is already the answer.
+"name": "the-decision-is-used-as-is-not-scaled-by-readyreplicas"
+"description": "the decision is used as-is, not scaled by readyReplicas"
+"inputs":
+  "active": "15"
+  "minReplicas": 1
+  "maxReplicas": 20
+  "readyReplicas": 8
+  "specReplicas": 8
+  "serviceId": "svc"
+  "namespace": "modelforge"
+  "apiVersion": "llmscaling.inference.x-k8s.io/v1alpha1"
+  "pinAPIVersion": true
+"expected":
+  "replicas": 15
diff --git a/internal/controller/autoscaling/testdata/helpers/recommend-default-prometheus/an-empty-metricprovider-still-queries-prometheus.yaml b/internal/controller/autoscaling/testdata/helpers/recommend-default-prometheus/an-empty-metricprovider-still-queries-prometheus.yaml
new file mode 100644
index 0000000..1e19cba
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/recommend-default-prometheus/an-empty-metricprovider-still-queries-prometheus.yaml
@@ -0,0 +1,16 @@
+# ceil(4 * 1.6/0.8) = 8, the per-replica ratio the Custom path skips.
+"name": "an-empty-metricprovider-still-queries-prometheus"
+"description": "an empty metricProvider still queries Prometheus"
+"inputs":
+  "sample": "1.6"
+  "readyReplicas": 4
+  "specReplicas": 4
+  "minReplicas": 1
+  "maxReplicas": 20
+  "metrics":
+  - "name": "kv-cache"
+    "query": "avg(kv)"
+    "target": "0.8"
+"expected":
+  "replicas": 8
+  "path": "/api/v1/query"
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/caps-a-burst-scale-down.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/caps-a-burst-scale-down.yaml
new file mode 100644
index 0000000..b03f424
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/caps-a-burst-scale-down.yaml
@@ -0,0 +1,8 @@
+"name": "caps-a-burst-scale-down"
+"description": "caps a burst scale-down"
+"inputs":
+  "current": 10
+  "desired": 1
+  "maxStep": 2
+"expected":
+  "want": 8
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-a-no-op-untouched.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-a-no-op-untouched.yaml
new file mode 100644
index 0000000..f5c8774
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-a-no-op-untouched.yaml
@@ -0,0 +1,8 @@
+"name": "leaves-a-no-op-untouched"
+"description": "leaves a no-op untouched"
+"inputs":
+  "current": 5
+  "desired": 5
+  "maxStep": 1
+"expected":
+  "want": 5
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-scale-up-untouched.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-scale-up-untouched.yaml
new file mode 100644
index 0000000..39eef98
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/leaves-scale-up-untouched.yaml
@@ -0,0 +1,8 @@
+"name": "leaves-scale-up-untouched"
+"description": "leaves scale-up untouched"
+"inputs":
+  "current": 2
+  "desired": 10
+  "maxStep": 1
+"expected":
+  "want": 10
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-exactly-at-the-cap.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-exactly-at-the-cap.yaml
new file mode 100644
index 0000000..72d38df
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-exactly-at-the-cap.yaml
@@ -0,0 +1,8 @@
+"name": "passes-through-a-step-exactly-at-the-cap"
+"description": "passes through a step exactly at the cap"
+"inputs":
+  "current": 10
+  "desired": 8
+  "maxStep": 2
+"expected":
+  "want": 8
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-within-the-cap.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-within-the-cap.yaml
new file mode 100644
index 0000000..e6123bd
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/passes-through-a-step-within-the-cap.yaml
@@ -0,0 +1,8 @@
+"name": "passes-through-a-step-within-the-cap"
+"description": "passes through a step within the cap"
+"inputs":
+  "current": 10
+  "desired": 9
+  "maxStep": 2
+"expected":
+  "want": 9
diff --git a/internal/controller/autoscaling/testdata/helpers/scale-guard/unlimited-when-maxstep-is-0.yaml b/internal/controller/autoscaling/testdata/helpers/scale-guard/unlimited-when-maxstep-is-0.yaml
new file mode 100644
index 0000000..a1aa396
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/scale-guard/unlimited-when-maxstep-is-0.yaml
@@ -0,0 +1,8 @@
+"name": "unlimited-when-maxstep-is-0"
+"description": "unlimited when maxStep is 0"
+"inputs":
+  "current": 10
+  "desired": 1
+  "maxStep": 0
+"expected":
+  "want": 1
diff --git a/internal/controller/autoscaling/testdata/helpers/stabilization/peak-holds-inside-the-window-then-expires.yaml b/internal/controller/autoscaling/testdata/helpers/stabilization/peak-holds-inside-the-window-then-expires.yaml
new file mode 100644
index 0000000..472901c
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/stabilization/peak-holds-inside-the-window-then-expires.yaml
@@ -0,0 +1,19 @@
+# First recommendation sticks, a dip inside the window holds the peak, the peak expires after the window, and a scale-up is immediate.
+"name": "peak-holds-inside-the-window-then-expires"
+"description": "peak holds inside the window then expires"
+"inputs":
+  "window": "60s"
+  "ops":
+  - "at": "0s"
+    "recommend": 5
+    "want": 5
+  - "at": "10s"
+    "recommend": 1
+    "want": 5
+  - "at": "61s"
+    "recommend": 1
+    "want": 1
+  - "at": "61s"
+    "recommend": 9
+    "want": 9
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/stabilization/pin-after-a-scale-down-paces-the-next-step.yaml b/internal/controller/autoscaling/testdata/helpers/stabilization/pin-after-a-scale-down-paces-the-next-step.yaml
new file mode 100644
index 0000000..d6a4768
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/stabilization/pin-after-a-scale-down-paces-the-next-step.yaml
@@ -0,0 +1,32 @@
+# Pinning the written count after a scale-down paces the next one a full window later.
+"name": "pin-after-a-scale-down-paces-the-next-step"
+"description": "pin after a scale-down paces the next step"
+"inputs":
+  "window": "300s"
+  "ops":
+  - "at": "0s"
+    "recommend": 8
+    "want": 8
+  - "at": "301s"
+    "recommend": 1
+    "want": 1
+  - "at": "301s"
+    "holdAt": 6
+  - "at": "311s"
+    "recommend": 1
+    "want": 6
+  - "at": "400s"
+    "recommend": 1
+    "want": 6
+  - "at": "600s"
+    "recommend": 1
+    "want": 6
+  - "at": "602s"
+    "recommend": 1
+    "want": 1
+  - "at": "700s"
+    "holdAt": 6
+  - "at": "710s"
+    "recommend": 9
+    "want": 9
+"expected": {}
diff --git a/internal/controller/autoscaling/testdata/helpers/stabilization/window-0-makes-the-pin-a-no-op.yaml b/internal/controller/autoscaling/testdata/helpers/stabilization/window-0-makes-the-pin-a-no-op.yaml
new file mode 100644
index 0000000..2354b7e
--- /dev/null
+++ b/internal/controller/autoscaling/testdata/helpers/stabilization/window-0-makes-the-pin-a-no-op.yaml
@@ -0,0 +1,12 @@
+# With stabilization off there is nothing to pace with, so the pin is a no-op.
+"name": "window-0-makes-the-pin-a-no-op"
+"description": "window 0 makes the pin a no-op"
+"inputs":
+  "window": "0s"
+  "ops":
+  - "at": "0s"
+    "holdAt": 6
+  - "at": "0s"
+    "recommend": 1
+    "want": 1
+"expected": {}
diff --git a/internal/controller/serving/llmservice_cases_test.go b/internal/controller/serving/llmservice_cases_test.go
new file mode 100644
index 0000000..69ccaf7
--- /dev/null
+++ b/internal/controller/serving/llmservice_cases_test.go
@@ -0,0 +1,892 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package serving
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"maps"
+	"time"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+	appsv1 "k8s.io/api/apps/v1"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	"k8s.io/apimachinery/pkg/api/meta"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/apimachinery/pkg/types"
+	"k8s.io/client-go/tools/events"
+	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
+	"sigs.k8s.io/controller-runtime/pkg/reconcile"
+
+	servingv1alpha1 "github.com/modelsphere/llm-operator/api/serving/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/helm"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+)
+
+// serviceCase is one LLMService controller spec. HistoryLimit defaults to 10.
+// PendingTimeout zero (omitted) leaves the controller default.
+type serviceCase struct {
+	Name           string          `json:"name,omitempty"`
+	Description    string          `json:"description,omitempty"`
+	HistoryLimit   int             `json:"historyLimit,omitempty"`
+	PendingTimeout *cases.Duration `json:"pendingTimeout,omitempty"`
+	Steps          []serviceStep   `json:"steps"`
+}
+
+type serviceStep struct {
+	Action         string            `json:"action"`
+	FreshNamespace bool              `json:"freshNamespace,omitempty"`
+	Name           string            `json:"name,omitempty"`
+	Version        string            `json:"version,omitempty"`
+	Annotations    map[string]string `json:"annotations,omitempty"`
+	Labels         map[string]string `json:"labels,omitempty"`
+	Finalizers     []string          `json:"finalizers,omitempty"`
+	ChartVersion   *string           `json:"chartVersion,omitempty"`
+	Suspend        *bool             `json:"suspend,omitempty"`
+	Layers         []layerSpec       `json:"layers,omitempty"`
+	Release        *releaseSpec      `json:"release,omitempty"`
+	Entries        []releaseSpec     `json:"entries,omitempty"`
+	Put            string            `json:"put,omitempty"`
+	Versions       []string          `json:"versions,omitempty"`
+	ListErr        *string           `json:"listErr,omitempty"`
+	InstallErr     *string           `json:"installErr,omitempty"`
+	UpgradeErr     *string           `json:"upgradeErr,omitempty"`
+	GetErr         *string           `json:"getErr,omitempty"`
+	KeepFailed     *bool             `json:"keepFailed,omitempty"`
+	Repos          []string          `json:"repos,omitempty"`
+	Status         *statusSpec       `json:"status,omitempty"`
+	Capture        []string          `json:"capture,omitempty"`
+	Expect         *expectSpec       `json:"expect,omitempty"`
+}
+
+type layerSpec struct {
+	Name   string `json:"name"`
+	Values string `json:"values,omitempty"`
+}
+
+type releaseSpec struct {
+	Revision     int               `json:"revision,omitempty"`
+	Status       string            `json:"status,omitempty"`
+	ChartName    string            `json:"chartName,omitempty"`
+	ChartVersion string            `json:"chartVersion,omitempty"`
+	Config       map[string]any    `json:"config,omitempty"`
+	Labels       map[string]string `json:"labels,omitempty"`
+	Owned        bool              `json:"owned,omitempty"`
+	Slot         bool              `json:"slot,omitempty"`
+	Updated      *cases.Duration   `json:"updated,omitempty"`
+}
+
+type statusSpec struct {
+	Phase      string          `json:"phase,omitempty"`
+	Conditions []conditionSpec `json:"conditions,omitempty"`
+}
+
+type conditionSpec struct {
+	Type               string `json:"type"`
+	Status             string `json:"status,omitempty"`
+	Reason             string `json:"reason,omitempty"`
+	Message            string `json:"message,omitempty"`
+	ObservedGeneration string `json:"observedGeneration,omitempty"`
+	Absent             bool   `json:"absent,omitempty"`
+}
+
+type expectSpec struct {
+	Mid                  *expectSpec      `json:"mid,omitempty"`
+	Error                *bool            `json:"error,omitempty"`
+	ErrorContains        string           `json:"errorContains,omitempty"`
+	Installs             *int             `json:"installs,omitempty"`
+	Upgrades             *int             `json:"upgrades,omitempty"`
+	Gets                 *int             `json:"gets,omitempty"`
+	Uninstalls           *int             `json:"uninstalls,omitempty"`
+	Rollbacks            *int             `json:"rollbacks,omitempty"`
+	LastRollback         *int             `json:"lastRollback,omitempty"`
+	HelmTouched          *int             `json:"helmTouched,omitempty"`
+	ListCalls            *int             `json:"listCalls,omitempty"`
+	LastForce            *bool            `json:"lastForce,omitempty"`
+	GetsUnchanged        *bool            `json:"getsUnchanged,omitempty"`
+	UpgradesUnchanged    *bool            `json:"upgradesUnchanged,omitempty"`
+	Phase                string           `json:"phase,omitempty"`
+	ChartVersion         string           `json:"chartVersion,omitempty"`
+	ReResolve            *string          `json:"reResolve,omitempty"`
+	AppliedHash          *hashExpect      `json:"appliedHash,omitempty"`
+	HelmRevision         *int             `json:"helmRevision,omitempty"`
+	ObservedGeneration   string           `json:"observedGeneration,omitempty"`
+	Generation           *int64           `json:"generation,omitempty"`
+	History              *historyExpect   `json:"history,omitempty"`
+	Conditions           []conditionSpec  `json:"conditions,omitempty"`
+	ReleaseUID           string           `json:"releaseUID,omitempty"`
+	Release              *releaseExpect   `json:"release,omitempty"`
+	NotFound             *bool            `json:"notFound,omitempty"`
+	Message              *messageExpect   `json:"message,omitempty"`
+	DistinctRevisions    *int             `json:"distinctRevisions,omitempty"`
+	Snapshots            []snapshotExpect `json:"snapshots,omitempty"`
+	Requeue              *requeueExpect   `json:"requeue,omitempty"`
+	EventContains        string           `json:"eventContains,omitempty"`
+	Finalizer            string           `json:"finalizer,omitempty"`
+	PlantedRevisionUsed  *bool            `json:"plantedRevisionUsed,omitempty"`
+	CapturedRevisionGone *bool            `json:"capturedRevisionGone,omitempty"`
+	LatestRevisionExists *bool            `json:"latestRevisionExists,omitempty"`
+}
+
+type hashExpect struct {
+	Prefix   string `json:"prefix,omitempty"`
+	Empty    *bool  `json:"empty,omitempty"`
+	Captured *bool  `json:"captured,omitempty"`
+}
+
+type historyExpect struct {
+	Length  *int                 `json:"length,omitempty"`
+	Entries []historyEntryExpect `json:"entries,omitempty"`
+}
+
+type historyEntryExpect struct {
+	Index              int      `json:"index"`
+	Action             string   `json:"action,omitempty"`
+	Revision           *int     `json:"revision,omitempty"`
+	AnnotationsPresent []string `json:"annotationsPresent,omitempty"`
+	AnnotationsAbsent  []string `json:"annotationsAbsent,omitempty"`
+	ReResolved         *bool    `json:"reResolved,omitempty"`
+	ForceConflicts     *bool    `json:"forceConflicts,omitempty"`
+	HashEqualsApplied  bool     `json:"hashEqualsApplied,omitempty"`
+	HashEqualsIndex    *int     `json:"hashEqualsIndex,omitempty"`
+}
+
+type releaseExpect struct {
+	Exists       *bool             `json:"exists,omitempty"`
+	Status       string            `json:"status,omitempty"`
+	Labels       map[string]string `json:"labels,omitempty"`
+	LabelsAbsent []string          `json:"labelsAbsent,omitempty"`
+}
+
+type messageExpect struct {
+	Contains    []string `json:"contains,omitempty"`
+	Equals      string   `json:"equals,omitempty"`
+	NotContains []string `json:"notContains,omitempty"`
+}
+
+type snapshotExpect struct {
+	Index             int               `json:"index"`
+	Annotations       map[string]string `json:"annotations,omitempty"`
+	AnnotationsAbsent []string          `json:"annotationsAbsent,omitempty"`
+	ChartName         string            `json:"chartName,omitempty"`
+	Labels            map[string]string `json:"labels,omitempty"`
+	ControllerOwner   *bool             `json:"controllerOwner,omitempty"`
+}
+
+type requeueExpect struct {
+	Exact   *cases.Duration `json:"exact,omitempty"`
+	AtLeast *cases.Duration `json:"atLeast,omitempty"`
+	Above   *cases.Duration `json:"above,omitempty"`
+	AtMost  *cases.Duration `json:"atMost,omitempty"`
+}
+
+type helmSnap struct {
+	gets     int
+	upgrades int
+}
+
+type caseRun struct {
+	tc                 serviceCase
+	ns                 string
+	name               string
+	h                  *fakeHelm
+	l                  *versionList
+	r                  *LLMServiceReconciler
+	events             chan string
+	capturedRevision   string
+	appliedHash        string
+	observedGeneration int64
+	generation         int64
+	planted            string
+}
+
+func runServiceCase(tc serviceCase) {
+	rec := events.NewFakeRecorder(32)
+	h := &fakeHelm{}
+	l := &versionList{versions: []string{ver086}}
+	limit := tc.HistoryLimit
+	if limit == 0 {
+		limit = 10
+	}
+	r := newReconciler(h, l, limit)
+	r.Recorder = rec
+	if tc.PendingTimeout != nil {
+		r.PendingTimeout = tc.PendingTimeout.Std()
+	}
+	run := &caseRun{
+		tc:     tc,
+		ns:     takeNS(),
+		name:   "qwen",
+		h:      h,
+		l:      l,
+		r:      r,
+		events: rec.Events,
+	}
+	Expect(tc.Steps).NotTo(BeEmpty(), tc.Name)
+	for i, step := range tc.Steps {
+		run.step(i, step)
+	}
+}
+
+func (c *caseRun) step(i int, step serviceStep) {
+	if step.FreshNamespace {
+		c.ns = takeNS()
+	}
+	if step.Name != "" {
+		c.name = step.Name
+	}
+	if step.Action == "" {
+		Fail(fmt.Sprintf("%s: step %d has no action", c.tc.Name, i+1))
+	}
+	if stepRequiresExpect(step.Action) && step.Expect == nil {
+		Fail(fmt.Sprintf("%s: step %d (%s) missing expect", c.tc.Name, i+1, step.Action))
+	}
+	switch step.Action {
+	case "create":
+		c.create(step)
+	case "reconcile":
+		c.reconcile(step.Expect)
+	case "update":
+		c.update(step)
+	case "delete":
+		c.deleteObj()
+	case "putRelease":
+		c.putRelease(step)
+	case "setVersions":
+		c.setVersions(step)
+	case "setHelm":
+		c.setHelm(step)
+	case "setAllowed":
+		c.r.AllowedChartRepos = append([]string(nil), step.Repos...)
+	case "deleteRelease":
+		delete(c.h.releases, c.h.slot(c.ns, c.name))
+	case "setHistory":
+		c.setHistory(step)
+	case "plantStale":
+		c.plantStale()
+	case "reconcileHeld":
+		c.reconcileHeld(step.Expect)
+	case "capture":
+		c.capture(step)
+	case "setStatus":
+		c.setStatus(step)
+	case "setStatusReleaseUID":
+		c.setStatusReleaseUID()
+	default:
+		Fail(fmt.Sprintf("%s: step %d unknown action %q", c.tc.Name, i+1, step.Action))
+	}
+	if step.Expect != nil && step.Action != "reconcile" && step.Action != "reconcileHeld" {
+		c.check(step.Expect, reconcile.Result{}, nil, helmSnap{}, false)
+	}
+}
+
+func stepRequiresExpect(action string) bool {
+	return action == "reconcile" || action == "reconcileHeld"
+}
+
+func (c *caseRun) create(step serviceStep) {
+	Expect(step.Version).NotTo(BeEmpty())
+	obj := newService(c.ns, c.name, step.Version)
+	maps.Copy(obj.Annotations, step.Annotations)
+	if step.Finalizers != nil {
+		obj.Finalizers = append([]string(nil), step.Finalizers...)
+	}
+	Expect(k8sClient.Create(ctx, obj)).To(Succeed())
+}
+
+func (c *caseRun) update(step serviceStep) {
+	obj := fetch(c.ns, c.name)
+	if step.Annotations != nil {
+		if obj.Annotations == nil {
+			obj.Annotations = map[string]string{}
+		}
+		maps.Copy(obj.Annotations, step.Annotations)
+	}
+	if step.Labels != nil {
+		if obj.Labels == nil {
+			obj.Labels = map[string]string{}
+		}
+		maps.Copy(obj.Labels, step.Labels)
+	}
+	if step.ChartVersion != nil {
+		obj.Spec.Chart.Version = *step.ChartVersion
+	}
+	if step.Suspend != nil {
+		obj.Spec.Suspend = *step.Suspend
+	}
+	if step.Layers != nil {
+		obj.Spec.Layers = layersFrom(step.Layers)
+	}
+	Expect(k8sClient.Update(ctx, obj)).To(Succeed())
+}
+
+func layersFrom(in []layerSpec) []servingv1alpha1.Layer {
+	out := make([]servingv1alpha1.Layer, len(in))
+	for i, l := range in {
+		out[i] = servingv1alpha1.Layer{Name: l.Name}
+		if l.Values != "" {
+			out[i].Values = &apiextensionsv1.JSON{Raw: []byte(l.Values)}
+		}
+	}
+	return out
+}
+
+func (c *caseRun) deleteObj() {
+	Expect(k8sClient.Delete(ctx, fetch(c.ns, c.name))).To(Succeed())
+}
+
+func (c *caseRun) reconcile(exp *expectSpec) {
+	before := helmSnap{gets: c.h.gets, upgrades: c.h.upgrades}
+	result, err := doReconcile(c.r, c.ns, c.name)
+	c.check(exp, result, err, before, true)
+}
+
+// reconcileHeld runs one reconcile that blocks inside helm, checks the
+// mid-apply status, then lets the install finish.
+func (c *caseRun) reconcileHeld(exp *expectSpec) {
+	Expect(exp).NotTo(BeNil())
+	Expect(exp.Mid).NotTo(BeNil())
+	c.h.hold = make(chan struct{})
+	c.h.entered = make(chan struct{}, 1)
+	released := false
+	release := func() {
+		if !released {
+			close(c.h.hold)
+			released = true
+		}
+	}
+	defer release()
+
+	done := make(chan error, 1)
+	go func() {
+		defer GinkgoRecover()
+		_, recErr := doReconcile(c.r, c.ns, c.name)
+		done <- recErr
+	}()
+	Eventually(c.h.entered).WithTimeout(5 * time.Second).Should(Receive())
+	c.check(exp.Mid, reconcile.Result{}, nil, helmSnap{}, false)
+	release()
+	var recErr error
+	Eventually(done).WithTimeout(5 * time.Second).Should(Receive(&recErr))
+	Expect(recErr).NotTo(HaveOccurred())
+	c.check(exp, reconcile.Result{}, recErr, helmSnap{}, true)
+}
+
+func (c *caseRun) putRelease(step serviceStep) {
+	Expect(step.Release).NotTo(BeNil())
+	rel := c.releaseFrom(*step.Release)
+	c.h.put(c.ns, c.name, &rel)
+}
+
+func (c *caseRun) setHistory(step serviceStep) {
+	Expect(step.Entries).NotTo(BeEmpty())
+	hist := make([]helm.Release, len(step.Entries))
+	for i, e := range step.Entries {
+		hist[i] = c.releaseFrom(e)
+	}
+	c.h.histories = map[string][]helm.Release{c.h.slot(c.ns, c.name): hist}
+	if step.Put == "" {
+		return
+	}
+	Expect(step.Put).To(Equal("last"))
+	last := hist[len(hist)-1]
+	c.h.put(c.ns, c.name, &last)
+}
+
+func (c *caseRun) releaseFrom(in releaseSpec) helm.Release {
+	if in.Slot {
+		cur := c.h.releases[c.h.slot(c.ns, c.name)]
+		Expect(cur).NotTo(BeNil())
+		return *cur
+	}
+	if in.Owned {
+		updated := time.Now()
+		if in.Updated != nil {
+			updated = time.Now().Add(in.Updated.Std())
+		}
+		return *ownedRel(fetch(c.ns, c.name), in.Revision, in.Status, updated)
+	}
+	rel := helm.Release{
+		Revision:     in.Revision,
+		Status:       in.Status,
+		ChartName:    in.ChartName,
+		ChartVersion: in.ChartVersion,
+		Config:       in.Config,
+		Labels:       in.Labels,
+	}
+	if in.Updated != nil {
+		rel.Updated = time.Now().Add(in.Updated.Std())
+	}
+	return rel
+}
+
+func (c *caseRun) setVersions(step serviceStep) {
+	if step.Versions != nil {
+		c.l.versions = append([]string(nil), step.Versions...)
+	}
+	if step.ListErr != nil {
+		c.l.err = errFrom(*step.ListErr)
+	}
+}
+
+func (c *caseRun) setHelm(step serviceStep) {
+	if step.InstallErr != nil {
+		c.h.installErr = errFrom(*step.InstallErr)
+	}
+	if step.UpgradeErr != nil {
+		c.h.upgradeErr = errFrom(*step.UpgradeErr)
+	}
+	if step.GetErr != nil {
+		c.h.getErr = errFrom(*step.GetErr)
+	}
+	if step.KeepFailed != nil {
+		c.h.keepFailed = *step.KeepFailed
+	}
+}
+
+func errFrom(msg string) error {
+	if msg == "" {
+		return nil
+	}
+	return errors.New(msg)
+}
+
+func (c *caseRun) setStatus(step serviceStep) {
+	Expect(step.Status).NotTo(BeNil())
+	obj := fetch(c.ns, c.name)
+	if step.Status.Phase != "" {
+		obj.Status.Phase = servingv1alpha1.Phase(step.Status.Phase)
+	}
+	for _, cond := range step.Status.Conditions {
+		meta.SetStatusCondition(&obj.Status.Conditions, conditionFrom(obj, cond))
+	}
+	Expect(k8sClient.Status().Update(ctx, obj)).To(Succeed())
+}
+
+func conditionFrom(obj *servingv1alpha1.LLMService, cond conditionSpec) metav1.Condition {
+	out := metav1.Condition{
+		Type:    cond.Type,
+		Status:  metav1.ConditionStatus(cond.Status),
+		Reason:  cond.Reason,
+		Message: cond.Message,
+	}
+	switch cond.ObservedGeneration {
+	case "":
+	case "current":
+		out.ObservedGeneration = obj.Generation
+	default:
+		Fail("unsupported observedGeneration " + cond.ObservedGeneration)
+	}
+	return out
+}
+
+func (c *caseRun) setStatusReleaseUID() {
+	obj := fetch(c.ns, c.name)
+	obj.Status.ReleaseUID = string(obj.UID)
+	Expect(k8sClient.Status().Update(ctx, obj)).To(Succeed())
+}
+
+func (c *caseRun) capture(step serviceStep) {
+	Expect(step.Capture).NotTo(BeEmpty())
+	obj := fetch(c.ns, c.name)
+	for _, what := range step.Capture {
+		switch what {
+		case "generation":
+			c.generation = obj.Generation
+		case "observedGeneration":
+			c.observedGeneration = obj.Status.ObservedGeneration
+		case "appliedHash":
+			c.appliedHash = obj.Status.AppliedHash
+		case "revision":
+			Expect(obj.Status.History).NotTo(BeEmpty())
+			c.capturedRevision = obj.Status.History[0].ControllerRevision
+		default:
+			Fail("unknown capture " + what)
+		}
+	}
+}
+
+// plantStale leaves a ControllerRevision under the name the next install
+// reuses, owned by the service, so the reconcile replaces that snapshot.
+func (c *caseRun) plantStale() {
+	obj := fetch(c.ns, c.name)
+	values, err := mergeLayers(obj.Spec.Layers)
+	Expect(err).NotTo(HaveOccurred())
+	hash, err := appliedHash(chartName, ver086, values)
+	Expect(err).NotTo(HaveOccurred())
+	stale := &appsv1.ControllerRevision{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      controllerRevisionName(obj.Name, hash, 1),
+			Namespace: c.ns,
+		},
+		Data:     runtime.RawExtension{Raw: []byte(`{"annotations":{"swiss.modelsphere.dev/action":"old"}}`)},
+		Revision: 1,
+	}
+	Expect(controllerutil.SetControllerReference(obj, stale, k8sClient.Scheme())).To(Succeed())
+	Expect(k8sClient.Create(ctx, stale)).To(Succeed())
+	c.planted = stale.Name
+}
+
+func (c *caseRun) check(exp *expectSpec, result reconcile.Result, err error, before helmSnap, reconciled bool) {
+	if exp == nil {
+		return
+	}
+	c.checkErr(exp, err, reconciled)
+	c.checkRequeue(exp.Requeue, result)
+	c.checkHelm(exp, before)
+	if exp.EventContains != "" {
+		Expect(c.events).To(Receive(ContainSubstring(exp.EventContains)))
+	}
+	if exp.NotFound != nil && *exp.NotFound {
+		c.checkMissing()
+	}
+	var got *servingv1alpha1.LLMService
+	if exp.wantsObject() {
+		got = fetch(c.ns, c.name)
+		c.checkObject(exp, got)
+	}
+	if exp.Release != nil {
+		c.checkRelease(exp.Release, got)
+	}
+}
+
+func (c *caseRun) checkErr(exp *expectSpec, err error, reconciled bool) {
+	if exp.ErrorContains != "" {
+		Expect(err).To(MatchError(ContainSubstring(exp.ErrorContains)))
+		return
+	}
+	if exp.Error != nil {
+		if *exp.Error {
+			Expect(err).To(HaveOccurred())
+		} else {
+			Expect(err).NotTo(HaveOccurred())
+		}
+		return
+	}
+	if reconciled {
+		Expect(err).NotTo(HaveOccurred())
+	}
+}
+
+func (c *caseRun) checkRequeue(exp *requeueExpect, result reconcile.Result) {
+	if exp == nil {
+		return
+	}
+	if exp.Exact != nil {
+		Expect(result.RequeueAfter).To(Equal(exp.Exact.Std()))
+	}
+	if exp.AtLeast != nil {
+		Expect(result.RequeueAfter).To(BeNumerically(">=", exp.AtLeast.Std()))
+	}
+	if exp.Above != nil {
+		Expect(result.RequeueAfter).To(BeNumerically(">", exp.Above.Std()))
+	}
+	if exp.AtMost != nil {
+		Expect(result.RequeueAfter).To(BeNumerically("<=", exp.AtMost.Std()))
+	}
+}
+
+func (c *caseRun) checkHelm(exp *expectSpec, before helmSnap) {
+	expectInt(c.h.installs, exp.Installs)
+	expectInt(c.h.upgrades, exp.Upgrades)
+	expectInt(c.h.gets, exp.Gets)
+	expectInt(c.h.uninstalls, exp.Uninstalls)
+	expectInt(c.h.rollbacks, exp.Rollbacks)
+	expectInt(c.h.lastRollback, exp.LastRollback)
+	expectInt(c.l.calls, exp.ListCalls)
+	if exp.HelmTouched != nil {
+		Expect(c.h.gets + c.h.installs + c.h.upgrades).To(Equal(*exp.HelmTouched))
+	}
+	expectBool(c.h.lastForce, exp.LastForce)
+	if exp.GetsUnchanged != nil && *exp.GetsUnchanged {
+		Expect(c.h.gets).To(Equal(before.gets))
+	}
+	if exp.UpgradesUnchanged != nil && *exp.UpgradesUnchanged {
+		Expect(c.h.upgrades).To(Equal(before.upgrades))
+	}
+}
+
+func expectInt(got int, want *int) {
+	if want != nil {
+		Expect(got).To(Equal(*want))
+	}
+}
+
+func expectBool(got bool, want *bool) {
+	if want == nil {
+		return
+	}
+	if *want {
+		Expect(got).To(BeTrue())
+	} else {
+		Expect(got).To(BeFalse())
+	}
+}
+
+func (e *expectSpec) wantsObject() bool {
+	if e.Phase != "" || e.ChartVersion != "" || e.ReResolve != nil || e.AppliedHash != nil {
+		return true
+	}
+	if e.HelmRevision != nil || e.ObservedGeneration != "" || e.Generation != nil || e.History != nil {
+		return true
+	}
+	if len(e.Conditions) > 0 || e.ReleaseUID != "" || e.Message != nil || e.Finalizer != "" {
+		return true
+	}
+	if e.DistinctRevisions != nil || len(e.Snapshots) > 0 || e.PlantedRevisionUsed != nil {
+		return true
+	}
+	return e.CapturedRevisionGone != nil || e.LatestRevisionExists != nil
+}
+
+func (c *caseRun) checkObject(exp *expectSpec, got *servingv1alpha1.LLMService) {
+	if exp.Phase != "" {
+		Expect(got.Status.Phase).To(Equal(servingv1alpha1.Phase(exp.Phase)))
+	}
+	if exp.ChartVersion != "" {
+		Expect(got.Status.Chart.Version).To(Equal(exp.ChartVersion))
+	}
+	if exp.ReResolve != nil {
+		Expect(got.Status.Chart.ReResolve).To(Equal(*exp.ReResolve))
+	}
+	c.checkHash(exp.AppliedHash, got)
+	if exp.HelmRevision != nil {
+		Expect(got.Status.Helm.Revision).To(Equal(*exp.HelmRevision))
+	}
+	c.checkGeneration(exp, got)
+	c.checkHistory(exp, got)
+	c.checkConditions(exp.Conditions, got)
+	c.checkReleaseUID(exp.ReleaseUID, got)
+	c.checkMessage(exp.Message, got)
+	c.checkFinalizer(exp.Finalizer, got)
+	c.checkRevisions(exp, got)
+}
+
+func (c *caseRun) checkHash(exp *hashExpect, got *servingv1alpha1.LLMService) {
+	if exp == nil {
+		return
+	}
+	if exp.Empty != nil && *exp.Empty {
+		Expect(got.Status.AppliedHash).To(BeEmpty())
+	}
+	if exp.Prefix != "" {
+		Expect(got.Status.AppliedHash).To(HavePrefix(exp.Prefix))
+	}
+	if exp.Captured != nil && *exp.Captured {
+		Expect(got.Status.AppliedHash).To(Equal(c.appliedHash))
+	}
+}
+
+func (c *caseRun) checkGeneration(exp *expectSpec, got *servingv1alpha1.LLMService) {
+	if exp.Generation != nil {
+		Expect(got.Generation).To(Equal(*exp.Generation))
+	}
+	switch exp.ObservedGeneration {
+	case "":
+	case "current":
+		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
+	case "captured":
+		Expect(got.Status.ObservedGeneration).To(Equal(c.observedGeneration))
+	case "nonzero":
+		Expect(got.Status.ObservedGeneration).NotTo(BeZero())
+	case "generation":
+		Expect(got.Generation).To(Equal(c.generation))
+	default:
+		Fail("unsupported observedGeneration " + exp.ObservedGeneration)
+	}
+}
+
+func (c *caseRun) checkHistory(exp *expectSpec, got *servingv1alpha1.LLMService) {
+	if exp.History == nil {
+		return
+	}
+	if exp.History.Length != nil {
+		Expect(got.Status.History).To(HaveLen(*exp.History.Length))
+	}
+	for _, e := range exp.History.Entries {
+		Expect(e.Index).To(BeNumerically(">=", 0))
+		Expect(e.Index).To(BeNumerically("<", len(got.Status.History)))
+		entry := got.Status.History[e.Index]
+		if e.Action != "" {
+			Expect(entry.Action).To(Equal(e.Action))
+		}
+		if e.Revision != nil {
+			Expect(entry.Revision).To(Equal(*e.Revision))
+		}
+		for _, k := range e.AnnotationsPresent {
+			Expect(entry.Annotations).To(HaveKey(k))
+		}
+		for _, k := range e.AnnotationsAbsent {
+			Expect(entry.Annotations).NotTo(HaveKey(k))
+		}
+		expectBool(entry.ReResolved, e.ReResolved)
+		expectBool(entry.ForceConflicts, e.ForceConflicts)
+		if e.HashEqualsApplied {
+			Expect(entry.Hash).To(Equal(got.Status.AppliedHash))
+		}
+		if e.HashEqualsIndex != nil {
+			Expect(entry.Hash).To(Equal(got.Status.History[*e.HashEqualsIndex].Hash))
+		}
+	}
+	if exp.DistinctRevisions != nil {
+		names := map[string]struct{}{}
+		for _, entry := range got.Status.History {
+			names[entry.ControllerRevision] = struct{}{}
+		}
+		Expect(names).To(HaveLen(*exp.DistinctRevisions))
+	}
+}
+
+func (c *caseRun) checkConditions(conds []conditionSpec, got *servingv1alpha1.LLMService) {
+	for _, want := range conds {
+		gotCond := condition(got, want.Type)
+		if want.Absent {
+			Expect(gotCond).To(BeNil())
+			continue
+		}
+		Expect(gotCond).NotTo(BeNil())
+		if want.Status != "" {
+			Expect(gotCond.Status).To(Equal(metav1.ConditionStatus(want.Status)))
+		}
+		if want.Reason != "" {
+			Expect(gotCond.Reason).To(Equal(want.Reason))
+		}
+	}
+}
+
+func (c *caseRun) checkReleaseUID(want string, got *servingv1alpha1.LLMService) {
+	switch want {
+	case "":
+	case "self":
+		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
+	case "empty":
+		Expect(got.Status.ReleaseUID).To(BeEmpty())
+	default:
+		Fail("unsupported releaseUID " + want)
+	}
+}
+
+func (c *caseRun) checkMessage(exp *messageExpect, got *servingv1alpha1.LLMService) {
+	if exp == nil {
+		return
+	}
+	if exp.Equals != "" {
+		Expect(got.Status.Message).To(Equal(exp.Equals))
+	}
+	for _, sub := range exp.Contains {
+		Expect(got.Status.Message).To(ContainSubstring(sub))
+	}
+	for _, sub := range exp.NotContains {
+		Expect(got.Status.Message).NotTo(ContainSubstring(sub))
+	}
+}
+
+func (c *caseRun) checkFinalizer(want string, got *servingv1alpha1.LLMService) {
+	switch want {
+	case "":
+	case "present":
+		Expect(got.Finalizers).To(ContainElement(servingv1alpha1.Finalizer))
+	case "absent":
+		Expect(got.Finalizers).NotTo(ContainElement(servingv1alpha1.Finalizer))
+	default:
+		Fail("unsupported finalizer " + want)
+	}
+}
+
+func (c *caseRun) checkRevisions(exp *expectSpec, got *servingv1alpha1.LLMService) {
+	if exp.PlantedRevisionUsed != nil && *exp.PlantedRevisionUsed {
+		Expect(got.Status.History).NotTo(BeEmpty())
+		Expect(got.Status.History[0].ControllerRevision).To(Equal(c.planted))
+	}
+	if exp.CapturedRevisionGone != nil && *exp.CapturedRevisionGone {
+		err := k8sClient.Get(ctx, types.NamespacedName{Namespace: c.ns, Name: c.capturedRevision}, &appsv1.ControllerRevision{})
+		Expect(apierrors.IsNotFound(err)).To(BeTrue())
+	}
+	if exp.LatestRevisionExists != nil && *exp.LatestRevisionExists {
+		Expect(got.Status.History).NotTo(BeEmpty())
+		Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: c.ns, Name: got.Status.History[0].ControllerRevision}, &appsv1.ControllerRevision{})).To(Succeed())
+	}
+	for _, sn := range exp.Snapshots {
+		Expect(sn.Index).To(BeNumerically("<", len(got.Status.History)))
+		c.checkSnapshot(got.Status.History[sn.Index].ControllerRevision, sn)
+	}
+}
+
+func (c *caseRun) checkSnapshot(name string, sn snapshotExpect) {
+	var cr appsv1.ControllerRevision
+	Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: c.ns, Name: name}, &cr)).To(Succeed())
+	if sn.ControllerOwner != nil && *sn.ControllerOwner {
+		Expect(cr.OwnerReferences).NotTo(BeEmpty())
+		Expect(cr.OwnerReferences[0].Controller).NotTo(BeNil())
+		Expect(*cr.OwnerReferences[0].Controller).To(BeTrue())
+	}
+	for k, v := range sn.Labels {
+		Expect(cr.Labels).To(HaveKeyWithValue(k, v))
+	}
+	var snap revisionSnapshot
+	Expect(json.Unmarshal(cr.Data.Raw, &snap)).To(Succeed())
+	for k, v := range sn.Annotations {
+		Expect(snap.Annotations).To(HaveKeyWithValue(k, v))
+	}
+	for _, k := range sn.AnnotationsAbsent {
+		Expect(snap.Annotations).NotTo(HaveKey(k))
+	}
+	if sn.ChartName != "" {
+		Expect(snap.Spec.Chart.Name).To(Equal(sn.ChartName))
+	}
+}
+
+func (c *caseRun) checkRelease(exp *releaseExpect, got *servingv1alpha1.LLMService) {
+	slot := c.h.slot(c.ns, c.name)
+	rel := c.h.releases[slot]
+	if exp.Exists != nil {
+		if *exp.Exists {
+			Expect(c.h.releases).To(HaveKey(slot))
+		} else {
+			Expect(c.h.releases).NotTo(HaveKey(slot))
+		}
+	}
+	if exp.Status != "" {
+		Expect(rel).NotTo(BeNil())
+		Expect(rel.Status).To(Equal(exp.Status))
+	}
+	if len(exp.Labels) == 0 && len(exp.LabelsAbsent) == 0 {
+		return
+	}
+	Expect(rel).NotTo(BeNil())
+	for k, v := range exp.Labels {
+		if v == "self" {
+			if got == nil {
+				got = fetch(c.ns, c.name)
+			}
+			v = string(got.UID)
+		}
+		Expect(rel.Labels).To(HaveKeyWithValue(k, v))
+	}
+	for _, k := range exp.LabelsAbsent {
+		Expect(rel.Labels).NotTo(HaveKey(k))
+	}
+}
+
+func (c *caseRun) checkMissing() {
+	err := k8sClient.Get(ctx, types.NamespacedName{Namespace: c.ns, Name: c.name}, &servingv1alpha1.LLMService{})
+	Expect(apierrors.IsNotFound(err)).To(BeTrue())
+}
diff --git a/internal/controller/serving/llmservice_controller_test.go b/internal/controller/serving/llmservice_controller_test.go
index b1eb8fe..c69e874 100644
--- a/internal/controller/serving/llmservice_controller_test.go
+++ b/internal/controller/serving/llmservice_controller_test.go
@@ -19,28 +19,23 @@ package serving
 import (
 	"cmp"
 	"context"
-	"encoding/json"
-	"errors"
 	"fmt"
 	"slices"
 	"time"
 
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
-	appsv1 "k8s.io/api/apps/v1"
 	corev1 "k8s.io/api/core/v1"
 	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
-	apierrors "k8s.io/apimachinery/pkg/api/errors"
 	"k8s.io/apimachinery/pkg/api/meta"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
-	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/tools/events"
-	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
 	"sigs.k8s.io/controller-runtime/pkg/reconcile"
 
 	servingv1alpha1 "github.com/modelsphere/llm-operator/api/serving/v1alpha1"
 	"github.com/modelsphere/llm-operator/internal/helm"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
 const (
@@ -270,917 +265,12 @@ func condition(obj *servingv1alpha1.LLMService, typ string) *metav1.Condition {
 }
 
 var _ = Describe("LLMService controller", func() {
-	It("installs and records a ControllerRevision", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		obj := newService(ns, "qwen", ">=0.8.0")
-		Expect(k8sClient.Create(ctx, obj)).To(Succeed())
-
-		_, err := doReconcile(newReconciler(h, l, 10), ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-
-		got := fetch(ns, "qwen")
-		Expect(got.Finalizers).To(ContainElement(servingv1alpha1.Finalizer))
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(got.Status.Chart.Version).To(Equal(ver086))
-		Expect(got.Status.AppliedHash).To(HavePrefix("sha256:"))
-		Expect(got.Status.Helm.Revision).To(Equal(1))
-		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
-		Expect(got.Status.History).To(HaveLen(1))
-		Expect(got.Status.History[0].Action).To(Equal("install"))
-		Expect(got.Status.History[0].Annotations).To(HaveKey("swiss.modelsphere.dev/note"))
-		Expect(got.Status.History[0].Annotations).NotTo(HaveKey("unrelated"))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Status).To(Equal(metav1.ConditionTrue))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonInstalled))
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-		Expect(h.releases[h.slot(ns, "qwen")].Labels).To(HaveKeyWithValue(llmserviceUIDLabel, string(got.UID)))
-		Expect(h.releases[h.slot(ns, "qwen")].Labels).To(HaveKeyWithValue(llmserviceLabel, "qwen"))
-
-		var cr appsv1.ControllerRevision
-		Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: got.Status.History[0].ControllerRevision}, &cr)).To(Succeed())
-		Expect(cr.Labels).To(HaveKeyWithValue(llmserviceLabel, "qwen"))
-		Expect(cr.OwnerReferences).NotTo(BeEmpty())
-		Expect(cr.OwnerReferences[0].Controller).NotTo(BeNil())
-		Expect(*cr.OwnerReferences[0].Controller).To(BeTrue())
-		var snap revisionSnapshot
-		Expect(json.Unmarshal(cr.Data.Raw, &snap)).To(Succeed())
-		Expect(snap.Annotations).To(HaveKeyWithValue("swiss.modelsphere.dev/action", "install"))
-		Expect(snap.Annotations).NotTo(HaveKey("unrelated"))
-		Expect(snap.Spec.Chart.Name).To(Equal("sglang"))
-	})
-
-	It("does not call helm when the spec is unchanged", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-
-		got := fetch(ns, "qwen")
-		got.Annotations["swiss.modelsphere.dev/note"] = "later"
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.gets).To(Equal(1))
-	})
-
-	It("upgrades and bounds history", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 2)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		first := fetch(ns, "qwen").Status.History[0].ControllerRevision
-
-		for _, body := range []string{`{"replicaCount":2}`, `{"replicaCount":3}`} {
-			got := fetch(ns, "qwen")
-			got.Spec.Layers = []servingv1alpha1.Layer{layer(body)}
-			Expect(k8sClient.Update(ctx, got)).To(Succeed())
-			_, err = doReconcile(r, ns, "qwen")
-			Expect(err).NotTo(HaveOccurred())
-		}
-		Expect(h.upgrades).To(Equal(2))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.History).To(HaveLen(2))
-		Expect(got.Status.History[0].Revision).To(Equal(3))
-		Expect(got.Status.History[1].Revision).To(Equal(2))
-		Expect(got.Status.Helm.Revision).To(Equal(3))
-		err = k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: first}, &appsv1.ControllerRevision{})
-		Expect(apierrors.IsNotFound(err)).To(BeTrue())
-		Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: got.Status.History[0].ControllerRevision}, &appsv1.ControllerRevision{})).To(Succeed())
-	})
-
-	It("keeps a resolved version while the range allows it", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(fetch(ns, "qwen").Status.Chart.Version).To(Equal(ver086))
-		Expect(l.calls).To(Equal(1))
-
-		l.versions = []string{ver086, "0.9.0"}
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(fetch(ns, "qwen").Status.Chart.Version).To(Equal(ver086))
-		Expect(l.calls).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-
-		got := fetch(ns, "qwen")
-		got.Spec.Chart.Version = ">=0.9.0"
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(fetch(ns, "qwen").Status.Chart.Version).To(Equal("0.9.0"))
-		Expect(l.calls).To(Equal(2))
-		Expect(h.upgrades).To(Equal(1))
-	})
-
-	It("re-resolves once per new re-resolve value", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		reResolve := func(v string) {
-			got := fetch(ns, "qwen")
-			got.Annotations[servingv1alpha1.AnnotationReResolve] = v
-			Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		}
-
-		// Nothing newer: honored without an upgrade.
-		reResolve("1")
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Chart.Version).To(Equal(ver086))
-		Expect(got.Status.Chart.ReResolve).To(Equal("1"))
-		Expect(l.calls).To(Equal(2))
-		Expect(h.upgrades).To(Equal(0))
-
-		// A newer chart under an honored value does not move the release.
-		l.versions = []string{ver086, "0.9.0"}
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(fetch(ns, "qwen").Status.Chart.Version).To(Equal(ver086))
-		Expect(l.calls).To(Equal(2))
-
-		// A failed upgrade leaves the value unhonored, so the retry re-resolves.
-		h.upgradeErr = errors.New("boom")
-		reResolve("2")
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).To(HaveOccurred())
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Chart.Version).To(Equal(ver086))
-		Expect(got.Status.Chart.ReResolve).To(Equal("1"))
-
-		h.upgradeErr = nil
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Chart.Version).To(Equal("0.9.0"))
-		Expect(got.Status.Chart.ReResolve).To(Equal("2"))
-		Expect(got.Status.History[0].ReResolved).To(BeTrue())
-		Expect(h.upgrades).To(Equal(2))
-
-		l.versions = []string{ver086, "0.9.0", "0.9.1"}
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(fetch(ns, "qwen").Status.Chart.Version).To(Equal("0.9.0"))
-		Expect(h.upgrades).To(Equal(2))
-	})
-
-	It("adopts a matching release and refuses drift", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		h.put(ns, "qwen", &helm.Release{
-			Revision: 7, Status: deployed, ChartName: chartName, ChartVersion: ver086,
-			Config: map[string]any{replicasKey: float64(1)},
-		})
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted).Status).To(Equal(metav1.ConditionTrue))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted).Reason).To(Equal(servingv1alpha1.ReasonAdopted))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonAdopted))
-		Expect(got.Status.History[0].Revision).To(Equal(7))
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-
-		ns2 := takeNS()
-		h.put(ns2, "qwen", &helm.Release{
-			Revision: 3, Status: deployed, ChartName: "other", ChartVersion: "0.1.0",
-			Config: map[string]any{replicasKey: float64(9)},
-		})
-		Expect(k8sClient.Create(ctx, newService(ns2, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err = doReconcile(r, ns2, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		_, err = doReconcile(r, ns2, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		drifted := fetch(ns2, "qwen")
-		Expect(drifted.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(drifted.Status.ObservedGeneration).To(Equal(drifted.Generation))
-		Expect(condition(drifted, servingv1alpha1.ConditionAdopted).Status).To(Equal(metav1.ConditionFalse))
-		Expect(condition(drifted, servingv1alpha1.ConditionAdopted).Reason).To(Equal(servingv1alpha1.ReasonDrift))
-		Expect(drifted.Status.Message).To(ContainSubstring("chart"))
-		Expect(drifted.Status.Message).To(ContainSubstring("values"))
-		Expect(drifted.Status.ReleaseUID).To(BeEmpty())
-	})
-
-	It("orphans a delete and uninstalls otherwise", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "keep", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "keep")
-		Expect(err).NotTo(HaveOccurred())
-		got := fetch(ns, "keep")
-		got.Annotations[servingv1alpha1.AnnotationDeletionPolicy] = servingv1alpha1.DeletionPolicyOrphan
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		Expect(k8sClient.Delete(ctx, got)).To(Succeed())
-		gets := h.gets
-		_, err = doReconcile(r, ns, "keep")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.gets).To(Equal(gets))
-		err = k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: "keep"}, &servingv1alpha1.LLMService{})
-		Expect(apierrors.IsNotFound(err)).To(BeTrue())
-
-		Expect(k8sClient.Create(ctx, newService(ns, "drop", ver086))).To(Succeed())
-		_, err = doReconcile(r, ns, "drop")
-		Expect(err).NotTo(HaveOccurred())
-		drop := fetch(ns, "drop")
-		Expect(k8sClient.Delete(ctx, drop)).To(Succeed())
-		_, err = doReconcile(r, ns, "drop")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(1))
-		err = k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: "drop"}, &servingv1alpha1.LLMService{})
-		Expect(apierrors.IsNotFound(err)).To(BeTrue())
-	})
-
-	It("drops the finalizer when the release is already absent", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		obj := newService(ns, "gone", ver086)
-		obj.Finalizers = []string{servingv1alpha1.Finalizer}
-		Expect(k8sClient.Create(ctx, obj)).To(Succeed())
-		Expect(k8sClient.Delete(ctx, obj)).To(Succeed())
-		_, err := doReconcile(newReconciler(h, l, 10), ns, "gone")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.gets).To(Equal(1))
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		err = k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: "gone"}, &servingv1alpha1.LLMService{})
-		Expect(apierrors.IsNotFound(err)).To(BeTrue())
-	})
-
-	It("keeps a release it never installed or adopted on delete", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-
-		// A foreign release with other values: adoption is refused as drift.
-		foreign := &helm.Release{
-			Revision: 3, Status: deployed, ChartName: chartName, ChartVersion: ver086,
-			Config: map[string]any{replicasKey: float64(9)},
-		}
-		h.put(ns, "qwen", foreign)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(condition(fetch(ns, "qwen"), servingv1alpha1.ConditionAdopted).Reason).To(Equal(servingv1alpha1.ReasonDrift))
-
-		Expect(k8sClient.Delete(ctx, fetch(ns, "qwen"))).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.releases).To(HaveKey(h.slot(ns, "qwen")))
-		err = k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: "qwen"}, &servingv1alpha1.LLMService{})
-		Expect(apierrors.IsNotFound(err)).To(BeTrue())
-
-		// status.releaseUID claimed before a first install that never stored a
-		// release, and a release since installed by hand: not this object's.
-		obj := newService(ns, "claimed", ver086)
-		obj.Finalizers = []string{servingv1alpha1.Finalizer}
-		Expect(k8sClient.Create(ctx, obj)).To(Succeed())
-		obj = fetch(ns, "claimed")
-		obj.Status.ReleaseUID = string(obj.UID)
-		Expect(k8sClient.Status().Update(ctx, obj)).To(Succeed())
-		h.put(ns, "claimed", &helm.Release{Revision: 1, Status: deployed, ChartName: chartName, ChartVersion: ver086})
-		Expect(k8sClient.Delete(ctx, obj)).To(Succeed())
-		_, err = doReconcile(r, ns, "claimed")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.releases).To(HaveKey(h.slot(ns, "claimed")))
-
-		// A release labelled for another LLMService of the same name.
-		obj = newService(ns, "other", ver086)
-		obj.Finalizers = []string{servingv1alpha1.Finalizer}
-		Expect(k8sClient.Create(ctx, obj)).To(Succeed())
-		h.put(ns, "other", &helm.Release{
-			Revision: 2, Status: deployed, ChartName: chartName, ChartVersion: ver086,
-			Labels: map[string]string{llmserviceLabel: "other", llmserviceUIDLabel: "some-other-uid"},
-		})
-		Expect(k8sClient.Delete(ctx, fetch(ns, "other"))).To(Succeed())
-		_, err = doReconcile(r, ns, "other")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(0))
-	})
-
-	It("uninstalls an adopted release on delete", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		h.put(ns, "qwen", &helm.Release{
-			Revision: 7, Status: deployed, ChartName: chartName, ChartVersion: ver086,
-			Config: map[string]any{replicasKey: float64(1)},
-		})
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(condition(fetch(ns, "qwen"), servingv1alpha1.ConditionAdopted).Status).To(Equal(metav1.ConditionTrue))
-		Expect(h.releases[h.slot(ns, "qwen")].Labels).NotTo(HaveKey(llmserviceUIDLabel))
-
-		Expect(k8sClient.Delete(ctx, fetch(ns, "qwen"))).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(1))
-	})
-
-	It("refuses a chart repository it was not started to install from", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		r.AllowedChartRepos = []string{"oci://ghcr.io/other"}
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.gets + h.installs + h.upgrades).To(Equal(0))
-		Expect(l.calls).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonChartRepoNotAllowed))
-		Expect(got.Status.Message).To(ContainSubstring("oci://ghcr.io/modelsphere/charts"))
-
-		r.AllowedChartRepos = []string{modelsphereOCI}
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(1))
-		Expect(fetch(ns, "qwen").Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-	})
-
-	It("leaves status unchanged while suspended", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		before := fetch(ns, "qwen")
-		before.Spec.Suspend = true
-		before.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":9}`)}
-		Expect(k8sClient.Update(ctx, before)).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(got.Status.AppliedHash).To(Equal(before.Status.AppliedHash))
-		Expect(got.Status.Helm.Revision).To(Equal(1))
-		Expect(got.Status.ObservedGeneration).To(Equal(before.Status.ObservedGeneration))
-	})
-
-	It("honors force-conflicts only for the current generation", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		obj := newService(ns, "qwen", ver086)
-		obj.Annotations[servingv1alpha1.AnnotationForceConflicts] = "1"
-		Expect(k8sClient.Create(ctx, obj)).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.lastForce).To(BeTrue())
-		Expect(fetch(ns, "qwen").Status.History[0].ForceConflicts).To(BeTrue())
-
-		got := fetch(ns, "qwen")
-		got.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":2}`)}
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		Expect(fetch(ns, "qwen").Generation).To(Equal(int64(2)))
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.lastForce).To(BeFalse())
-		Expect(fetch(ns, "qwen").Status.History[0].ForceConflicts).To(BeFalse())
-	})
-
-	It("records a helm failure", func() {
-		ns := takeNS()
-		h := &fakeHelm{installErr: errors.New("boom")}
-		l := &versionList{versions: []string{ver086}}
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(newReconciler(h, l, 10), ns, "qwen")
-		Expect(err).To(MatchError(ContainSubstring("boom")))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(got.Status.Message).To(ContainSubstring("boom"))
-		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonApplyFailed))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Status).To(Equal(metav1.ConditionFalse))
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-	})
-
-	It("installs again after a failed install and a transient chart list error", func() {
-		ns := takeNS()
-		h := &fakeHelm{installErr: errors.New("boom"), keepFailed: true}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).To(MatchError(ContainSubstring("boom")))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.AppliedHash).To(BeEmpty())
-		Expect(got.Status.History).To(BeEmpty())
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-		Expect(h.releases[h.slot(ns, "qwen")].Labels).To(HaveKeyWithValue(llmserviceUIDLabel, string(got.UID)))
-		Expect(h.releases[h.slot(ns, "qwen")].Status).To(Equal(statusFailed))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonApplyFailed))
-
-		l.err = errors.New("registry down")
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).To(MatchError(ContainSubstring("registry down")))
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonChartResolveFailed))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted)).To(BeNil())
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-
-		l.err = nil
-		h.installErr = nil
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(2))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(1))
-		Expect(h.rollbacks).To(Equal(0))
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonInstalled))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted)).To(BeNil())
-		Expect(got.Status.Message).NotTo(ContainSubstring("Drift"))
-	})
-
-	It("requeues a helm get error instead of drifting a failed release", func() {
-		ns := takeNS()
-		h := &fakeHelm{installErr: errors.New("boom"), keepFailed: true}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).To(MatchError(ContainSubstring("boom")))
-
-		h.installErr = nil
-		h.getErr = errors.New("apiserver timeout")
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).To(MatchError(ContainSubstring("apiserver timeout")))
-		got := fetch(ns, "qwen")
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonApplyFailed))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted)).To(BeNil())
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.installs).To(Equal(1))
-
-		h.getErr = nil
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(1))
-		Expect(h.installs).To(Equal(2))
-		Expect(h.upgrades).To(Equal(0))
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonInstalled))
-	})
-
-	It("upgrades an owned failed release that still has a deployed revision", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		got := fetch(ns, "qwen")
-		slot := h.slot(ns, "qwen")
-		failed := ownedRel(got, 2, statusFailed, time.Now())
-		h.histories = map[string][]helm.Release{
-			slot: {*h.releases[slot], *failed},
-		}
-		h.put(ns, "qwen", failed)
-		got.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":2}`)}
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.upgrades).To(Equal(1))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.installs).To(Equal(1))
-		Expect(fetch(ns, "qwen").Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(fetch(ns, "qwen"), servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonUpgraded))
-	})
-
-	It("does not adopt a release that is not deployed", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		h.put(ns, "qwen", &helm.Release{
-			Revision: 1, Status: statusFailed, ChartName: chartName, ChartVersion: ver086,
-			Config: map[string]any{replicasKey: float64(1)},
-		})
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(got.Status.Message).To(Equal("release is failed, not deployed"))
-		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted).Status).To(Equal(metav1.ConditionFalse))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted).Reason).To(Equal(servingv1alpha1.ReasonDrift))
-		Expect(got.Status.ReleaseUID).To(BeEmpty())
-	})
-
-	It("still drifts a foreign failed release when Applied is ApplyFailed", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		got := fetch(ns, "qwen")
-		got.Status.Phase = servingv1alpha1.PhaseFailed
-		meta.SetStatusCondition(&got.Status.Conditions, metav1.Condition{
-			Type:               servingv1alpha1.ConditionApplied,
-			Status:             metav1.ConditionFalse,
-			Reason:             servingv1alpha1.ReasonApplyFailed,
-			Message:            "boom",
-			ObservedGeneration: got.Generation,
-		})
-		Expect(k8sClient.Status().Update(ctx, got)).To(Succeed())
-		h.put(ns, "qwen", &helm.Release{
-			Revision: 1, Status: statusFailed, ChartName: chartName, ChartVersion: ver086,
-			Config: map[string]any{replicasKey: float64(1)},
-		})
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		got = fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted).Reason).To(Equal(servingv1alpha1.ReasonDrift))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonDrift))
-	})
-
-	It("keeps observedGeneration while install is in progress", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		prev := fetch(ns, "qwen").Status.ObservedGeneration
-		Expect(prev).NotTo(BeZero())
-
-		delete(h.releases, h.slot(ns, "qwen"))
-		got := fetch(ns, "qwen")
-		got.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":4}`)}
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-
-		h.hold = make(chan struct{})
-		h.entered = make(chan struct{}, 1)
-		done := make(chan error, 1)
-		go func() {
-			_, recErr := doReconcile(r, ns, "qwen")
-			done <- recErr
-		}()
-		Eventually(h.entered).WithTimeout(5 * time.Second).Should(Receive())
-		mid := fetch(ns, "qwen")
-		Expect(mid.Status.Phase).To(Equal(servingv1alpha1.PhaseApplying))
-		Expect(mid.Status.ObservedGeneration).To(Equal(prev))
-		close(h.hold)
-		Eventually(done).WithTimeout(5 * time.Second).Should(Receive(BeNil()))
-		Expect(h.installs).To(Equal(2))
-		Expect(h.upgrades).To(Equal(0))
-		doneObj := fetch(ns, "qwen")
-		Expect(doneObj.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(doneObj.Status.ObservedGeneration).To(Equal(doneObj.Generation))
-	})
-
-	It("records a rollback to earlier values as a ControllerRevision of its own", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-
-		got := fetch(ns, "qwen")
-		got.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":2}`)}
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-
-		// What swiss sends for a rollback: the earlier values, with its own
-		// annotations, which the snapshot records -- the data a reused
-		// ControllerRevision name would have had to change.
-		got = fetch(ns, "qwen")
-		got.Spec.Layers = []servingv1alpha1.Layer{layer(`{"replicaCount":1}`)}
-		got.Annotations["swiss.modelsphere.dev/action"] = "rollback"
-		got.Annotations["swiss.modelsphere.dev/plan-hash"] = "p3"
-		Expect(k8sClient.Update(ctx, got)).To(Succeed())
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-
-		got = fetch(ns, "qwen")
-		Expect(got.Status.History).To(HaveLen(3))
-		Expect(got.Status.AppliedHash).To(Equal(got.Status.History[0].Hash))
-		Expect(got.Status.History[0].Hash).To(Equal(got.Status.History[2].Hash))
-		Expect(got.Status.History[0].Action).To(Equal("rollback"))
-		names := map[string]bool{}
-		for _, e := range got.Status.History {
-			names[e.ControllerRevision] = true
-		}
-		Expect(names).To(HaveLen(3))
-		Expect(got.Status.History[0].Revision).To(Equal(3))
-
-		snapshot := func(name string) revisionSnapshot {
-			var cr appsv1.ControllerRevision
-			Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: name}, &cr)).To(Succeed())
-			var snap revisionSnapshot
-			Expect(json.Unmarshal(cr.Data.Raw, &snap)).To(Succeed())
-			return snap
-		}
-		Expect(snapshot(got.Status.History[0].ControllerRevision).Annotations).To(HaveKeyWithValue("swiss.modelsphere.dev/action", "rollback"))
-		Expect(snapshot(got.Status.History[2].ControllerRevision).Annotations).To(HaveKeyWithValue("swiss.modelsphere.dev/action", "install"))
-
-		// Applied: the next reconcile has nothing to do.
-		upgrades := h.upgrades
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.upgrades).To(Equal(upgrades))
-	})
-
-	It("replaces its own stale snapshot under a reused revision name", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-
-		// What an uninstall leaves when helm's numbering restarts: a snapshot
-		// of this object's, under the name the next install will use.
-		values, err := mergeLayers(obj.Spec.Layers)
-		Expect(err).NotTo(HaveOccurred())
-		hash, err := appliedHash(chartName, ver086, values)
-		Expect(err).NotTo(HaveOccurred())
-		stale := &appsv1.ControllerRevision{
-			ObjectMeta: metav1.ObjectMeta{Name: controllerRevisionName("qwen", hash, 1), Namespace: ns},
-			Data:       runtime.RawExtension{Raw: []byte(`{"annotations":{"swiss.modelsphere.dev/action":"old"}}`)},
-			Revision:   1,
-		}
-		Expect(controllerutil.SetControllerReference(obj, stale, k8sClient.Scheme())).To(Succeed())
-		Expect(k8sClient.Create(ctx, stale)).To(Succeed())
-
-		_, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(got.Status.History[0].ControllerRevision).To(Equal(stale.Name))
-		var cr appsv1.ControllerRevision
-		Expect(k8sClient.Get(ctx, types.NamespacedName{Namespace: ns, Name: stale.Name}, &cr)).To(Succeed())
-		var snap revisionSnapshot
-		Expect(json.Unmarshal(cr.Data.Raw, &snap)).To(Succeed())
-		Expect(snap.Annotations).To(HaveKeyWithValue("swiss.modelsphere.dev/action", "install"))
-	})
-
-	It("fails when the chart version cannot be resolved", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{err: errors.New("registry down")}
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ">=0.8.0"))).To(Succeed())
-		_, err := doReconcile(newReconciler(h, l, 10), ns, "qwen")
-		Expect(err).To(HaveOccurred())
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(got.Status.Message).To(ContainSubstring("registry down"))
-		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonChartResolveFailed))
-		Expect(h.installs).To(Equal(0))
-	})
-
-	It("recovers a stale owned pending-install and installs", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		rec := events.NewFakeRecorder(8)
-		r := newReconciler(h, l, 10)
-		r.Recorder = rec
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		h.put(ns, "qwen", ownedRel(obj, 1, statusPendingInstall, time.Now().Add(-time.Hour)))
-
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.uninstalls).To(Equal(1))
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.rollbacks).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonInstalled))
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-		Expect(rec.Events).To(Receive(ContainSubstring(servingv1alpha1.ReasonRecoveredPendingRelease)))
-	})
-
-	It("waits out a fresh owned pending-upgrade", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		r.PendingTimeout = 2 * time.Minute
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		h.put(ns, "qwen", ownedRel(obj, 2, statusPendingUpgrade, time.Now()))
-		h.histories = map[string][]helm.Release{
-			h.slot(ns, "qwen"): {ownedRelVal(obj, 1, deployed, time.Now().Add(-time.Hour)), *h.releases[h.slot(ns, "qwen")]},
-		}
-
-		result, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(result.RequeueAfter).To(BeNumerically(">=", minPendingRequeue))
-		Expect(result.RequeueAfter).To(BeNumerically(">", time.Minute))
-		Expect(result.RequeueAfter).To(BeNumerically("<=", 2*time.Minute))
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.rollbacks).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplying))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonOperationInProgress))
-		Expect(got.Status.Message).To(ContainSubstring(statusPendingUpgrade))
-
-		result, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(result.RequeueAfter).To(BeNumerically(">=", minPendingRequeue))
-		Expect(h.rollbacks).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-	})
-
-	It("clamps a short pending wait to 15s", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		r.PendingTimeout = 10 * time.Second
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		h.put(ns, "qwen", ownedRel(obj, 1, statusPendingRollback, time.Now().Add(-8*time.Second)))
-
-		result, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(result.RequeueAfter).To(Equal(minPendingRequeue))
-		Expect(h.rollbacks).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(condition(fetch(ns, "qwen"), servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonOperationInProgress))
-	})
-
-	It("rolls back a stale owned pending-upgrade and upgrades", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		rec := events.NewFakeRecorder(8)
-		r := newReconciler(h, l, 10)
-		r.Recorder = rec
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		deployedRel := ownedRelVal(obj, 4, deployed, time.Now().Add(-2*time.Hour))
-		pending := ownedRel(obj, 5, statusPendingUpgrade, time.Now().Add(-time.Hour))
-		h.put(ns, "qwen", pending)
-		h.histories = map[string][]helm.Release{
-			h.slot(ns, "qwen"): {
-				ownedRelVal(obj, 2, deployed, time.Now().Add(-3*time.Hour)),
-				deployedRel,
-				*pending,
-			},
-		}
-
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.rollbacks).To(Equal(1))
-		Expect(h.lastRollback).To(Equal(4))
-		Expect(h.upgrades).To(Equal(1))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.installs).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonUpgraded))
-		Expect(rec.Events).To(Receive(ContainSubstring(servingv1alpha1.ReasonRecoveredPendingRelease)))
-	})
-
-	It("uninstalls a stale owned pending-rollback that has no deployed revision", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		pending := ownedRel(obj, 2, statusPendingRollback, time.Now().Add(-time.Hour))
-		h.put(ns, "qwen", pending)
-		h.histories = map[string][]helm.Release{
-			h.slot(ns, "qwen"): {ownedRelVal(obj, 1, statusFailed, time.Now().Add(-2*time.Hour)), *pending},
-		}
-
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.rollbacks).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(1))
-		Expect(h.installs).To(Equal(1))
-		Expect(h.upgrades).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonInstalled))
-	})
-
-	It("does not write a pending release it does not own", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		h.put(ns, "qwen", &helm.Release{
-			Revision: 3, Status: statusPendingUpgrade, ChartName: chartName, ChartVersion: ver086,
-			Updated: time.Now().Add(-time.Hour),
+	for _, f := range cases.MustLoad[serviceCase]("testdata/controller") {
+		tc := f.Case
+		It(tc.Description, func() {
+			runServiceCase(tc)
 		})
-
-		result, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(result.RequeueAfter).To(Equal(time.Minute))
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.rollbacks).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseFailed))
-		Expect(got.Status.Message).To(Equal("release is pending-upgrade; another client is operating on it"))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Status).To(Equal(metav1.ConditionFalse))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonPendingRelease))
-		Expect(got.Status.ObservedGeneration).To(Equal(got.Generation))
-
-		result, err = doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(result.RequeueAfter).To(Equal(time.Minute))
-		Expect(h.installs).To(Equal(0))
-		Expect(h.upgrades).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		Expect(h.rollbacks).To(Equal(0))
-	})
-
-	It("upgrades an owned deployed release whose status was lost", func() {
-		ns := takeNS()
-		h := &fakeHelm{}
-		l := &versionList{versions: []string{ver086}}
-		r := newReconciler(h, l, 10)
-		Expect(k8sClient.Create(ctx, newService(ns, "qwen", ver086))).To(Succeed())
-		obj := fetch(ns, "qwen")
-		h.put(ns, "qwen", ownedRel(obj, 4, deployed, time.Now().Add(-time.Hour)))
-
-		_, err := doReconcile(r, ns, "qwen")
-		Expect(err).NotTo(HaveOccurred())
-		Expect(h.upgrades).To(Equal(1))
-		Expect(h.installs).To(Equal(0))
-		Expect(h.uninstalls).To(Equal(0))
-		got := fetch(ns, "qwen")
-		Expect(got.Status.Phase).To(Equal(servingv1alpha1.PhaseApplied))
-		Expect(condition(got, servingv1alpha1.ConditionApplied).Reason).To(Equal(servingv1alpha1.ReasonUpgraded))
-		Expect(condition(got, servingv1alpha1.ConditionAdopted)).To(BeNil())
-		Expect(got.Status.ReleaseUID).To(Equal(string(got.UID)))
-		Expect(got.Status.Helm.Revision).To(Equal(5))
-	})
+	}
 })
 
 func ownedRel(obj *servingv1alpha1.LLMService, rev int, status string, updated time.Time) *helm.Release {
diff --git a/internal/controller/serving/predicate_test.go b/internal/controller/serving/predicate_test.go
index 673ded5..7c558e2 100644
--- a/internal/controller/serving/predicate_test.go
+++ b/internal/controller/serving/predicate_test.go
@@ -7,39 +7,55 @@ import (
 	"sigs.k8s.io/controller-runtime/pkg/event"
 
 	servingv1alpha1 "github.com/modelsphere/llm-operator/api/serving/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-func TestLLMServicePredicate(t *testing.T) {
-	p := llmServicePredicate{}
-	base := &servingv1alpha1.LLMService{}
-	base.Generation = 1
-	base.Annotations = map[string]string{"a": "1"}
-
-	same := base.DeepCopy()
-	same.Status.Phase = servingv1alpha1.PhaseApplied
-	if p.Update(event.UpdateEvent{ObjectOld: base, ObjectNew: same}) {
-		t.Fatal("status-only update reconciled")
-	}
-
-	gen := base.DeepCopy()
-	gen.Generation = 2
-	if !p.Update(event.UpdateEvent{ObjectOld: base, ObjectNew: gen}) {
-		t.Fatal("generation change ignored")
-	}
-
-	ann := base.DeepCopy()
-	ann.Annotations = map[string]string{"a": "2"}
-	if !p.Update(event.UpdateEvent{ObjectOld: base, ObjectNew: ann}) {
-		t.Fatal("annotation change ignored")
-	}
+// predicateCase is one llmServicePredicate input. The base object is generation
+// 1 with annotation a=1; the case describes what changes on the new object.
+type predicateCase struct {
+	Name        string            `json:"name,omitempty"`
+	Description string            `json:"description,omitempty"`
+	Op          string            `json:"op"`
+	Phase       string            `json:"phase,omitempty"`
+	Generation  int64             `json:"generation,omitempty"`
+	Annotations map[string]string `json:"annotations,omitempty"`
+	Deleting    bool              `json:"deleting,omitempty"`
+	Want        bool              `json:"want"`
+}
 
-	now := metav1.Now()
-	del := base.DeepCopy()
-	del.DeletionTimestamp = &now
-	if !p.Update(event.UpdateEvent{ObjectOld: base, ObjectNew: del}) {
-		t.Fatal("deletion ignored")
-	}
-	if !p.Create(event.CreateEvent{Object: base}) || !p.Delete(event.DeleteEvent{Object: base}) {
-		t.Fatal("create or delete ignored")
-	}
+func TestLLMServicePredicate(t *testing.T) {
+	cases.Run(t, "testdata/predicate", func(t *testing.T, tc predicateCase) {
+		p := llmServicePredicate{}
+		base := &servingv1alpha1.LLMService{}
+		base.Generation = 1
+		base.Annotations = map[string]string{"a": "1"}
+		var got bool
+		switch tc.Op {
+		case "update":
+			next := base.DeepCopy()
+			if tc.Phase != "" {
+				next.Status.Phase = servingv1alpha1.Phase(tc.Phase)
+			}
+			if tc.Generation != 0 {
+				next.Generation = tc.Generation
+			}
+			if tc.Annotations != nil {
+				next.Annotations = tc.Annotations
+			}
+			if tc.Deleting {
+				now := metav1.Now()
+				next.DeletionTimestamp = &now
+			}
+			got = p.Update(event.UpdateEvent{ObjectOld: base, ObjectNew: next})
+		case "create":
+			got = p.Create(event.CreateEvent{Object: base})
+		case "delete":
+			got = p.Delete(event.DeleteEvent{Object: base})
+		default:
+			t.Fatalf("%s: unknown op %q", tc.Name, tc.Op)
+		}
+		if got != tc.Want {
+			t.Fatalf("%s: reconcile = %v, want %v", tc.Description, got, tc.Want)
+		}
+	})
 }
diff --git a/internal/controller/serving/repo_test.go b/internal/controller/serving/repo_test.go
index 978354f..63567c1 100644
--- a/internal/controller/serving/repo_test.go
+++ b/internal/controller/serving/repo_test.go
@@ -1,26 +1,30 @@
 package serving
 
-import "testing"
+import (
+	"testing"
 
-const modelsphereOCI = "oci://ghcr.io/modelsphere"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+)
+
+// repoCase is one chartRepoAllowed input. NilAllowlist passes a nil allowlist,
+// which permits every repository. Otherwise Allowed is the allowlist.
+type repoCase struct {
+	Name         string   `json:"name,omitempty"`
+	Description  string   `json:"description,omitempty"`
+	Allowed      []string `json:"allowed,omitempty"`
+	NilAllowlist bool     `json:"nilAllowlist,omitempty"`
+	Repo         string   `json:"repo"`
+	Want         bool     `json:"want"`
+}
 
 func TestChartRepoAllowed(t *testing.T) {
-	allowed := []string{modelsphereOCI, "https://modelsphere.github.io/helm-charts/"}
-	for repo, want := range map[string]bool{
-		modelsphereOCI:                              true,
-		"oci://ghcr.io/modelsphere/charts":          true,
-		"oci://ghcr.io/modelsphere/charts/":         true,
-		"https://modelsphere.github.io/helm-charts": true,
-		"oci://ghcr.io/modelsphere-x/charts":        false,
-		"oci://ghcr.io/other/charts":                false,
-		"https://modelsphere.github.io":             false,
-		"":                                          false,
-	} {
-		if got := chartRepoAllowed(allowed, repo); got != want {
-			t.Errorf("%q: got %v, want %v", repo, got, want)
+	cases.Run(t, "testdata/repo", func(t *testing.T, tc repoCase) {
+		var allowed []string
+		if !tc.NilAllowlist {
+			allowed = tc.Allowed
+		}
+		if got := chartRepoAllowed(allowed, tc.Repo); got != tc.Want {
+			t.Errorf("%s: %q: got %v, want %v", tc.Description, tc.Repo, got, tc.Want)
 		}
-	}
-	if !chartRepoAllowed(nil, "https://anything.example.com/charts") {
-		t.Error("no allowlist allows any repository")
-	}
+	})
 }
diff --git a/internal/controller/serving/testdata/controller/adopts-a-matching-release-and-refuses-drift.yaml b/internal/controller/serving/testdata/controller/adopts-a-matching-release-and-refuses-drift.yaml
new file mode 100644
index 0000000..a4e070c
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/adopts-a-matching-release-and-refuses-drift.yaml
@@ -0,0 +1,57 @@
+name: adopts-a-matching-release-and-refuses-drift
+description: adopts a matching release and refuses drift
+steps:
+  - action: putRelease
+    release:
+      revision: 7
+      status: deployed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      config:
+        replicaCount: 1
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect:
+      installs: 0
+      upgrades: 0
+      phase: Applied
+      conditions:
+        - type: Adopted
+          status: "True"
+          reason: Adopted
+        - type: Applied
+          reason: Adopted
+      history:
+        entries:
+          - index: 0
+            revision: 7
+      releaseUID: self
+  - action: putRelease
+    freshNamespace: true
+    release:
+      revision: 3
+      status: deployed
+      chartName: other
+      chartVersion: "0.1.0"
+      config:
+        replicaCount: 9
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect: {}
+  - action: reconcile
+    expect:
+      installs: 0
+      upgrades: 0
+      phase: Failed
+      observedGeneration: current
+      conditions:
+        - type: Adopted
+          status: "False"
+          reason: Drift
+      message:
+        contains:
+          - chart
+          - values
+      releaseUID: empty
diff --git a/internal/controller/serving/testdata/controller/clamps-a-short-pending-wait-to-15s.yaml b/internal/controller/serving/testdata/controller/clamps-a-short-pending-wait-to-15s.yaml
new file mode 100644
index 0000000..88615e6
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/clamps-a-short-pending-wait-to-15s.yaml
@@ -0,0 +1,21 @@
+name: clamps-a-short-pending-wait-to-15s
+description: clamps a short pending wait to 15s
+pendingTimeout: "10s"
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 1
+      status: pending-rollback
+      updated: "-8s"
+  - action: reconcile
+    expect:
+      requeue:
+        exact: "15s"
+      rollbacks: 0
+      uninstalls: 0
+      conditions:
+        - type: Applied
+          reason: OperationInProgress
diff --git a/internal/controller/serving/testdata/controller/does-not-adopt-a-release-that-is-not-deployed.yaml b/internal/controller/serving/testdata/controller/does-not-adopt-a-release-that-is-not-deployed.yaml
new file mode 100644
index 0000000..a2f447c
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/does-not-adopt-a-release-that-is-not-deployed.yaml
@@ -0,0 +1,28 @@
+name: does-not-adopt-a-release-that-is-not-deployed
+description: does not adopt a release that is not deployed
+steps:
+  - action: putRelease
+    release:
+      revision: 1
+      status: failed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      config:
+        replicaCount: 1
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: reconcile
+    expect:
+      installs: 0
+      upgrades: 0
+      phase: Failed
+      message:
+        equals: release is failed, not deployed
+      observedGeneration: current
+      conditions:
+        - type: Adopted
+          status: "False"
+          reason: Drift
+      releaseUID: empty
diff --git a/internal/controller/serving/testdata/controller/does-not-call-helm-when-the-spec-is-unchanged.yaml b/internal/controller/serving/testdata/controller/does-not-call-helm-when-the-spec-is-unchanged.yaml
new file mode 100644
index 0000000..07f0891
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/does-not-call-helm-when-the-spec-is-unchanged.yaml
@@ -0,0 +1,15 @@
+name: does-not-call-helm-when-the-spec-is-unchanged
+description: does not call helm when the spec is unchanged
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: update
+    annotations:
+      swiss.modelsphere.dev/note: later
+  - action: reconcile
+    expect:
+      installs: 1
+      upgrades: 0
+      gets: 1
diff --git a/internal/controller/serving/testdata/controller/does-not-write-a-pending-release-it-does-not-own.yaml b/internal/controller/serving/testdata/controller/does-not-write-a-pending-release-it-does-not-own.yaml
new file mode 100644
index 0000000..43e8d2d
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/does-not-write-a-pending-release-it-does-not-own.yaml
@@ -0,0 +1,36 @@
+name: does-not-write-a-pending-release-it-does-not-own
+description: does not write a pending release it does not own
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      revision: 3
+      status: pending-upgrade
+      chartName: sglang
+      chartVersion: "0.8.6"
+      updated: "-1h"
+  - action: reconcile
+    expect:
+      requeue:
+        exact: "1m"
+      installs: 0
+      upgrades: 0
+      uninstalls: 0
+      rollbacks: 0
+      phase: Failed
+      message:
+        equals: "release is pending-upgrade; another client is operating on it"
+      conditions:
+        - type: Applied
+          status: "False"
+          reason: PendingRelease
+      observedGeneration: current
+  - action: reconcile
+    expect:
+      requeue:
+        exact: "1m"
+      installs: 0
+      upgrades: 0
+      uninstalls: 0
+      rollbacks: 0
diff --git a/internal/controller/serving/testdata/controller/drops-the-finalizer-when-the-release-is-already-absent.yaml b/internal/controller/serving/testdata/controller/drops-the-finalizer-when-the-release-is-already-absent.yaml
new file mode 100644
index 0000000..6c46727
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/drops-the-finalizer-when-the-release-is-already-absent.yaml
@@ -0,0 +1,16 @@
+name: drops-the-finalizer-when-the-release-is-already-absent
+description: drops the finalizer when the release is already absent
+steps:
+  - action: create
+    name: gone
+    version: "0.8.6"
+    finalizers:
+      - serving.modelsphere.dev/finalizer
+  - action: delete
+  - action: reconcile
+    expect:
+      gets: 1
+      installs: 0
+      upgrades: 0
+      uninstalls: 0
+      notFound: true
diff --git a/internal/controller/serving/testdata/controller/fails-when-the-chart-version-cannot-be-resolved.yaml b/internal/controller/serving/testdata/controller/fails-when-the-chart-version-cannot-be-resolved.yaml
new file mode 100644
index 0000000..157a185
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/fails-when-the-chart-version-cannot-be-resolved.yaml
@@ -0,0 +1,19 @@
+name: fails-when-the-chart-version-cannot-be-resolved
+description: fails when the chart version cannot be resolved
+steps:
+  - action: setVersions
+    listErr: registry down
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect:
+      error: true
+      phase: Failed
+      message:
+        contains:
+          - registry down
+      observedGeneration: current
+      conditions:
+        - type: Applied
+          reason: ChartResolveFailed
+      installs: 0
diff --git a/internal/controller/serving/testdata/controller/honors-force-conflicts-only-for-the-current-generation.yaml b/internal/controller/serving/testdata/controller/honors-force-conflicts-only-for-the-current-generation.yaml
new file mode 100644
index 0000000..c75d686
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/honors-force-conflicts-only-for-the-current-generation.yaml
@@ -0,0 +1,27 @@
+name: honors-force-conflicts-only-for-the-current-generation
+description: honors force-conflicts only for the current generation
+steps:
+  - action: create
+    version: "0.8.6"
+    annotations:
+      serving.modelsphere.dev/force-conflicts: "1"
+  - action: reconcile
+    expect:
+      lastForce: true
+      history:
+        entries:
+          - index: 0
+            forceConflicts: true
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":2}'
+    expect:
+      generation: 2
+  - action: reconcile
+    expect:
+      lastForce: false
+      history:
+        entries:
+          - index: 0
+            forceConflicts: false
diff --git a/internal/controller/serving/testdata/controller/installs-again-after-a-failed-install-and-a-transient-chart-list-error.yaml b/internal/controller/serving/testdata/controller/installs-again-after-a-failed-install-and-a-transient-chart-list-error.yaml
new file mode 100644
index 0000000..77cc169
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/installs-again-after-a-failed-install-and-a-transient-chart-list-error.yaml
@@ -0,0 +1,57 @@
+name: installs-again-after-a-failed-install-and-a-transient-chart-list-error
+description: installs again after a failed install and a transient chart list error
+steps:
+  - action: setHelm
+    installErr: boom
+    keepFailed: true
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect:
+      errorContains: boom
+      appliedHash:
+        empty: true
+      history:
+        length: 0
+      releaseUID: self
+      release:
+        labels:
+          serving.modelsphere.dev/llmservice-uid: self
+        status: failed
+      conditions:
+        - type: Applied
+          reason: ApplyFailed
+  - action: setVersions
+    listErr: registry down
+  - action: reconcile
+    expect:
+      errorContains: registry down
+      phase: Failed
+      conditions:
+        - type: Applied
+          reason: ChartResolveFailed
+        - type: Adopted
+          absent: true
+      releaseUID: self
+      installs: 1
+      upgrades: 0
+      uninstalls: 0
+  - action: setVersions
+    listErr: ""
+  - action: setHelm
+    installErr: ""
+  - action: reconcile
+    expect:
+      installs: 2
+      upgrades: 0
+      uninstalls: 1
+      rollbacks: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Installed
+        - type: Adopted
+          absent: true
+      message:
+        notContains:
+          - Drift
diff --git a/internal/controller/serving/testdata/controller/installs-and-records-a-controllerrevision.yaml b/internal/controller/serving/testdata/controller/installs-and-records-a-controllerrevision.yaml
new file mode 100644
index 0000000..48dffcc
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/installs-and-records-a-controllerrevision.yaml
@@ -0,0 +1,44 @@
+name: installs-and-records-a-controllerrevision
+description: installs and records a ControllerRevision
+steps:
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect:
+      installs: 1
+      upgrades: 0
+      finalizer: present
+      phase: Applied
+      chartVersion: "0.8.6"
+      appliedHash:
+        prefix: "sha256:"
+      helmRevision: 1
+      observedGeneration: current
+      history:
+        length: 1
+        entries:
+          - index: 0
+            action: install
+            annotationsPresent:
+              - swiss.modelsphere.dev/note
+            annotationsAbsent:
+              - unrelated
+      conditions:
+        - type: Applied
+          status: "True"
+          reason: Installed
+      releaseUID: self
+      release:
+        labels:
+          serving.modelsphere.dev/llmservice-uid: self
+          serving.modelsphere.dev/llmservice: qwen
+      snapshots:
+        - index: 0
+          labels:
+            serving.modelsphere.dev/llmservice: qwen
+          controllerOwner: true
+          annotations:
+            swiss.modelsphere.dev/action: install
+          annotationsAbsent:
+            - unrelated
+          chartName: sglang
diff --git a/internal/controller/serving/testdata/controller/keeps-a-release-it-never-installed-or-adopted-on-delete.yaml b/internal/controller/serving/testdata/controller/keeps-a-release-it-never-installed-or-adopted-on-delete.yaml
new file mode 100644
index 0000000..4f761d9
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/keeps-a-release-it-never-installed-or-adopted-on-delete.yaml
@@ -0,0 +1,61 @@
+name: keeps-a-release-it-never-installed-or-adopted-on-delete
+description: keeps a release it never installed or adopted on delete
+steps:
+  - action: putRelease
+    release:
+      revision: 3
+      status: deployed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      config:
+        replicaCount: 9
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect:
+      conditions:
+        - type: Adopted
+          reason: Drift
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 0
+      release:
+        exists: true
+      notFound: true
+  - action: create
+    name: claimed
+    version: "0.8.6"
+    finalizers:
+      - serving.modelsphere.dev/finalizer
+  - action: setStatusReleaseUID
+  - action: putRelease
+    release:
+      revision: 1
+      status: deployed
+      chartName: sglang
+      chartVersion: "0.8.6"
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 0
+      release:
+        exists: true
+  - action: create
+    name: other
+    version: "0.8.6"
+    finalizers:
+      - serving.modelsphere.dev/finalizer
+  - action: putRelease
+    release:
+      revision: 2
+      status: deployed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      labels:
+        serving.modelsphere.dev/llmservice: other
+        serving.modelsphere.dev/llmservice-uid: some-other-uid
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 0
diff --git a/internal/controller/serving/testdata/controller/keeps-a-resolved-version-while-the-range-allows-it.yaml b/internal/controller/serving/testdata/controller/keeps-a-resolved-version-while-the-range-allows-it.yaml
new file mode 100644
index 0000000..17f928a
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/keeps-a-resolved-version-while-the-range-allows-it.yaml
@@ -0,0 +1,25 @@
+name: keeps-a-resolved-version-while-the-range-allows-it
+description: keeps a resolved version while the range allows it
+steps:
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect:
+      chartVersion: "0.8.6"
+      listCalls: 1
+  - action: setVersions
+    versions:
+      - "0.8.6"
+      - "0.9.0"
+  - action: reconcile
+    expect:
+      chartVersion: "0.8.6"
+      listCalls: 1
+      upgrades: 0
+  - action: update
+    chartVersion: ">=0.9.0"
+  - action: reconcile
+    expect:
+      chartVersion: "0.9.0"
+      listCalls: 2
+      upgrades: 1
diff --git a/internal/controller/serving/testdata/controller/keeps-observedgeneration-while-install-is-in-progress.yaml b/internal/controller/serving/testdata/controller/keeps-observedgeneration-while-install-is-in-progress.yaml
new file mode 100644
index 0000000..d588c1b
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/keeps-observedgeneration-while-install-is-in-progress.yaml
@@ -0,0 +1,26 @@
+name: keeps-observedgeneration-while-install-is-in-progress
+description: keeps observedGeneration while install is in progress
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: capture
+    capture:
+      - observedGeneration
+    expect:
+      observedGeneration: nonzero
+  - action: deleteRelease
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":4}'
+  - action: reconcileHeld
+    expect:
+      mid:
+        phase: Applying
+        observedGeneration: captured
+      installs: 2
+      upgrades: 0
+      phase: Applied
+      observedGeneration: current
diff --git a/internal/controller/serving/testdata/controller/leaves-status-unchanged-while-suspended.yaml b/internal/controller/serving/testdata/controller/leaves-status-unchanged-while-suspended.yaml
new file mode 100644
index 0000000..e3472a5
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/leaves-status-unchanged-while-suspended.yaml
@@ -0,0 +1,25 @@
+name: leaves-status-unchanged-while-suspended
+description: leaves status unchanged while suspended
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: capture
+    capture:
+      - appliedHash
+      - observedGeneration
+  - action: update
+    suspend: true
+    layers:
+      - name: site
+        values: '{"replicaCount":9}'
+  - action: reconcile
+    expect:
+      installs: 1
+      upgrades: 0
+      phase: Applied
+      appliedHash:
+        captured: true
+      helmRevision: 1
+      observedGeneration: captured
diff --git a/internal/controller/serving/testdata/controller/orphans-a-delete-and-uninstalls-otherwise.yaml b/internal/controller/serving/testdata/controller/orphans-a-delete-and-uninstalls-otherwise.yaml
new file mode 100644
index 0000000..104599b
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/orphans-a-delete-and-uninstalls-otherwise.yaml
@@ -0,0 +1,27 @@
+name: orphans-a-delete-and-uninstalls-otherwise
+description: orphans a delete and uninstalls otherwise
+steps:
+  - action: create
+    name: keep
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: update
+    annotations:
+      serving.modelsphere.dev/deletion-policy: Orphan
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 0
+      getsUnchanged: true
+      notFound: true
+  - action: create
+    name: drop
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 1
+      notFound: true
diff --git a/internal/controller/serving/testdata/controller/re-resolves-once-per-new-re-resolve-value.yaml b/internal/controller/serving/testdata/controller/re-resolves-once-per-new-re-resolve-value.yaml
new file mode 100644
index 0000000..d0eb4c1
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/re-resolves-once-per-new-re-resolve-value.yaml
@@ -0,0 +1,54 @@
+name: re-resolves-once-per-new-re-resolve-value
+description: re-resolves once per new re-resolve value
+steps:
+  - action: create
+    version: ">=0.8.0"
+  - action: reconcile
+    expect: {}
+  - action: update
+    annotations:
+      serving.modelsphere.dev/re-resolve: "1"
+  - action: reconcile
+    expect:
+      chartVersion: "0.8.6"
+      reResolve: "1"
+      listCalls: 2
+      upgrades: 0
+  - action: setVersions
+    versions:
+      - "0.8.6"
+      - "0.9.0"
+  - action: reconcile
+    expect:
+      chartVersion: "0.8.6"
+      listCalls: 2
+  - action: setHelm
+    upgradeErr: boom
+  - action: update
+    annotations:
+      serving.modelsphere.dev/re-resolve: "2"
+  - action: reconcile
+    expect:
+      errorContains: boom
+      chartVersion: "0.8.6"
+      reResolve: "1"
+  - action: setHelm
+    upgradeErr: ""
+  - action: reconcile
+    expect:
+      chartVersion: "0.9.0"
+      reResolve: "2"
+      history:
+        entries:
+          - index: 0
+            reResolved: true
+      upgrades: 2
+  - action: setVersions
+    versions:
+      - "0.8.6"
+      - "0.9.0"
+      - "0.9.1"
+  - action: reconcile
+    expect:
+      chartVersion: "0.9.0"
+      upgrades: 2
diff --git a/internal/controller/serving/testdata/controller/records-a-helm-failure.yaml b/internal/controller/serving/testdata/controller/records-a-helm-failure.yaml
new file mode 100644
index 0000000..b997750
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/records-a-helm-failure.yaml
@@ -0,0 +1,20 @@
+name: records-a-helm-failure
+description: records a helm failure
+steps:
+  - action: setHelm
+    installErr: boom
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect:
+      errorContains: boom
+      phase: Failed
+      message:
+        contains:
+          - boom
+      observedGeneration: current
+      conditions:
+        - type: Applied
+          reason: ApplyFailed
+          status: "False"
+      releaseUID: self
diff --git a/internal/controller/serving/testdata/controller/records-a-rollback-to-earlier-values-as-a-controllerrevision-of-its-own.yaml b/internal/controller/serving/testdata/controller/records-a-rollback-to-earlier-values-as-a-controllerrevision-of-its-own.yaml
new file mode 100644
index 0000000..e5ac402
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/records-a-rollback-to-earlier-values-as-a-controllerrevision-of-its-own.yaml
@@ -0,0 +1,41 @@
+name: records-a-rollback-to-earlier-values-as-a-controllerrevision-of-its-own
+description: records a rollback to earlier values as a ControllerRevision of its own
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":2}'
+  - action: reconcile
+    expect: {}
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":1}'
+    annotations:
+      swiss.modelsphere.dev/action: rollback
+      swiss.modelsphere.dev/plan-hash: p3
+  - action: reconcile
+    expect:
+      history:
+        length: 3
+        entries:
+          - index: 0
+            action: rollback
+            revision: 3
+            hashEqualsApplied: true
+            hashEqualsIndex: 2
+      distinctRevisions: 3
+      snapshots:
+        - index: 0
+          annotations:
+            swiss.modelsphere.dev/action: rollback
+        - index: 2
+          annotations:
+            swiss.modelsphere.dev/action: install
+  - action: reconcile
+    expect:
+      upgradesUnchanged: true
diff --git a/internal/controller/serving/testdata/controller/recovers-a-stale-owned-pending-install-and-installs.yaml b/internal/controller/serving/testdata/controller/recovers-a-stale-owned-pending-install-and-installs.yaml
new file mode 100644
index 0000000..487b71f
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/recovers-a-stale-owned-pending-install-and-installs.yaml
@@ -0,0 +1,23 @@
+name: recovers-a-stale-owned-pending-install-and-installs
+description: recovers a stale owned pending-install and installs
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 1
+      status: pending-install
+      updated: "-1h"
+  - action: reconcile
+    expect:
+      uninstalls: 1
+      installs: 1
+      upgrades: 0
+      rollbacks: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Installed
+      releaseUID: self
+      eventContains: RecoveredPendingRelease
diff --git a/internal/controller/serving/testdata/controller/refuses-a-chart-repository-it-was-not-started-to-install-from.yaml b/internal/controller/serving/testdata/controller/refuses-a-chart-repository-it-was-not-started-to-install-from.yaml
new file mode 100644
index 0000000..64075d2
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/refuses-a-chart-repository-it-was-not-started-to-install-from.yaml
@@ -0,0 +1,26 @@
+name: refuses-a-chart-repository-it-was-not-started-to-install-from
+description: refuses a chart repository it was not started to install from
+steps:
+  - action: setAllowed
+    repos:
+      - oci://ghcr.io/other
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect:
+      helmTouched: 0
+      listCalls: 0
+      phase: Failed
+      conditions:
+        - type: Applied
+          reason: ChartRepoNotAllowed
+      message:
+        contains:
+          - oci://ghcr.io/modelsphere/charts
+  - action: setAllowed
+    repos:
+      - oci://ghcr.io/modelsphere
+  - action: reconcile
+    expect:
+      installs: 1
+      phase: Applied
diff --git a/internal/controller/serving/testdata/controller/replaces-its-own-stale-snapshot-under-a-reused-revision-name.yaml b/internal/controller/serving/testdata/controller/replaces-its-own-stale-snapshot-under-a-reused-revision-name.yaml
new file mode 100644
index 0000000..652fd73
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/replaces-its-own-stale-snapshot-under-a-reused-revision-name.yaml
@@ -0,0 +1,14 @@
+name: replaces-its-own-stale-snapshot-under-a-reused-revision-name
+description: replaces its own stale snapshot under a reused revision name
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: plantStale
+  - action: reconcile
+    expect:
+      phase: Applied
+      plantedRevisionUsed: true
+      snapshots:
+        - index: 0
+          annotations:
+            swiss.modelsphere.dev/action: install
diff --git a/internal/controller/serving/testdata/controller/requeues-a-helm-get-error-instead-of-drifting-a-failed-release.yaml b/internal/controller/serving/testdata/controller/requeues-a-helm-get-error-instead-of-drifting-a-failed-release.yaml
new file mode 100644
index 0000000..0c469a0
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/requeues-a-helm-get-error-instead-of-drifting-a-failed-release.yaml
@@ -0,0 +1,35 @@
+name: requeues-a-helm-get-error-instead-of-drifting-a-failed-release
+description: requeues a helm get error instead of drifting a failed release
+steps:
+  - action: setHelm
+    installErr: boom
+    keepFailed: true
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect:
+      errorContains: boom
+  - action: setHelm
+    installErr: ""
+    getErr: apiserver timeout
+  - action: reconcile
+    expect:
+      errorContains: apiserver timeout
+      conditions:
+        - type: Applied
+          reason: ApplyFailed
+        - type: Adopted
+          absent: true
+      uninstalls: 0
+      installs: 1
+  - action: setHelm
+    getErr: ""
+  - action: reconcile
+    expect:
+      uninstalls: 1
+      installs: 2
+      upgrades: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Installed
diff --git a/internal/controller/serving/testdata/controller/rolls-back-a-stale-owned-pending-upgrade-and-upgrades.yaml b/internal/controller/serving/testdata/controller/rolls-back-a-stale-owned-pending-upgrade-and-upgrades.yaml
new file mode 100644
index 0000000..c32f519
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/rolls-back-a-stale-owned-pending-upgrade-and-upgrades.yaml
@@ -0,0 +1,34 @@
+name: rolls-back-a-stale-owned-pending-upgrade-and-upgrades
+description: rolls back a stale owned pending-upgrade and upgrades
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 5
+      status: pending-upgrade
+      updated: "-1h"
+  - action: setHistory
+    entries:
+      - owned: true
+        revision: 2
+        status: deployed
+        updated: "-3h"
+      - owned: true
+        revision: 4
+        status: deployed
+        updated: "-2h"
+      - slot: true
+  - action: reconcile
+    expect:
+      rollbacks: 1
+      lastRollback: 4
+      upgrades: 1
+      uninstalls: 0
+      installs: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Upgraded
+      eventContains: RecoveredPendingRelease
diff --git a/internal/controller/serving/testdata/controller/still-drifts-a-foreign-failed-release-when-applied-is-applyfailed.yaml b/internal/controller/serving/testdata/controller/still-drifts-a-foreign-failed-release-when-applied-is-applyfailed.yaml
new file mode 100644
index 0000000..cc1d5d1
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/still-drifts-a-foreign-failed-release-when-applied-is-applyfailed.yaml
@@ -0,0 +1,33 @@
+name: still-drifts-a-foreign-failed-release-when-applied-is-applyfailed
+description: still drifts a foreign failed release when Applied is ApplyFailed
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: setStatus
+    status:
+      phase: Failed
+      conditions:
+        - type: Applied
+          status: "False"
+          reason: ApplyFailed
+          message: boom
+          observedGeneration: current
+  - action: putRelease
+    release:
+      revision: 1
+      status: failed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      config:
+        replicaCount: 1
+  - action: reconcile
+    expect:
+      installs: 0
+      upgrades: 0
+      uninstalls: 0
+      phase: Failed
+      conditions:
+        - type: Adopted
+          reason: Drift
+        - type: Applied
+          reason: Drift
diff --git a/internal/controller/serving/testdata/controller/uninstalls-a-stale-owned-pending-rollback-that-has-no-deployed-revision.yaml b/internal/controller/serving/testdata/controller/uninstalls-a-stale-owned-pending-rollback-that-has-no-deployed-revision.yaml
new file mode 100644
index 0000000..e0787c7
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/uninstalls-a-stale-owned-pending-rollback-that-has-no-deployed-revision.yaml
@@ -0,0 +1,28 @@
+name: uninstalls-a-stale-owned-pending-rollback-that-has-no-deployed-revision
+description: uninstalls a stale owned pending-rollback that has no deployed revision
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 2
+      status: pending-rollback
+      updated: "-1h"
+  - action: setHistory
+    entries:
+      - owned: true
+        revision: 1
+        status: failed
+        updated: "-2h"
+      - slot: true
+  - action: reconcile
+    expect:
+      rollbacks: 0
+      uninstalls: 1
+      installs: 1
+      upgrades: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Installed
diff --git a/internal/controller/serving/testdata/controller/uninstalls-an-adopted-release-on-delete.yaml b/internal/controller/serving/testdata/controller/uninstalls-an-adopted-release-on-delete.yaml
new file mode 100644
index 0000000..5e9fc95
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/uninstalls-an-adopted-release-on-delete.yaml
@@ -0,0 +1,25 @@
+name: uninstalls-an-adopted-release-on-delete
+description: uninstalls an adopted release on delete
+steps:
+  - action: putRelease
+    release:
+      revision: 7
+      status: deployed
+      chartName: sglang
+      chartVersion: "0.8.6"
+      config:
+        replicaCount: 1
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect:
+      conditions:
+        - type: Adopted
+          status: "True"
+      release:
+        labelsAbsent:
+          - serving.modelsphere.dev/llmservice-uid
+  - action: delete
+  - action: reconcile
+    expect:
+      uninstalls: 1
diff --git a/internal/controller/serving/testdata/controller/upgrades-an-owned-deployed-release-whose-status-was-lost.yaml b/internal/controller/serving/testdata/controller/upgrades-an-owned-deployed-release-whose-status-was-lost.yaml
new file mode 100644
index 0000000..d909358
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/upgrades-an-owned-deployed-release-whose-status-was-lost.yaml
@@ -0,0 +1,24 @@
+name: upgrades-an-owned-deployed-release-whose-status-was-lost
+description: upgrades an owned deployed release whose status was lost
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 4
+      status: deployed
+      updated: "-1h"
+  - action: reconcile
+    expect:
+      upgrades: 1
+      installs: 0
+      uninstalls: 0
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Upgraded
+        - type: Adopted
+          absent: true
+      releaseUID: self
+      helmRevision: 5
diff --git a/internal/controller/serving/testdata/controller/upgrades-an-owned-failed-release-that-still-has-a-deployed-revision.yaml b/internal/controller/serving/testdata/controller/upgrades-an-owned-failed-release-that-still-has-a-deployed-revision.yaml
new file mode 100644
index 0000000..138122d
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/upgrades-an-owned-failed-release-that-still-has-a-deployed-revision.yaml
@@ -0,0 +1,28 @@
+name: upgrades-an-owned-failed-release-that-still-has-a-deployed-revision
+description: upgrades an owned failed release that still has a deployed revision
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: setHistory
+    entries:
+      - slot: true
+      - owned: true
+        revision: 2
+        status: failed
+        updated: "0s"
+    put: last
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":2}'
+  - action: reconcile
+    expect:
+      upgrades: 1
+      uninstalls: 0
+      installs: 1
+      phase: Applied
+      conditions:
+        - type: Applied
+          reason: Upgraded
diff --git a/internal/controller/serving/testdata/controller/upgrades-and-bounds-history.yaml b/internal/controller/serving/testdata/controller/upgrades-and-bounds-history.yaml
new file mode 100644
index 0000000..6be49e4
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/upgrades-and-bounds-history.yaml
@@ -0,0 +1,34 @@
+name: upgrades-and-bounds-history
+description: upgrades and bounds history
+historyLimit: 2
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: reconcile
+    expect: {}
+  - action: capture
+    capture:
+      - revision
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":2}'
+  - action: reconcile
+    expect: {}
+  - action: update
+    layers:
+      - name: site
+        values: '{"replicaCount":3}'
+  - action: reconcile
+    expect:
+      upgrades: 2
+      history:
+        length: 2
+        entries:
+          - index: 0
+            revision: 3
+          - index: 1
+            revision: 2
+      helmRevision: 3
+      capturedRevisionGone: true
+      latestRevisionExists: true
diff --git a/internal/controller/serving/testdata/controller/waits-out-a-fresh-owned-pending-upgrade.yaml b/internal/controller/serving/testdata/controller/waits-out-a-fresh-owned-pending-upgrade.yaml
new file mode 100644
index 0000000..0b8a08a
--- /dev/null
+++ b/internal/controller/serving/testdata/controller/waits-out-a-fresh-owned-pending-upgrade.yaml
@@ -0,0 +1,42 @@
+name: waits-out-a-fresh-owned-pending-upgrade
+description: waits out a fresh owned pending-upgrade
+pendingTimeout: "2m"
+steps:
+  - action: create
+    version: "0.8.6"
+  - action: putRelease
+    release:
+      owned: true
+      revision: 2
+      status: pending-upgrade
+      updated: "0s"
+  - action: setHistory
+    entries:
+      - owned: true
+        revision: 1
+        status: deployed
+        updated: "-1h"
+      - slot: true
+  - action: reconcile
+    expect:
+      requeue:
+        atLeast: "15s"
+        above: "1m"
+        atMost: "2m"
+      installs: 0
+      upgrades: 0
+      uninstalls: 0
+      rollbacks: 0
+      phase: Applying
+      conditions:
+        - type: Applied
+          reason: OperationInProgress
+      message:
+        contains:
+          - pending-upgrade
+  - action: reconcile
+    expect:
+      requeue:
+        atLeast: "15s"
+      rollbacks: 0
+      upgrades: 0
diff --git a/internal/controller/serving/testdata/predicate/annotation-change.yaml b/internal/controller/serving/testdata/predicate/annotation-change.yaml
new file mode 100644
index 0000000..6fb070d
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/annotation-change.yaml
@@ -0,0 +1,6 @@
+name: annotation-change
+description: annotation change
+op: update
+annotations:
+  a: "2"
+want: true
diff --git a/internal/controller/serving/testdata/predicate/create.yaml b/internal/controller/serving/testdata/predicate/create.yaml
new file mode 100644
index 0000000..81a4441
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/create.yaml
@@ -0,0 +1,4 @@
+name: create
+description: create reconciles
+op: create
+want: true
diff --git a/internal/controller/serving/testdata/predicate/delete.yaml b/internal/controller/serving/testdata/predicate/delete.yaml
new file mode 100644
index 0000000..c81d4aa
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/delete.yaml
@@ -0,0 +1,4 @@
+name: delete
+description: delete reconciles
+op: delete
+want: true
diff --git a/internal/controller/serving/testdata/predicate/deletion-timestamp.yaml b/internal/controller/serving/testdata/predicate/deletion-timestamp.yaml
new file mode 100644
index 0000000..78803dc
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/deletion-timestamp.yaml
@@ -0,0 +1,5 @@
+name: deletion-timestamp
+description: deletion timestamp
+op: update
+deleting: true
+want: true
diff --git a/internal/controller/serving/testdata/predicate/generation-change.yaml b/internal/controller/serving/testdata/predicate/generation-change.yaml
new file mode 100644
index 0000000..53b8bd0
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/generation-change.yaml
@@ -0,0 +1,5 @@
+name: generation-change
+description: generation change
+op: update
+generation: 2
+want: true
diff --git a/internal/controller/serving/testdata/predicate/status-only-ignored.yaml b/internal/controller/serving/testdata/predicate/status-only-ignored.yaml
new file mode 100644
index 0000000..8504500
--- /dev/null
+++ b/internal/controller/serving/testdata/predicate/status-only-ignored.yaml
@@ -0,0 +1,5 @@
+name: status-only-ignored
+description: status-only ignored
+op: update
+phase: Applied
+want: false
diff --git a/internal/controller/serving/testdata/repo/empty.yaml b/internal/controller/serving/testdata/repo/empty.yaml
new file mode 100644
index 0000000..cec8688
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/empty.yaml
@@ -0,0 +1,7 @@
+name: empty
+description: empty repository is refused
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: ""
+want: false
diff --git a/internal/controller/serving/testdata/repo/https-modelsphere-github-io-helm-charts.yaml b/internal/controller/serving/testdata/repo/https-modelsphere-github-io-helm-charts.yaml
new file mode 100644
index 0000000..c912e68
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/https-modelsphere-github-io-helm-charts.yaml
@@ -0,0 +1,7 @@
+name: https-modelsphere-github-io-helm-charts
+description: "https://modelsphere.github.io/helm-charts"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "https://modelsphere.github.io/helm-charts"
+want: true
diff --git a/internal/controller/serving/testdata/repo/https-modelsphere-github-io.yaml b/internal/controller/serving/testdata/repo/https-modelsphere-github-io.yaml
new file mode 100644
index 0000000..025bfdd
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/https-modelsphere-github-io.yaml
@@ -0,0 +1,7 @@
+name: https-modelsphere-github-io
+description: "https://modelsphere.github.io"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "https://modelsphere.github.io"
+want: false
diff --git a/internal/controller/serving/testdata/repo/nil-allowlist.yaml b/internal/controller/serving/testdata/repo/nil-allowlist.yaml
new file mode 100644
index 0000000..9a153b2
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/nil-allowlist.yaml
@@ -0,0 +1,5 @@
+name: nil-allowlist
+description: no allowlist allows any repository
+nilAllowlist: true
+repo: https://anything.example.com/charts
+want: true
diff --git a/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts-slash.yaml b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts-slash.yaml
new file mode 100644
index 0000000..b6185ce
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts-slash.yaml
@@ -0,0 +1,7 @@
+name: oci-ghcr-io-modelsphere-charts-slash
+description: "oci://ghcr.io/modelsphere/charts/"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "oci://ghcr.io/modelsphere/charts/"
+want: true
diff --git a/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts.yaml b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts.yaml
new file mode 100644
index 0000000..1257aa8
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-charts.yaml
@@ -0,0 +1,7 @@
+name: oci-ghcr-io-modelsphere-charts
+description: "oci://ghcr.io/modelsphere/charts"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "oci://ghcr.io/modelsphere/charts"
+want: true
diff --git a/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-x-charts.yaml b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-x-charts.yaml
new file mode 100644
index 0000000..d1896fd
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere-x-charts.yaml
@@ -0,0 +1,7 @@
+name: oci-ghcr-io-modelsphere-x-charts
+description: "oci://ghcr.io/modelsphere-x/charts"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "oci://ghcr.io/modelsphere-x/charts"
+want: false
diff --git a/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere.yaml b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere.yaml
new file mode 100644
index 0000000..8af55ed
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/oci-ghcr-io-modelsphere.yaml
@@ -0,0 +1,7 @@
+name: oci-ghcr-io-modelsphere
+description: "oci://ghcr.io/modelsphere"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "oci://ghcr.io/modelsphere"
+want: true
diff --git a/internal/controller/serving/testdata/repo/oci-ghcr-io-other-charts.yaml b/internal/controller/serving/testdata/repo/oci-ghcr-io-other-charts.yaml
new file mode 100644
index 0000000..ffde8a6
--- /dev/null
+++ b/internal/controller/serving/testdata/repo/oci-ghcr-io-other-charts.yaml
@@ -0,0 +1,7 @@
+name: oci-ghcr-io-other-charts
+description: "oci://ghcr.io/other/charts"
+allowed:
+  - oci://ghcr.io/modelsphere
+  - https://modelsphere.github.io/helm-charts/
+repo: "oci://ghcr.io/other/charts"
+want: false
diff --git a/internal/controller/serving/testdata/values/hash/stable-and-sensitive.yaml b/internal/controller/serving/testdata/values/hash/stable-and-sensitive.yaml
new file mode 100644
index 0000000..212e663
--- /dev/null
+++ b/internal/controller/serving/testdata/values/hash/stable-and-sensitive.yaml
@@ -0,0 +1,32 @@
+name: stable-and-sensitive
+description: order-stable sha256 and sensitive to version and values
+checks:
+  - kind: stable
+    chart: sglang
+    version: "0.8.6"
+    prefix: "sha256:"
+    hashLen: 71
+    left:
+      b: 1
+      a:
+        d: 1
+        c: 2
+    right:
+      a:
+        c: 2
+        d: 1
+      b: 1
+  - kind: nilEqualsEmpty
+  - kind: version
+    chart: sglang
+    version: "0.8.7"
+    values:
+      b: 1
+      a:
+        d: 1
+        c: 2
+  - kind: values
+    chart: sglang
+    version: "0.8.6"
+    values:
+      b: 2
diff --git a/internal/controller/serving/testdata/values/merge/flat.yaml b/internal/controller/serving/testdata/values/merge/flat.yaml
new file mode 100644
index 0000000..ebe0575
--- /dev/null
+++ b/internal/controller/serving/testdata/values/merge/flat.yaml
@@ -0,0 +1,23 @@
+name: flat
+description: flat merge of four layers
+layers:
+  - name: a
+    values: '{"x":1,"m":{"a":1,"b":2},"s":[1],"keep":true}'
+  - name: b
+    values: '{"x":2,"m":{"b":null,"c":3},"s":[2,3],"n":null}'
+  - name: empty
+  - name: scalar
+    values: '{"m":"replaced"}'
+checks:
+  - path: ["x"]
+    equals: 2
+  - path: ["n"]
+    absent: true
+  - path: ["keep"]
+    equals: true
+  - path: ["m"]
+    equals: replaced
+  - path: ["s"]
+    equals:
+      - 2
+      - 3
diff --git a/internal/controller/serving/testdata/values/merge/nested.yaml b/internal/controller/serving/testdata/values/merge/nested.yaml
new file mode 100644
index 0000000..8c7291e
--- /dev/null
+++ b/internal/controller/serving/testdata/values/merge/nested.yaml
@@ -0,0 +1,12 @@
+name: nested
+description: nested maps merge recursively
+layers:
+  - name: a
+    values: '{"m":{"a":1,"b":{"c":1}}}'
+  - name: b
+    values: '{"m":{"b":{"d":2}}}'
+checks:
+  - path: ["m", "b", "c"]
+    equals: 1
+  - path: ["m", "b", "d"]
+    equals: 2
diff --git a/internal/controller/serving/values_test.go b/internal/controller/serving/values_test.go
index 82b7f64..4980a58 100644
--- a/internal/controller/serving/values_test.go
+++ b/internal/controller/serving/values_test.go
@@ -1,80 +1,125 @@
 package serving
 
 import (
+	"reflect"
 	"strings"
 	"testing"
 
-	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
-
-	servingv1alpha1 "github.com/modelsphere/llm-operator/api/serving/v1alpha1"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-func raw(s string) *apiextensionsv1.JSON {
-	return &apiextensionsv1.JSON{Raw: []byte(s)}
+// mergeCase is one mergeLayers call and the values it must produce.
+// JSON numbers in a layer are float64, and checks compare with the same type.
+type mergeCase struct {
+	Name        string       `json:"name,omitempty"`
+	Description string       `json:"description,omitempty"`
+	Layers      []layerSpec  `json:"layers"`
+	Checks      []mergeCheck `json:"checks"`
+}
+
+type mergeCheck struct {
+	Path   []string `json:"path"`
+	Absent bool     `json:"absent,omitempty"`
+	Equals any      `json:"equals,omitempty"`
 }
 
 func TestMergeLayers(t *testing.T) {
-	got, err := mergeLayers([]servingv1alpha1.Layer{
-		{Name: "a", Values: raw(`{"x":1,"m":{"a":1,"b":2},"s":[1],"keep":true}`)},
-		{Name: "b", Values: raw(`{"x":2,"m":{"b":null,"c":3},"s":[2,3],"n":null}`)},
-		{Name: "empty"},
-		{Name: "scalar", Values: raw(`{"m":"replaced"}`)},
+	cases.Run(t, "testdata/values/merge", func(t *testing.T, tc mergeCase) {
+		got, err := mergeLayers(layersFrom(tc.Layers))
+		if err != nil {
+			t.Fatal(err)
+		}
+		for _, check := range tc.Checks {
+			val, ok := lookupPath(got, check.Path)
+			if check.Absent {
+				if ok {
+					t.Errorf("%s: %v was kept: %#v", tc.Description, check.Path, val)
+				}
+				continue
+			}
+			if !ok || !reflect.DeepEqual(val, check.Equals) {
+				t.Errorf("%s: %v = %#v, want %#v", tc.Description, check.Path, val, check.Equals)
+			}
+		}
 	})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if got["x"] != float64(2) {
-		t.Errorf("x = %#v", got["x"])
-	}
-	if _, ok := got["n"]; ok {
-		t.Errorf("null key n was kept: %#v", got["n"])
-	}
-	if got["keep"] != true {
-		t.Errorf("keep = %#v", got["keep"])
-	}
-	m, ok := got["m"].(string)
-	if !ok || m != "replaced" {
-		t.Errorf("map replaced by scalar = %#v", got["m"])
-	}
-	s, ok := got["s"].([]any)
-	if !ok || len(s) != 2 || s[0] != float64(2) || s[1] != float64(3) {
-		t.Errorf("slice = %#v", got["s"])
-	}
+}
 
-	nested, err := mergeLayers([]servingv1alpha1.Layer{
-		{Name: "a", Values: raw(`{"m":{"a":1,"b":{"c":1}}}`)},
-		{Name: "b", Values: raw(`{"m":{"b":{"d":2}}}`)},
-	})
-	if err != nil {
-		t.Fatal(err)
-	}
-	inner := nested["m"].(map[string]any)["b"].(map[string]any)
-	if inner["c"] != float64(1) || inner["d"] != float64(2) {
-		t.Errorf("nested merge = %#v", nested["m"])
+func lookupPath(root map[string]any, path []string) (any, bool) {
+	var cur any = root
+	for _, p := range path {
+		m, ok := cur.(map[string]any)
+		if !ok {
+			return nil, false
+		}
+		cur, ok = m[p]
+		if !ok {
+			return nil, false
+		}
 	}
+	return cur, true
+}
+
+// hashCase runs ordered appliedHash checks. The stable check records the
+// digest the later checks compare against, so it has to come first.
+type hashCase struct {
+	Name        string      `json:"name,omitempty"`
+	Description string      `json:"description,omitempty"`
+	Checks      []hashCheck `json:"checks"`
+}
+
+type hashCheck struct {
+	Kind    string         `json:"kind"`
+	Chart   string         `json:"chart,omitempty"`
+	Version string         `json:"version,omitempty"`
+	Left    map[string]any `json:"left,omitempty"`
+	Right   map[string]any `json:"right,omitempty"`
+	Values  map[string]any `json:"values,omitempty"`
+	Prefix  string         `json:"prefix,omitempty"`
+	HashLen int            `json:"hashLen,omitempty"`
 }
 
 func TestAppliedHashIsStableAndSensitive(t *testing.T) {
-	a, err := appliedHash("sglang", "0.8.6", map[string]any{"b": 1, "a": map[string]any{"d": 1, "c": 2}})
-	if err != nil {
-		t.Fatal(err)
-	}
-	b, err := appliedHash("sglang", "0.8.6", map[string]any{"a": map[string]any{"c": 2, "d": 1}, "b": 1})
-	if err != nil {
-		t.Fatal(err)
-	}
-	if a != b || !strings.HasPrefix(a, "sha256:") || len(a) != len("sha256:")+64 {
-		t.Fatalf("hash = %s and %s", a, b)
-	}
-	if same, err := jsonEqual(nil, map[string]any{}); err != nil || !same {
-		t.Fatalf("nil and empty map: %v %v", same, err)
-	}
-	changed, err := appliedHash("sglang", "0.8.7", map[string]any{"b": 1, "a": map[string]any{"d": 1, "c": 2}})
-	if err != nil || changed == a {
-		t.Fatalf("version change did not move the hash: %s", changed)
-	}
-	changed, err = appliedHash("sglang", "0.8.6", map[string]any{"b": 2})
-	if err != nil || changed == a {
-		t.Fatalf("values change did not move the hash: %s", changed)
-	}
+	cases.Run(t, "testdata/values/hash", func(t *testing.T, tc hashCase) {
+		var base string
+		for _, check := range tc.Checks {
+			switch check.Kind {
+			case "stable":
+				a, err := appliedHash(check.Chart, check.Version, check.Left)
+				if err != nil {
+					t.Fatal(err)
+				}
+				b, err := appliedHash(check.Chart, check.Version, check.Right)
+				if err != nil {
+					t.Fatal(err)
+				}
+				if a != b || !strings.HasPrefix(a, check.Prefix) || len(a) != check.HashLen {
+					t.Fatalf("hash = %s and %s", a, b)
+				}
+				base = a
+			case "nilEqualsEmpty":
+				same, err := jsonEqual(nil, map[string]any{})
+				if err != nil || !same {
+					t.Fatalf("nil and empty map: %v %v", same, err)
+				}
+			case "version":
+				if base == "" {
+					t.Fatal("stable hash was not recorded")
+				}
+				changed, err := appliedHash(check.Chart, check.Version, check.Values)
+				if err != nil || changed == base {
+					t.Fatalf("version change did not move the hash: %s", changed)
+				}
+			case "values":
+				if base == "" {
+					t.Fatal("stable hash was not recorded")
+				}
+				changed, err := appliedHash(check.Chart, check.Version, check.Values)
+				if err != nil || changed == base {
+					t.Fatalf("values change did not move the hash: %s", changed)
+				}
+			default:
+				t.Fatalf("%s: unknown check %q", tc.Name, check.Kind)
+			}
+		}
+	})
 }
diff --git a/internal/helm/guard_test.go b/internal/helm/guard_test.go
index db45f51..4702947 100644
--- a/internal/helm/guard_test.go
+++ b/internal/helm/guard_test.go
@@ -6,73 +6,111 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-const clusterScoped = "cluster-scoped"
+// namespaceCase is one post-render. An empty refused string means the
+// manifest is returned unchanged and must equal the input bytes.
+type namespaceCase struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Manifest    string `json:"manifest"`
+	Refused     string `json:"refused,omitempty"`
+}
 
 func TestNamespaceGuard(t *testing.T) {
-	g := namespaceGuard{namespace: "models", mapper: testMapper()}
-	for name, tc := range map[string]struct {
-		manifest, refused string
-	}{
-		"own namespace":     {"apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: a\n  namespace: models\n", ""},
-		"no namespace":      {"apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: a\n", ""},
-		"comments only":     {"# Source: x/templates/empty.yaml\n---\n", ""},
-		"another namespace": {"apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: a\n  namespace: kube-system\n", "into namespace kube-system"},
-		clusterScoped:       {"apiVersion: rbac.authorization.k8s.io/v1\nkind: ClusterRole\nmetadata:\n  name: a\n", "ClusterRole \"a\", which is cluster-scoped"},
-		"namespace object":  {"apiVersion: v1\nkind: Namespace\nmetadata:\n  name: a\n", clusterScoped},
-		"inside a list": {"apiVersion: v1\nkind: List\nitems:\n- apiVersion: v1\n  kind: ConfigMap\n  metadata:\n    name: a\n" +
-			"- apiVersion: rbac.authorization.k8s.io/v1\n  kind: Role\n  metadata:\n    name: b\n    namespace: kube-system\n", "Role \"b\" into namespace kube-system"},
-		"after a good one": {"apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: a\n---\napiVersion: rbac.authorization.k8s.io/v1\nkind: ClusterRole\nmetadata:\n  name: b\n", clusterScoped},
-		"unknown kind":     {"apiVersion: example.com/v1\nkind: Widget\nmetadata:\n  name: a\n", "cannot be told"},
-		"no kind":          {"replicaCount: 1\n", "an object with no kind"},
-	} {
-		out, err := g.Run(bytes.NewBufferString(tc.manifest))
-		switch {
-		case tc.refused == "" && err != nil:
-			t.Errorf("%s: refused: %v", name, err)
-		case tc.refused == "" && out.String() != tc.manifest:
-			t.Errorf("%s: manifests changed:\n%s", name, out)
-		case tc.refused != "" && (err == nil || !strings.Contains(err.Error(), tc.refused)):
-			t.Errorf("%s: got %v, want an error with %q", name, err, tc.refused)
+	cases.Run(t, "testdata/guard/namespace", func(t *testing.T, tc namespaceCase) {
+		g := namespaceGuard{namespace: "models", mapper: testMapper()}
+		out, err := g.Run(bytes.NewBufferString(tc.Manifest))
+		if tc.Refused == "" {
+			if err != nil {
+				t.Fatalf("refused: %v", err)
+			}
+			if out == nil || out.String() != tc.Manifest {
+				got := ""
+				if out != nil {
+					got = out.String()
+				}
+				t.Fatalf("manifests changed:\ngot %q\nwant %q", got, tc.Manifest)
+			}
+			return
 		}
-	}
+		if err == nil || !strings.Contains(err.Error(), tc.Refused) {
+			t.Fatalf("got %v, want an error with %q", err, tc.Refused)
+		}
+	})
+}
+
+// applyCase is one install or upgrade sequence on a fresh client.
+// The refused template and hook installs do not share the later upgrade.
+type applyCase struct {
+	Name        string      `json:"name,omitempty"`
+	Description string      `json:"description,omitempty"`
+	Steps       []applyStep `json:"steps"`
+}
+
+type applyStep struct {
+	Write         string `json:"write,omitempty"`
+	Op            string `json:"op,omitempty"`
+	ErrorContains string `json:"errorContains,omitempty"`
+	Absent        bool   `json:"absent,omitempty"`
 }
 
-// Install and Upgrade run the guard, hooks included, and store nothing a
-// refused chart rendered.
 func TestApplyRefusesAnotherNamespace(t *testing.T) {
-	for name, tpl := range map[string]string{
-		"template": "apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: escape\n  namespace: kube-system\n",
-		"hook": "apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: escape\n  namespace: kube-system\n" +
-			"  annotations:\n    helm.sh/hook: pre-install,pre-upgrade\n",
-	} {
-		dir := writeTinyChart(t)
-		if err := os.WriteFile(filepath.Join(dir, "templates", "escape.yaml"), []byte(tpl), 0o644); err != nil {
-			t.Fatal(err)
-		}
-		c, _ := testClient(t, dir)
-		ref := ChartRef{Repo: testRepo, Name: tinyChartName, Version: tinyVersion}
-		if _, err := c.Install(t.Context(), testNamespace, tinyChartName, ref, nil, ApplyOpts{}); err == nil || !strings.Contains(err.Error(), "kube-system") {
-			t.Errorf("%s: install: got %v, want a refusal", name, err)
-		}
-		if got, err := c.Get(t.Context(), testNamespace, tinyChartName); err != nil || got != nil {
-			t.Errorf("%s: a refused install stored %#v, %v", name, got, err)
-		}
-	}
+	cases.Run(t, "testdata/guard/apply", runApplyCase)
+}
 
-	// A chart that changes to write elsewhere is refused on upgrade.
+func runApplyCase(t *testing.T, tc applyCase) {
+	t.Helper()
 	dir := writeTinyChart(t)
 	c, _ := testClient(t, dir)
 	ref := ChartRef{Repo: testRepo, Name: tinyChartName, Version: tinyVersion}
-	if _, err := c.Install(t.Context(), testNamespace, tinyChartName, ref, nil, ApplyOpts{}); err != nil {
-		t.Fatal(err)
-	}
-	cluster := "apiVersion: rbac.authorization.k8s.io/v1\nkind: ClusterRole\nmetadata:\n  name: escape\n"
-	if err := os.WriteFile(filepath.Join(dir, "templates", "escape.yaml"), []byte(cluster), 0o644); err != nil {
-		t.Fatal(err)
-	}
-	if _, err := c.Upgrade(t.Context(), testNamespace, tinyChartName, ref, nil, ApplyOpts{}); err == nil || !strings.Contains(err.Error(), clusterScoped) {
-		t.Errorf("upgrade: got %v, want a refusal", err)
+	ctx := t.Context()
+	for _, step := range tc.Steps {
+		if step.Write != "" {
+			err := os.WriteFile(filepath.Join(dir, "templates", "escape.yaml"), []byte(step.Write), 0o644)
+			if err != nil {
+				t.Fatal(err)
+			}
+		}
+		switch step.Op {
+		case "":
+			if step.Write == "" {
+				t.Fatal("step has no write and no op")
+			}
+		case "install":
+			_, err := c.Install(ctx, testNamespace, tinyChartName, ref, nil, ApplyOpts{})
+			if step.ErrorContains == "" {
+				if err != nil {
+					t.Fatal(err)
+				}
+				continue
+			}
+			if err == nil || !strings.Contains(err.Error(), step.ErrorContains) {
+				t.Errorf("install: got %v, want a refusal", err)
+			}
+		case "upgrade":
+			_, err := c.Upgrade(ctx, testNamespace, tinyChartName, ref, nil, ApplyOpts{})
+			if step.ErrorContains == "" {
+				if err != nil {
+					t.Fatal(err)
+				}
+				continue
+			}
+			if err == nil || !strings.Contains(err.Error(), step.ErrorContains) {
+				t.Errorf("upgrade: got %v, want a refusal", err)
+			}
+		case "get":
+			got, err := c.Get(ctx, testNamespace, tinyChartName)
+			if !step.Absent {
+				t.Fatal("get step expects absent")
+			}
+			if err != nil || got != nil {
+				t.Errorf("a refused install stored %#v, %v", got, err)
+			}
+		default:
+			t.Fatalf("op %q", step.Op)
+		}
 	}
 }
diff --git a/internal/helm/helm_test.go b/internal/helm/helm_test.go
index 061754c..a375383 100644
--- a/internal/helm/helm_test.go
+++ b/internal/helm/helm_test.go
@@ -19,16 +19,15 @@ import (
 	"helm.sh/helm/v4/pkg/storage/driver"
 	"k8s.io/apimachinery/pkg/api/meta"
 	"k8s.io/apimachinery/pkg/runtime/schema"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
 const (
-	tinyVersion    = "0.1.0"
-	tinyChartName  = "tiny"
-	testNamespace  = "default"
-	testReleaseUID = "uid-1"
-	uidLabelKey    = "serving.modelsphere.dev/llmservice-uid"
-	nameLabelKey   = "serving.modelsphere.dev/llmservice"
-	testRepo       = "oci://example.com/charts"
+	tinyVersion   = "0.1.0"
+	tinyChartName = "tiny"
+	testNamespace = "default"
+	testRepo      = "oci://example.com/charts"
 )
 
 func testClient(t *testing.T, chartDir string) (*Client, *action.Configuration) {
@@ -80,196 +79,227 @@ func writeTinyChart(t *testing.T) string {
 	return dir
 }
 
-func TestGetInstallUpgradeUninstall(t *testing.T) {
-	const valueOne = "one"
-	c, _ := testClient(t, writeTinyChart(t))
-	ctx := t.Context()
-	const ns, name = testNamespace, tinyChartName
-	ref := ChartRef{Repo: testRepo, Name: name, Version: tinyVersion}
-	labels := testLabels(name)
-	before := time.Now().Add(-time.Second)
-
-	got, err := c.Get(ctx, ns, name)
-	if err != nil || got != nil {
-		t.Fatalf("absent get = %#v, %v", got, err)
+func storeRelease(t *testing.T, cfg *action.Configuration, rel *releasev1.Release) {
+	t.Helper()
+	if err := cfg.Releases.Create(rel); err != nil {
+		t.Fatal(err)
 	}
+}
 
-	installed, err := c.Install(ctx, ns, name, ref, map[string]any{"v": valueOne}, ApplyOpts{ForceConflicts: true, Labels: labels})
-	if err != nil {
-		t.Fatalf("install: %v", err)
-	}
-	if installed.Revision != 1 || installed.Status != "deployed" || installed.ChartName != name || installed.ChartVersion != tinyVersion {
-		t.Fatalf("install release = %#v", installed)
-	}
-	if installed.Config["v"] != valueOne {
-		t.Fatalf("install config = %#v", installed.Config)
-	}
-	if installed.Labels[uidLabelKey] != testReleaseUID || installed.Updated.Before(before) {
-		t.Fatalf("install labels/updated = %#v", installed)
+func storedRelease(name string, rev int, status common.Status, updated time.Time, labels map[string]string) *releasev1.Release {
+	return &releasev1.Release{
+		Name:      name,
+		Namespace: testNamespace,
+		Version:   rev,
+		Labels:    labels,
+		Config:    map[string]any{"v": "one"},
+		Manifest:  "apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: " + tinyChartName + "\n",
+		Chart:     &chartv2.Chart{Metadata: &chartv2.Metadata{Name: tinyChartName, Version: tinyVersion}},
+		Info: &releasev1.Info{
+			Status:        status,
+			LastDeployed:  updated,
+			FirstDeployed: updated,
+		},
 	}
+}
 
-	got, err = c.Get(ctx, ns, name)
-	if err != nil || got == nil || got.Revision != 1 || got.Config["v"] != valueOne {
-		t.Fatalf("get after install = %#v, %v", got, err)
-	}
-	if got.Labels[uidLabelKey] != testReleaseUID || got.Updated.Before(before) {
-		t.Fatalf("get labels/updated = %#v", got)
-	}
+// clientCase is one client sequence. Seed times are offsets from a base
+// captured at the start, stored as base.Add(offset).UTC().
+// updatedAfterStart mirrors !Updated.Before(base.Add(-time.Second)).
+type clientCase struct {
+	Name         string        `json:"name,omitempty"`
+	Description  string        `json:"description,omitempty"`
+	Namespace    string        `json:"namespace,omitempty"`
+	ChartRepo    string        `json:"chartRepo,omitempty"`
+	ChartName    string        `json:"chartName,omitempty"`
+	ChartVersion string        `json:"chartVersion,omitempty"`
+	Seed         []seedRelease `json:"seed,omitempty"`
+	Steps        []clientStep  `json:"steps"`
+}
 
-	upgraded, err := c.Upgrade(ctx, ns, name, ref, map[string]any{"v": "two"}, ApplyOpts{Labels: labels})
-	if err != nil {
-		t.Fatalf("upgrade: %v", err)
-	}
-	if upgraded.Revision != 2 || upgraded.Config["v"] != "two" || upgraded.Status != "deployed" {
-		t.Fatalf("upgrade release = %#v", upgraded)
-	}
-	if upgraded.Labels[uidLabelKey] != testReleaseUID || upgraded.Updated.Before(before) {
-		t.Fatalf("upgrade labels/updated = %#v", upgraded)
-	}
+type seedRelease struct {
+	Name     string            `json:"name"`
+	Revision int               `json:"revision"`
+	Status   string            `json:"status"`
+	Updated  cases.Duration    `json:"updated"`
+	Labels   map[string]string `json:"labels,omitempty"`
+}
 
-	if err := c.Uninstall(ctx, ns, name); err != nil {
-		t.Fatalf("uninstall: %v", err)
-	}
-	got, err = c.Get(ctx, ns, name)
-	if err != nil || got != nil {
-		t.Fatalf("get after uninstall = %#v, %v", got, err)
-	}
-	if err := c.Uninstall(ctx, ns, name); err != nil {
-		t.Fatalf("uninstall absent: %v", err)
-	}
+type clientStep struct {
+	Op             string            `json:"op"`
+	Release        string            `json:"release"`
+	Values         map[string]any    `json:"values,omitempty"`
+	ForceConflicts bool              `json:"forceConflicts,omitempty"`
+	Labels         map[string]string `json:"labels,omitempty"`
+	Revision       int               `json:"revision,omitempty"`
+	Absent         bool              `json:"absent,omitempty"`
+	Expect         *releaseExpect    `json:"expect,omitempty"`
+	History        []releaseExpect   `json:"history,omitempty"`
+}
+
+type releaseExpect struct {
+	Revision          *int              `json:"revision,omitempty"`
+	Status            string            `json:"status,omitempty"`
+	ChartName         string            `json:"chartName,omitempty"`
+	ChartVersion      string            `json:"chartVersion,omitempty"`
+	Config            map[string]string `json:"config,omitempty"`
+	Labels            map[string]string `json:"labels,omitempty"`
+	Updated           *cases.Duration   `json:"updated,omitempty"`
+	UpdatedAfterStart bool              `json:"updatedAfterStart,omitempty"`
+}
+
+func TestGetInstallUpgradeUninstall(t *testing.T) {
+	cases.Run(t, "testdata/client/get-install-upgrade-uninstall", runClientCase)
 }
 
 func TestHistoryNewestFirstAndRollbackPendingUpgrade(t *testing.T) {
-	c, cfg := testClient(t, writeTinyChart(t))
-	ctx := t.Context()
-	const ns, name = testNamespace, tinyChartName
-	labels := testLabels(name)
-	deployedAt := time.Now().Add(-2 * time.Hour).UTC()
-	failedAt := time.Now().Add(-time.Hour).UTC()
-	pendingAt := time.Now().Add(-30 * time.Minute).UTC()
-	storeRelease(t, cfg, storedRelease(name, 1, common.StatusDeployed, deployedAt, labels))
-	storeRelease(t, cfg, storedRelease(name, 2, common.StatusFailed, failedAt, labels))
-	storeRelease(t, cfg, storedRelease(name, 3, common.StatusPendingUpgrade, pendingAt, labels))
+	cases.Run(t, "testdata/client/history-newest-first", runClientCase)
+}
 
-	got, err := c.Get(ctx, ns, name)
-	if err != nil {
-		t.Fatalf("get: %v", err)
-	}
-	if got.Revision != 3 || got.Status != string(common.StatusPendingUpgrade) || !got.Updated.Equal(pendingAt) {
-		t.Fatalf("get pending = %#v", got)
-	}
-	if got.Labels[uidLabelKey] != testReleaseUID {
-		t.Fatalf("get labels = %#v", got.Labels)
-	}
+func TestPendingInstallFailedAndAbsentHistory(t *testing.T) {
+	cases.Run(t, "testdata/client/pending-install-failed", runClientCase)
+}
+
+func TestRollbackPendingRollback(t *testing.T) {
+	cases.Run(t, "testdata/client/rollback-pending-rollback", runClientCase)
+}
 
-	hist, err := c.History(ctx, ns, name)
-	if err != nil {
-		t.Fatalf("history: %v", err)
+func runClientCase(t *testing.T, tc clientCase) {
+	t.Helper()
+	base := time.Now()
+	ns := tc.Namespace
+	if ns == "" {
+		ns = testNamespace
 	}
-	if len(hist) != 3 || hist[0].Revision != 3 || hist[1].Revision != 2 || hist[2].Revision != 1 {
-		t.Fatalf("history = %#v", hist)
+	c, cfg := testClient(t, writeTinyChart(t))
+	for _, seed := range tc.Seed {
+		at := base.Add(seed.Updated.Std()).UTC()
+		rel := storedRelease(seed.Name, seed.Revision, common.Status(seed.Status), at, seed.Labels)
+		storeRelease(t, cfg, rel)
 	}
-	if hist[0].Status != string(common.StatusPendingUpgrade) || hist[1].Status != string(common.StatusFailed) || hist[2].Status != string(common.StatusDeployed) {
-		t.Fatalf("history status = %#v", hist)
+	ref := ChartRef{Repo: tc.ChartRepo, Name: tc.ChartName, Version: tc.ChartVersion}
+	if ref.Repo == "" {
+		ref.Repo = testRepo
 	}
-	if !hist[2].Updated.Equal(deployedAt) || hist[2].Labels[uidLabelKey] != testReleaseUID {
-		t.Fatalf("deployed history = %#v", hist[2])
+	if ref.Name == "" {
+		ref.Name = tinyChartName
 	}
-
-	if err := c.Rollback(ctx, ns, name, 1); err != nil {
-		t.Fatalf("rollback: %v", err)
+	if ref.Version == "" {
+		ref.Version = tinyVersion
 	}
-	got, err = c.Get(ctx, ns, name)
-	if err != nil {
-		t.Fatalf("get after rollback: %v", err)
-	}
-	if got.Status != string(common.StatusDeployed) || got.Revision != 4 || got.Labels[uidLabelKey] != testReleaseUID || got.ChartName != name {
-		t.Fatalf("rolled back = %#v", got)
+	ctx := t.Context()
+	for _, step := range tc.Steps {
+		runClientStep(t, ctx, c, ns, ref, base, step)
 	}
 }
 
-func TestPendingInstallFailedAndAbsentHistory(t *testing.T) {
-	c, cfg := testClient(t, writeTinyChart(t))
-	ctx := t.Context()
-	const ns = testNamespace
-	labels := testLabels("fresh")
-	failedAt := time.Now().Add(-time.Hour).UTC()
-	installAt := time.Now().Add(-3 * time.Hour).UTC()
-	storeRelease(t, cfg, storedRelease("fresh", 1, common.StatusPendingInstall, installAt, labels))
-	storeRelease(t, cfg, storedRelease("broke", 1, common.StatusFailed, failedAt, nil))
+func runClientStep(t *testing.T, ctx context.Context, c *Client, ns string, ref ChartRef, base time.Time, step clientStep) {
+	t.Helper()
+	if step.Release == "" {
+		t.Fatal("step has no release")
+	}
+	switch step.Op {
+	case "get":
+		got, err := c.Get(ctx, ns, step.Release)
+		if step.Absent {
+			if err != nil || got != nil {
+				t.Fatalf("absent get = %#v, %v", got, err)
+			}
+			return
+		}
+		if err != nil {
+			t.Fatalf("get: %v", err)
+		}
+		checkRelease(t, "get", got, step.Expect, base)
+	case "install":
+		got, err := c.Install(ctx, ns, step.Release, ref, step.Values, ApplyOpts{
+			ForceConflicts: step.ForceConflicts,
+			Labels:         step.Labels,
+		})
+		if err != nil {
+			t.Fatalf("install: %v", err)
+		}
+		checkRelease(t, "install", got, step.Expect, base)
+	case "upgrade":
+		got, err := c.Upgrade(ctx, ns, step.Release, ref, step.Values, ApplyOpts{
+			ForceConflicts: step.ForceConflicts,
+			Labels:         step.Labels,
+		})
+		if err != nil {
+			t.Fatalf("upgrade: %v", err)
+		}
+		checkRelease(t, "upgrade", got, step.Expect, base)
+	case "uninstall":
+		if err := c.Uninstall(ctx, ns, step.Release); err != nil {
+			t.Fatalf("uninstall: %v", err)
+		}
+	case "history":
+		hist, err := c.History(ctx, ns, step.Release)
+		if err != nil {
+			t.Fatalf("history: %v", err)
+		}
+		if step.Absent {
+			if hist != nil {
+				t.Fatalf("absent history = %#v, %v", hist, err)
+			}
+			return
+		}
+		if len(hist) != len(step.History) {
+			t.Fatalf("history = %#v", hist)
+		}
+		for i := range step.History {
+			checkRelease(t, "history", &hist[i], &step.History[i], base)
+		}
+	case "rollback":
+		if step.Revision == 0 {
+			t.Fatal("rollback has no revision")
+		}
+		if err := c.Rollback(ctx, ns, step.Release, step.Revision); err != nil {
+			t.Fatalf("rollback: %v", err)
+		}
+	default:
+		t.Fatalf("op %q", step.Op)
+	}
+}
 
-	hist, err := c.History(ctx, ns, "fresh")
-	if err != nil || len(hist) != 1 || hist[0].Status != string(common.StatusPendingInstall) || !hist[0].Updated.Equal(installAt) {
-		t.Fatalf("pending-install history = %#v, %v", hist, err)
+func checkRelease(t *testing.T, what string, rel *Release, exp *releaseExpect, base time.Time) {
+	t.Helper()
+	if rel == nil {
+		t.Fatalf("%s = nil", what)
 	}
-	if err := c.Uninstall(ctx, ns, "fresh"); err != nil {
-		t.Fatalf("uninstall pending-install: %v", err)
+	if exp == nil {
+		t.Fatalf("%s has no expect", what)
 	}
-	got, err := c.Get(ctx, ns, "fresh")
-	if err != nil || got != nil {
-		t.Fatalf("get after uninstall pending-install = %#v, %v", got, err)
+	if exp.Revision != nil && rel.Revision != *exp.Revision {
+		t.Fatalf("%s revision = %d, want %d", what, rel.Revision, *exp.Revision)
 	}
-
-	got, err = c.Get(ctx, ns, "broke")
-	if err != nil || got == nil || got.Status != string(common.StatusFailed) || !got.Updated.Equal(failedAt) {
-		t.Fatalf("failed get = %#v, %v", got, err)
+	if exp.Status != "" && rel.Status != exp.Status {
+		t.Fatalf("%s status = %q, want %q", what, rel.Status, exp.Status)
 	}
-	hist, err = c.History(ctx, ns, "broke")
-	if err != nil || len(hist) != 1 || hist[0].Status != string(common.StatusFailed) || hist[0].Revision != 1 {
-		t.Fatalf("failed history = %#v, %v", hist, err)
+	if exp.ChartName != "" && rel.ChartName != exp.ChartName {
+		t.Fatalf("%s chart = %q, want %q", what, rel.ChartName, exp.ChartName)
 	}
-
-	missing, err := c.History(ctx, ns, "absent")
-	if err != nil || missing != nil {
-		t.Fatalf("absent history = %#v, %v", missing, err)
+	if exp.ChartVersion != "" && rel.ChartVersion != exp.ChartVersion {
+		t.Fatalf("%s chart version = %q, want %q", what, rel.ChartVersion, exp.ChartVersion)
 	}
-}
-
-func TestRollbackPendingRollback(t *testing.T) {
-	c, cfg := testClient(t, writeTinyChart(t))
-	ctx := t.Context()
-	const ns, name = testNamespace, "back"
-	deployedAt := time.Now().Add(-2 * time.Hour).UTC()
-	pendingAt := time.Now().Add(-30 * time.Minute).UTC()
-	storeRelease(t, cfg, storedRelease(name, 1, common.StatusDeployed, deployedAt, testLabels(name)))
-	storeRelease(t, cfg, storedRelease(name, 2, common.StatusPendingRollback, pendingAt, testLabels(name)))
-
-	if err := c.Rollback(ctx, ns, name, 1); err != nil {
-		t.Fatalf("rollback pending-rollback: %v", err)
+	for k, want := range exp.Config {
+		got, ok := rel.Config[k].(string)
+		if !ok || got != want {
+			t.Fatalf("%s config[%s] = %#v, want %q", what, k, rel.Config[k], want)
+		}
 	}
-	got, err := c.Get(ctx, ns, name)
-	if err != nil || got == nil || got.Status != string(common.StatusDeployed) || got.Revision != 3 {
-		t.Fatalf("pending-rollback recovered = %#v, %v", got, err)
+	for k, want := range exp.Labels {
+		if rel.Labels[k] != want {
+			t.Fatalf("%s label %s = %q, want %q", what, k, rel.Labels[k], want)
+		}
 	}
-}
-
-func testLabels(name string) map[string]string {
-	return map[string]string{
-		uidLabelKey:  testReleaseUID,
-		nameLabelKey: name,
-	}
-}
-
-func storeRelease(t *testing.T, cfg *action.Configuration, rel *releasev1.Release) {
-	t.Helper()
-	if err := cfg.Releases.Create(rel); err != nil {
-		t.Fatal(err)
+	if exp.Updated != nil {
+		want := base.Add(exp.Updated.Std()).UTC()
+		if !rel.Updated.Equal(want) {
+			t.Fatalf("%s updated = %s, want %s", what, rel.Updated, want)
+		}
 	}
-}
-
-func storedRelease(name string, rev int, status common.Status, updated time.Time, labels map[string]string) *releasev1.Release {
-	return &releasev1.Release{
-		Name:      name,
-		Namespace: testNamespace,
-		Version:   rev,
-		Labels:    labels,
-		Config:    map[string]any{"v": "one"},
-		Manifest:  "apiVersion: v1\nkind: ConfigMap\nmetadata:\n  name: " + tinyChartName + "\n",
-		Chart:     &chartv2.Chart{Metadata: &chartv2.Metadata{Name: tinyChartName, Version: tinyVersion}},
-		Info: &releasev1.Info{
-			Status:        status,
-			LastDeployed:  updated,
-			FirstDeployed: updated,
-		},
+	if exp.UpdatedAfterStart && rel.Updated.Before(base.Add(-time.Second)) {
+		t.Fatalf("%s updated = %s, want after %s", what, rel.Updated, base.Add(-time.Second))
 	}
 }
diff --git a/internal/helm/testdata/client/get-install-upgrade-uninstall/get-install-upgrade-uninstall.yaml b/internal/helm/testdata/client/get-install-upgrade-uninstall/get-install-upgrade-uninstall.yaml
new file mode 100644
index 0000000..057a679
--- /dev/null
+++ b/internal/helm/testdata/client/get-install-upgrade-uninstall/get-install-upgrade-uninstall.yaml
@@ -0,0 +1,58 @@
+description: get, install, upgrade, and uninstall
+namespace: default
+chartRepo: oci://example.com/charts
+chartName: tiny
+chartVersion: "0.1.0"
+steps:
+  - op: get
+    release: tiny
+    absent: true
+  - op: install
+    release: tiny
+    forceConflicts: true
+    values:
+      v: one
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: tiny
+    expect:
+      revision: 1
+      status: deployed
+      chartName: tiny
+      chartVersion: "0.1.0"
+      config:
+        v: one
+      labels:
+        serving.modelsphere.dev/llmservice-uid: uid-1
+      updatedAfterStart: true
+  - op: get
+    release: tiny
+    expect:
+      revision: 1
+      config:
+        v: one
+      labels:
+        serving.modelsphere.dev/llmservice-uid: uid-1
+      updatedAfterStart: true
+  - op: upgrade
+    release: tiny
+    values:
+      v: two
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: tiny
+    expect:
+      revision: 2
+      status: deployed
+      config:
+        v: two
+      labels:
+        serving.modelsphere.dev/llmservice-uid: uid-1
+      updatedAfterStart: true
+  - op: uninstall
+    release: tiny
+  - op: get
+    release: tiny
+    absent: true
+  - op: uninstall
+    release: tiny
diff --git a/internal/helm/testdata/client/history-newest-first/history-newest-first.yaml b/internal/helm/testdata/client/history-newest-first/history-newest-first.yaml
new file mode 100644
index 0000000..012f654
--- /dev/null
+++ b/internal/helm/testdata/client/history-newest-first/history-newest-first.yaml
@@ -0,0 +1,56 @@
+description: history newest first and rollback of a pending upgrade
+namespace: default
+seed:
+  - name: tiny
+    revision: 1
+    status: deployed
+    updated: "-2h"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: tiny
+  - name: tiny
+    revision: 2
+    status: failed
+    updated: "-1h"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: tiny
+  - name: tiny
+    revision: 3
+    status: pending-upgrade
+    updated: "-30m"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: tiny
+steps:
+  - op: get
+    release: tiny
+    expect:
+      revision: 3
+      status: pending-upgrade
+      updated: "-30m"
+      labels:
+        serving.modelsphere.dev/llmservice-uid: uid-1
+  - op: history
+    release: tiny
+    history:
+      - revision: 3
+        status: pending-upgrade
+      - revision: 2
+        status: failed
+      - revision: 1
+        status: deployed
+        updated: "-2h"
+        labels:
+          serving.modelsphere.dev/llmservice-uid: uid-1
+  - op: rollback
+    release: tiny
+    revision: 1
+  - op: get
+    release: tiny
+    expect:
+      revision: 4
+      status: deployed
+      chartName: tiny
+      labels:
+        serving.modelsphere.dev/llmservice-uid: uid-1
diff --git a/internal/helm/testdata/client/pending-install-failed/pending-install-failed.yaml b/internal/helm/testdata/client/pending-install-failed/pending-install-failed.yaml
new file mode 100644
index 0000000..31731c6
--- /dev/null
+++ b/internal/helm/testdata/client/pending-install-failed/pending-install-failed.yaml
@@ -0,0 +1,38 @@
+description: pending install, a failed release, and absent history
+namespace: default
+seed:
+  - name: fresh
+    revision: 1
+    status: pending-install
+    updated: "-3h"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: fresh
+  - name: broke
+    revision: 1
+    status: failed
+    updated: "-1h"
+steps:
+  - op: history
+    release: fresh
+    history:
+      - status: pending-install
+        updated: "-3h"
+  - op: uninstall
+    release: fresh
+  - op: get
+    release: fresh
+    absent: true
+  - op: get
+    release: broke
+    expect:
+      status: failed
+      updated: "-1h"
+  - op: history
+    release: broke
+    history:
+      - revision: 1
+        status: failed
+  - op: history
+    release: absent
+    absent: true
diff --git a/internal/helm/testdata/client/rollback-pending-rollback/rollback-pending-rollback.yaml b/internal/helm/testdata/client/rollback-pending-rollback/rollback-pending-rollback.yaml
new file mode 100644
index 0000000..2e6327f
--- /dev/null
+++ b/internal/helm/testdata/client/rollback-pending-rollback/rollback-pending-rollback.yaml
@@ -0,0 +1,26 @@
+description: rollback of a pending rollback
+namespace: default
+seed:
+  - name: back
+    revision: 1
+    status: deployed
+    updated: "-2h"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: back
+  - name: back
+    revision: 2
+    status: pending-rollback
+    updated: "-30m"
+    labels:
+      serving.modelsphere.dev/llmservice-uid: uid-1
+      serving.modelsphere.dev/llmservice: back
+steps:
+  - op: rollback
+    release: back
+    revision: 1
+  - op: get
+    release: back
+    expect:
+      revision: 3
+      status: deployed
diff --git a/internal/helm/testdata/guard/apply/hook.yaml b/internal/helm/testdata/guard/apply/hook.yaml
new file mode 100644
index 0000000..aca78ab
--- /dev/null
+++ b/internal/helm/testdata/guard/apply/hook.yaml
@@ -0,0 +1,14 @@
+description: hook install refused
+steps:
+  - write: |
+      apiVersion: v1
+      kind: ConfigMap
+      metadata:
+        name: escape
+        namespace: kube-system
+        annotations:
+          helm.sh/hook: pre-install,pre-upgrade
+  - op: install
+    errorContains: kube-system
+  - op: get
+    absent: true
diff --git a/internal/helm/testdata/guard/apply/template.yaml b/internal/helm/testdata/guard/apply/template.yaml
new file mode 100644
index 0000000..2a06b43
--- /dev/null
+++ b/internal/helm/testdata/guard/apply/template.yaml
@@ -0,0 +1,12 @@
+description: template install refused
+steps:
+  - write: |
+      apiVersion: v1
+      kind: ConfigMap
+      metadata:
+        name: escape
+        namespace: kube-system
+  - op: install
+    errorContains: kube-system
+  - op: get
+    absent: true
diff --git a/internal/helm/testdata/guard/apply/upgrade.yaml b/internal/helm/testdata/guard/apply/upgrade.yaml
new file mode 100644
index 0000000..d7ff57a
--- /dev/null
+++ b/internal/helm/testdata/guard/apply/upgrade.yaml
@@ -0,0 +1,10 @@
+description: upgrade refused when the chart becomes cluster-scoped
+steps:
+  - op: install
+  - write: |
+      apiVersion: rbac.authorization.k8s.io/v1
+      kind: ClusterRole
+      metadata:
+        name: escape
+  - op: upgrade
+    errorContains: cluster-scoped
diff --git a/internal/helm/testdata/guard/namespace/after-a-good-one.yaml b/internal/helm/testdata/guard/namespace/after-a-good-one.yaml
new file mode 100644
index 0000000..8c938ab
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/after-a-good-one.yaml
@@ -0,0 +1,12 @@
+description: after a good one
+manifest: |
+  apiVersion: v1
+  kind: ConfigMap
+  metadata:
+    name: a
+  ---
+  apiVersion: rbac.authorization.k8s.io/v1
+  kind: ClusterRole
+  metadata:
+    name: b
+refused: cluster-scoped
diff --git a/internal/helm/testdata/guard/namespace/another-namespace.yaml b/internal/helm/testdata/guard/namespace/another-namespace.yaml
new file mode 100644
index 0000000..c82c7f1
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/another-namespace.yaml
@@ -0,0 +1,8 @@
+description: another namespace
+manifest: |
+  apiVersion: v1
+  kind: ConfigMap
+  metadata:
+    name: a
+    namespace: kube-system
+refused: into namespace kube-system
diff --git a/internal/helm/testdata/guard/namespace/cluster-scoped.yaml b/internal/helm/testdata/guard/namespace/cluster-scoped.yaml
new file mode 100644
index 0000000..373d6dc
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/cluster-scoped.yaml
@@ -0,0 +1,7 @@
+description: cluster-scoped
+manifest: |
+  apiVersion: rbac.authorization.k8s.io/v1
+  kind: ClusterRole
+  metadata:
+    name: a
+refused: 'ClusterRole "a", which is cluster-scoped'
diff --git a/internal/helm/testdata/guard/namespace/comments-only.yaml b/internal/helm/testdata/guard/namespace/comments-only.yaml
new file mode 100644
index 0000000..00eee77
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/comments-only.yaml
@@ -0,0 +1,5 @@
+description: comments only
+manifest: |
+  # Source: x/templates/empty.yaml
+  ---
+refused: ""
diff --git a/internal/helm/testdata/guard/namespace/inside-a-list.yaml b/internal/helm/testdata/guard/namespace/inside-a-list.yaml
new file mode 100644
index 0000000..1360362
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/inside-a-list.yaml
@@ -0,0 +1,15 @@
+description: inside a list
+manifest: |
+  apiVersion: v1
+  kind: List
+  items:
+  - apiVersion: v1
+    kind: ConfigMap
+    metadata:
+      name: a
+  - apiVersion: rbac.authorization.k8s.io/v1
+    kind: Role
+    metadata:
+      name: b
+      namespace: kube-system
+refused: 'Role "b" into namespace kube-system'
diff --git a/internal/helm/testdata/guard/namespace/namespace-object.yaml b/internal/helm/testdata/guard/namespace/namespace-object.yaml
new file mode 100644
index 0000000..741d335
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/namespace-object.yaml
@@ -0,0 +1,7 @@
+description: namespace object
+manifest: |
+  apiVersion: v1
+  kind: Namespace
+  metadata:
+    name: a
+refused: cluster-scoped
diff --git a/internal/helm/testdata/guard/namespace/no-kind.yaml b/internal/helm/testdata/guard/namespace/no-kind.yaml
new file mode 100644
index 0000000..5a2f967
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/no-kind.yaml
@@ -0,0 +1,4 @@
+description: no kind
+manifest: |
+  replicaCount: 1
+refused: an object with no kind
diff --git a/internal/helm/testdata/guard/namespace/no-namespace.yaml b/internal/helm/testdata/guard/namespace/no-namespace.yaml
new file mode 100644
index 0000000..772e290
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/no-namespace.yaml
@@ -0,0 +1,7 @@
+description: no namespace
+manifest: |
+  apiVersion: v1
+  kind: ConfigMap
+  metadata:
+    name: a
+refused: ""
diff --git a/internal/helm/testdata/guard/namespace/own-namespace.yaml b/internal/helm/testdata/guard/namespace/own-namespace.yaml
new file mode 100644
index 0000000..9022b29
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/own-namespace.yaml
@@ -0,0 +1,8 @@
+description: own namespace
+manifest: |
+  apiVersion: v1
+  kind: ConfigMap
+  metadata:
+    name: a
+    namespace: models
+refused: ""
diff --git a/internal/helm/testdata/guard/namespace/unknown-kind.yaml b/internal/helm/testdata/guard/namespace/unknown-kind.yaml
new file mode 100644
index 0000000..9b0d1fe
--- /dev/null
+++ b/internal/helm/testdata/guard/namespace/unknown-kind.yaml
@@ -0,0 +1,7 @@
+description: unknown kind
+manifest: |
+  apiVersion: example.com/v1
+  kind: Widget
+  metadata:
+    name: a
+refused: cannot be told
diff --git a/internal/testutil/cases/cases.go b/internal/testutil/cases/cases.go
new file mode 100644
index 0000000..038dd7a
--- /dev/null
+++ b/internal/testutil/cases/cases.go
@@ -0,0 +1,236 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+// Package cases loads one YAML document per test case.
+//
+// Files are discovered recursively under a directory. Only names ending in
+// ".yaml" are cases; anything else (a README, a note) is ignored. The result
+// is ordered by the slash-separated path relative to that directory, so a
+// rename or a new file changes order only by that path.
+//
+// Each file is decoded with sigs.k8s.io/yaml in strict mode: an unknown field,
+// a duplicate key, or a value of the wrong type fails the load. An empty
+// directory, a missing directory, and an unreadable file fail as well.
+//
+// A decoded case may carry Name and Description. Name, when set, must equal the
+// file stem (the base name without ".yaml"). Description is the sentence
+// runners use as the subtest or Ginkgo entry text; when it is empty, the
+// stem is used instead.
+//
+// Durations in a case are Go duration strings ("30s", "1h", "-2h") decoded
+// into Duration. A nil pointer is YAML null or an omitted field; a pointer
+// to zero is the string "0s".
+//
+// Keys are decoded with YAML 1.1 rules. An unquoted n, y, yes, no, on, or off
+// is a boolean, not a string, and strict decoding then reports it as an
+// unknown field. Quote any such key.
+package cases
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"reflect"
+	"slices"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/onsi/ginkgo/v2"
+	"sigs.k8s.io/yaml"
+)
+
+// Duration is a time.Duration stored in YAML as a Go duration string.
+type Duration time.Duration
+
+// Std returns the duration as a time.Duration.
+func (d Duration) Std() time.Duration { return time.Duration(d) }
+
+// UnmarshalJSON accepts a JSON string in Go duration syntax.
+func (d *Duration) UnmarshalJSON(b []byte) error {
+	var s string
+	if err := json.Unmarshal(b, &s); err != nil {
+		return fmt.Errorf("duration: %w", err)
+	}
+	parsed, err := time.ParseDuration(s)
+	if err != nil {
+		return fmt.Errorf("duration %q: %w", s, err)
+	}
+	*d = Duration(parsed)
+	return nil
+}
+
+// File is one decoded case and where it was read from.
+type File[T any] struct {
+	// Rel is the path relative to the loaded directory, with forward slashes.
+	Rel string
+	// Stem is the file name without the .yaml suffix.
+	Stem string
+	// Case is the decoded document.
+	Case T
+}
+
+// Load reads and decodes every case file under dir.
+func Load[T any](dir string) ([]T, error) {
+	files, err := LoadFiles[T](dir)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]T, len(files))
+	for i, f := range files {
+		out[i] = f.Case
+	}
+	return out, nil
+}
+
+// LoadFiles is Load with the relative path and stem of each file.
+func LoadFiles[T any](dir string) ([]File[T], error) {
+	rels, err := yamlRels(dir)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]File[T], 0, len(rels))
+	for _, rel := range rels {
+		path := filepath.Join(dir, filepath.FromSlash(rel))
+		body, err := os.ReadFile(path)
+		if err != nil {
+			return nil, fmt.Errorf("cases: read %s: %w", rel, err)
+		}
+		var tc T
+		if err := yaml.UnmarshalStrict(body, &tc); err != nil {
+			return nil, fmt.Errorf("cases: decode %s: %w", rel, err)
+		}
+		stem := strings.TrimSuffix(filepath.Base(rel), ".yaml")
+		if err := checkName(tc, stem, rel); err != nil {
+			return nil, err
+		}
+		out = append(out, File[T]{Rel: rel, Stem: stem, Case: tc})
+	}
+	return out, nil
+}
+
+// MustLoad is LoadFiles for a Ginkgo container, which has no testing.T yet.
+// A bad directory panics so the suite fails instead of running zero cases.
+func MustLoad[T any](dir string) []File[T] {
+	files, err := LoadFiles[T](dir)
+	if err != nil {
+		panic(err)
+	}
+	return files
+}
+
+// Run loads dir and calls fn as one subtest per file.
+// The subtest name is the case Description when that field is set, and the
+// file stem otherwise, so an existing sentence stays the name go test prints.
+func Run[T any](t *testing.T, dir string, fn func(t *testing.T, tc T)) {
+	t.Helper()
+	files, err := LoadFiles[T](dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, f := range files {
+		t.Run(SpecName(f.Case, f.Stem), func(t *testing.T) {
+			fn(t, f.Case)
+		})
+	}
+}
+
+// Entries loads dir as Ginkgo table entries, one Entry per file.
+// description supplies the entry text; a nil description, or a result of "",
+// uses the file stem. A load error panics so suite setup fails loudly.
+func Entries[T any](dir string, description func(T) string) []ginkgo.TableEntry {
+	files, err := LoadFiles[T](dir)
+	if err != nil {
+		panic(err)
+	}
+	entries := make([]ginkgo.TableEntry, len(files))
+	for i, f := range files {
+		text := f.Stem
+		if description != nil {
+			if d := description(f.Case); d != "" {
+				text = d
+			}
+		}
+		entries[i] = ginkgo.Entry(text, f.Case)
+	}
+	return entries
+}
+
+// SpecName is the subtest or entry text for a case: its Description, or stem.
+func SpecName[T any](tc T, stem string) string {
+	if d, ok := stringField(tc, "Description"); ok && d != "" {
+		return d
+	}
+	return stem
+}
+
+func yamlRels(dir string) ([]string, error) {
+	var rels []string
+	err := filepath.WalkDir(dir, func(path string, d os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.IsDir() || !strings.HasSuffix(d.Name(), ".yaml") {
+			return nil
+		}
+		rel, err := filepath.Rel(dir, path)
+		if err != nil {
+			return err
+		}
+		rels = append(rels, filepath.ToSlash(rel))
+		return nil
+	})
+	if err != nil {
+		return nil, fmt.Errorf("cases: read %s: %w", dir, err)
+	}
+	if len(rels) == 0 {
+		return nil, fmt.Errorf("cases: %s has no yaml files", dir)
+	}
+	slices.Sort(rels)
+	return rels, nil
+}
+
+func checkName(tc any, stem, rel string) error {
+	name, ok := stringField(tc, "Name")
+	if !ok || name == "" {
+		return nil
+	}
+	if name != stem {
+		return fmt.Errorf("cases: %s: name %q must equal the file stem %q", rel, name, stem)
+	}
+	return nil
+}
+
+// stringField reads an exported string field. ok is false when the case has
+// no such field, so a case type is not required to carry Name or Description.
+func stringField(tc any, field string) (string, bool) {
+	v := reflect.ValueOf(tc)
+	if v.Kind() == reflect.Pointer {
+		if v.IsNil() {
+			return "", false
+		}
+		v = v.Elem()
+	}
+	if v.Kind() != reflect.Struct {
+		return "", false
+	}
+	f := v.FieldByName(field)
+	if !f.IsValid() || f.Kind() != reflect.String || !f.CanInterface() {
+		return "", false
+	}
+	return f.String(), true
+}
diff --git a/internal/testutil/cases/cases_test.go b/internal/testutil/cases/cases_test.go
new file mode 100644
index 0000000..75c9cf9
--- /dev/null
+++ b/internal/testutil/cases/cases_test.go
@@ -0,0 +1,152 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package cases
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// sample is the document shape under testdata/. The happy-path files are the
+// cases; the other directories are documents this loader must reject.
+type sample struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Value int       `json:"value"`
+		Wait  Duration  `json:"wait,omitempty"`
+		Maybe *Duration `json:"maybe,omitempty"`
+		Empty *Duration `json:"empty,omitempty"`
+	} `json:"inputs"`
+	Expected struct {
+		Value int `json:"value"`
+	} `json:"expected"`
+}
+
+func TestLoadOrdersByRelativePath(t *testing.T) {
+	files, err := LoadFiles[sample]("testdata/ok")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(files) != 2 {
+		t.Fatalf("got %d files", len(files))
+	}
+	if files[0].Rel != "a/c.yaml" || files[1].Rel != "b.yaml" {
+		t.Fatalf("order = %q, %q", files[0].Rel, files[1].Rel)
+	}
+	if files[0].Case.Inputs.Value != 1 || files[1].Case.Expected.Value != 2 {
+		t.Fatalf("decoded %+v %+v", files[0].Case, files[1].Case)
+	}
+	if files[0].Case.Inputs.Wait.Std() != 30*time.Second {
+		t.Fatalf("wait = %s", files[0].Case.Inputs.Wait.Std())
+	}
+	if files[0].Case.Inputs.Maybe == nil || files[0].Case.Inputs.Maybe.Std() != time.Second {
+		t.Fatalf("maybe = %v", files[0].Case.Inputs.Maybe)
+	}
+	if files[0].Case.Inputs.Empty != nil {
+		t.Fatal("null duration decoded as set")
+	}
+
+	got, err := Load[sample]("testdata/ok")
+	if err != nil || len(got) != 2 || got[0].Description != "first" {
+		t.Fatalf("Load = %+v, %v", got, err)
+	}
+}
+
+func TestRunUsesDescription(t *testing.T) {
+	var seen []string
+	Run(t, "testdata/ok", func(t *testing.T, tc sample) {
+		seen = append(seen, tc.Description)
+		if !strings.HasSuffix(t.Name(), "/"+tc.Description) {
+			t.Errorf("subtest %q does not end with %q", t.Name(), tc.Description)
+		}
+	})
+	if strings.Join(seen, ",") != "first,second" {
+		t.Fatalf("ran %v", seen)
+	}
+}
+
+func TestEntriesFollowsDescription(t *testing.T) {
+	entries := Entries("testdata/ok", func(s sample) string { return s.Description })
+	if len(entries) != 2 {
+		t.Fatalf("got %d entries", len(entries))
+	}
+	entries = Entries[sample]("testdata/ok", nil)
+	if len(entries) != 2 {
+		t.Fatalf("nil description: got %d entries", len(entries))
+	}
+}
+
+func TestRejectsBadDocuments(t *testing.T) {
+	checks := []struct {
+		dir  string
+		want string
+	}{
+		{"testdata/unknown-field", "unknown field"},
+		{"testdata/name-mismatch", "must equal the file stem"},
+		{"testdata/bad-yaml", "decode"},
+		{"testdata/no-yaml", "has no yaml files"},
+		{"testdata/missing", "read"},
+	}
+	for _, tc := range checks {
+		t.Run(tc.dir, func(t *testing.T) {
+			_, err := Load[sample](tc.dir)
+			if err == nil || !strings.Contains(err.Error(), tc.want) {
+				t.Fatalf("got %v, want it to contain %q", err, tc.want)
+			}
+		})
+	}
+}
+
+// An empty directory cannot be stored, so this one failure mode is built in a
+// temp dir. The other failures live as files under testdata/.
+func TestEmptyDirFails(t *testing.T) {
+	dir := t.TempDir()
+	if _, err := Load[sample](dir); err == nil || !strings.Contains(err.Error(), "has no yaml files") {
+		t.Fatalf("got %v", err)
+	}
+}
+
+func TestMustLoadPanics(t *testing.T) {
+	defer func() {
+		if recover() == nil {
+			t.Fatal("MustLoad returned")
+		}
+	}()
+	MustLoad[sample](filepath.Join(t.TempDir(), "absent"))
+}
+
+func TestUnreadableFileFails(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "gone.yaml")
+	if err := os.WriteFile(path, []byte("name: gone\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Chmod(path, 0); err != nil {
+		t.Fatal(err)
+	}
+	// Root ignores mode bits. Skip when the file is still readable.
+	if _, err := os.ReadFile(path); err == nil {
+		t.Skip("process can read a mode-0 file")
+	}
+	if _, err := Load[sample](dir); err == nil || !strings.Contains(err.Error(), "read") {
+		t.Fatalf("got %v", err)
+	}
+}
diff --git a/internal/testutil/cases/testdata/bad-yaml/x.yaml b/internal/testutil/cases/testdata/bad-yaml/x.yaml
new file mode 100644
index 0000000..89a1e6b
--- /dev/null
+++ b/internal/testutil/cases/testdata/bad-yaml/x.yaml
@@ -0,0 +1 @@
+name: [unterminated
diff --git a/internal/testutil/cases/testdata/name-mismatch/nope.yaml b/internal/testutil/cases/testdata/name-mismatch/nope.yaml
new file mode 100644
index 0000000..2121d0e
--- /dev/null
+++ b/internal/testutil/cases/testdata/name-mismatch/nope.yaml
@@ -0,0 +1,6 @@
+name: other
+description: name does not match the file
+inputs:
+  value: 1
+expected:
+  value: 1
diff --git a/internal/testutil/cases/testdata/no-yaml/note.txt b/internal/testutil/cases/testdata/no-yaml/note.txt
new file mode 100644
index 0000000..0622888
--- /dev/null
+++ b/internal/testutil/cases/testdata/no-yaml/note.txt
@@ -0,0 +1 @@
+not a case
diff --git a/internal/testutil/cases/testdata/ok/a/c.yaml b/internal/testutil/cases/testdata/ok/a/c.yaml
new file mode 100644
index 0000000..9269483
--- /dev/null
+++ b/internal/testutil/cases/testdata/ok/a/c.yaml
@@ -0,0 +1,9 @@
+name: c
+description: first
+inputs:
+  value: 1
+  wait: "30s"
+  maybe: "1s"
+  empty: null
+expected:
+  value: 1
diff --git a/internal/testutil/cases/testdata/ok/b.yaml b/internal/testutil/cases/testdata/ok/b.yaml
new file mode 100644
index 0000000..ee3db7b
--- /dev/null
+++ b/internal/testutil/cases/testdata/ok/b.yaml
@@ -0,0 +1,6 @@
+name: b
+description: second
+inputs:
+  value: 2
+expected:
+  value: 2
diff --git a/internal/testutil/cases/testdata/unknown-field/bad.yaml b/internal/testutil/cases/testdata/unknown-field/bad.yaml
new file mode 100644
index 0000000..c0a7637
--- /dev/null
+++ b/internal/testutil/cases/testdata/unknown-field/bad.yaml
@@ -0,0 +1,7 @@
+name: bad
+description: unknown field
+inputs:
+  value: 1
+expected:
+  value: 1
+  extra: true

From 6391612c0cfaad85f2d0fb6a2576ce325fec5f91 Mon Sep 17 00:00:00 2001
From: huskcass <333684917+huskcass@users.noreply.github.com>
Date: Thu, 22 Oct 2026 06:00:50 -0700
Subject: [PATCH 2/2] test: reject case files that are not exactly one mapping

The case loader now requires exactly one non-null YAML mapping per
file. An empty, comment-only, null or "---"-only file, and a file with
more than one document, fail the load and name the file. A .yml file
fails the load instead of being skipped.

Keys must match the json name exactly; encoding/json alone matched
them case-insensitively. Types that decode themselves (durations,
quantities and the like) are checked as leaves.

The Ginkgo entry helper moves to cases/ginkgocases so the core loader
no longer pulls Ginkgo flags into the chart and helm test binaries.

Runners no longer treat an omitted field as "anything passes":
decision server cases need a status, error cases need errorContains,
chart repo cases set allowed (or nilAllowlist) explicitly, and an
empty entry in any substring list fails the case. Condition message
and observedGeneration are asserted when set. The unused
"capture: generation" step and update labels are removed, and the
duration docs now match what the loader accepts.
---
 DEVELOPMENT.md                                |   9 +-
 internal/chart/repo_test.go                   |  44 +++--
 .../autoscaling/customprovider_test.go        |  22 ++-
 .../autoscaling/llmscaler_casefile_test.go    |   4 +
 .../autoscaling/llmscaler_controller_test.go  |  19 +--
 .../autoscaling/llmscaler_helpers_test.go     |   4 +
 .../autoscaling/llmscaler_scenario_test.go    |  10 ++
 .../a-body-that-is-not-json-is-an-error.yaml  |   1 +
 ...-int32-is-refused-rather-than-wrapped.yaml |   1 +
 ...on-for-another-service-does-not-match.yaml |   1 +
 ...on-with-no-active-count-is-not-a-zero.yaml |   1 +
 ...decision-resolves-to-its-active-count.yaml |   1 +
 .../a-negative-count-is-refused.yaml          |   1 +
 ...payload-with-no-apiversion-is-refused.yaml |   1 +
 .../an-empty-decision-list-is-not-a-zero.yaml |   1 +
 ...chema-is-refused-not-parsed-hopefully.yaml |   1 +
 ...licates-are-an-error-not-a-pick-first.yaml |   1 +
 ...lter-is-reported-when-nothing-matches.yaml |   1 +
 ...ter-picks-between-same-named-services.yaml |   1 +
 .../zero-is-a-decision-not-a-missing-one.yaml |   1 +
 ...ecision-server-payload-resolves-to-15.yaml |   1 +
 .../serving/llmservice_cases_test.go          |  33 ++--
 internal/controller/serving/repo_test.go      |  10 +-
 internal/testutil/cases/cases.go              | 146 ++++++++++++-----
 internal/testutil/cases/cases_test.go         | 150 ++++++++++++++++--
 .../testutil/cases/ginkgocases/entries.go     |  49 ++++++
 .../cases/ginkgocases/entries_test.go         |  49 ++++++
 internal/testutil/cases/keys.go               | 133 ++++++++++++++++
 28 files changed, 602 insertions(+), 94 deletions(-)
 create mode 100644 internal/testutil/cases/ginkgocases/entries.go
 create mode 100644 internal/testutil/cases/ginkgocases/entries_test.go
 create mode 100644 internal/testutil/cases/keys.go

diff --git a/DEVELOPMENT.md b/DEVELOPMENT.md
index 103fa9a..354166d 100644
--- a/DEVELOPMENT.md
+++ b/DEVELOPMENT.md
@@ -147,13 +147,16 @@ internal/<pkg>/testdata/<suite>/<case-name>.yaml
 
 The suite directory groups a table (`guards/hold`, `controller/metric-scaling`). The file name is the case name in kebab-case. A document has:
 
+- exactly one YAML mapping. An empty file, a comment-only file, `null`, a bare `---`, or a second document fails the load. A trailing `---` counts as a second document.
 - `name`, optional. When set it must equal the file name without `.yaml`.
 - `description`, the sentence `go test -v` and Ginkgo print. When it is empty, the file name is used.
-- `inputs` and `expected`, decoded into a Go struct. An unknown field, a duplicate key, or a value of the wrong type fails the load.
+- `inputs` and `expected`, decoded into a Go struct. An unknown field, a duplicate key, a key in the wrong case, or a value of the wrong type fails the load. Mapping keys must match the json name of the field, including case.
 
-Durations are Go duration strings (`30s`, `1m1s`, `-2h`). A nil pointer is an omitted field or `null`. A pointer to zero is `0` or `"0s"`, not an omission. Times relative to a fixed origin are offsets from that origin; the runner adds them.
+Only `*.yaml` is loaded. A `*.yml` file fails the load, so two extensions cannot share a stem. Other names are ignored. The suffix check is case-sensitive.
 
-The loader is `internal/testutil/cases`. It reads every `*.yaml` under the directory, in path order, and rejects an empty directory. `cases.Run` starts one subtest per file. Ginkgo tables use `cases.Entries`, one entry per file.
+Durations are Go duration strings (`30s`, `1m1s`, `-2h`, `0s`). Only a pointer to a duration may be omitted or `null`. A non-pointer duration rejects `null` and treats an omission as zero. A pointer to zero is `"0s"`, not a bare `0`. Times relative to a fixed origin are offsets from that origin; the runner adds them.
+
+The loader is `internal/testutil/cases`. It reads every `*.yaml` under the directory, in path order, and rejects an empty directory. `cases.Run` starts one subtest per file. Ginkgo tables use `ginkgocases.Entries` in `internal/testutil/cases/ginkgocases`, so a test binary that does not import that package does not register Ginkgo flags.
 
 To add a case, copy a nearby file, change the inputs and the expected result, and run `make test`.
 
diff --git a/internal/chart/repo_test.go b/internal/chart/repo_test.go
index b2d658e..3f75bbc 100644
--- a/internal/chart/repo_test.go
+++ b/internal/chart/repo_test.go
@@ -200,23 +200,43 @@ func runChartCall(t *testing.T, srv *httptest.Server, serverURL, basic string, c
 				t.Errorf("got %v", err)
 			}
 		}
-		for _, sub := range call.ErrorContains {
-			want := subst(sub, serverURL, repo, indexURL, callBasic)
-			if err == nil || !strings.Contains(err.Error(), want) {
-				t.Errorf("%q does not say %q", err, want)
-			}
-		}
-		for _, sub := range call.ErrorNotContains {
-			want := subst(sub, serverURL, repo, indexURL, callBasic)
-			if err != nil && strings.Contains(err.Error(), want) {
-				t.Errorf("%q names the token endpoint rather than the repository", err)
-			}
-		}
+		checkContained(t, err, call.ErrorContains, serverURL, repo, indexURL, callBasic)
+		checkAbsent(t, err, call.ErrorNotContains, serverURL, repo, indexURL, callBasic)
 	default:
 		t.Fatal("call has no expectation")
 	}
 }
 
+// checkContained fails when err does not include every substring.
+// An empty substring matches every error, so it is rejected on its own.
+func checkContained(t *testing.T, err error, subs []string, serverURL, repo, indexURL, callBasic string) {
+	t.Helper()
+	for _, sub := range subs {
+		if sub == "" {
+			t.Fatal("errorContains entry is empty")
+		}
+		want := subst(sub, serverURL, repo, indexURL, callBasic)
+		if err == nil || !strings.Contains(err.Error(), want) {
+			t.Errorf("%q does not say %q", err, want)
+		}
+	}
+}
+
+// checkAbsent fails when err includes a substring that must stay out of it.
+// An empty substring matches every error, so it is rejected on its own.
+func checkAbsent(t *testing.T, err error, subs []string, serverURL, repo, indexURL, callBasic string) {
+	t.Helper()
+	for _, sub := range subs {
+		if sub == "" {
+			t.Fatal("errorNotContains entry is empty")
+		}
+		want := subst(sub, serverURL, repo, indexURL, callBasic)
+		if err != nil && strings.Contains(err.Error(), want) {
+			t.Errorf("%q names the token endpoint rather than the repository", err)
+		}
+	}
+}
+
 func (call chartCall) source(t *testing.T, srv *httptest.Server) (repo, path string) {
 	t.Helper()
 	switch call.RepoKind {
diff --git a/internal/controller/autoscaling/customprovider_test.go b/internal/controller/autoscaling/customprovider_test.go
index 5b93e59..fc16894 100644
--- a/internal/controller/autoscaling/customprovider_test.go
+++ b/internal/controller/autoscaling/customprovider_test.go
@@ -71,17 +71,19 @@ func TestQueryDecisionReplicasSamplePayload(t *testing.T) {
 
 func assertDecision(t *testing.T, tc decisionFile) {
 	t.Helper()
-	status := tc.Inputs.Status
-	if status == 0 {
-		status = http.StatusOK
+	// Zero is an omitted status, not HTTP 200. WriteHeader panics on it.
+	if tc.Inputs.Status < 100 || tc.Inputs.Status > 999 {
+		t.Fatalf("status is required, got %d", tc.Inputs.Status)
 	}
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-		w.WriteHeader(status)
+		w.WriteHeader(tc.Inputs.Status)
 		_, _ = w.Write([]byte(tc.Inputs.Body))
 	}))
 	defer srv.Close()
 
 	got, err := queryDecisionReplicas(srv.URL, "", tc.Inputs.ServiceID, tc.Inputs.Namespace, nil)
+	// An empty errorContains means the call succeeds. A case that expects a
+	// failure names the substring; an empty string would match every error.
 	if tc.Expected.ErrorContains == "" {
 		if err != nil {
 			t.Fatalf("unexpected error: %v", err)
@@ -171,6 +173,14 @@ type decisionNon200File struct {
 // when its body happens to be a well-formed decision payload.
 func TestQueryDecisionReplicasNon200(t *testing.T) {
 	cases.Run(t, "testdata/helpers/decision-non-200", func(t *testing.T, tc decisionNon200File) {
+		// Zero is an omitted status, and WriteHeader panics on it.
+		if tc.Inputs.Status < 100 || tc.Inputs.Status > 999 {
+			t.Fatalf("status is required, got %d", tc.Inputs.Status)
+		}
+		// An empty substring matches every error.
+		if tc.Expected.ErrorContains == "" {
+			t.Fatal("errorContains is required")
+		}
 		srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
 			w.WriteHeader(tc.Inputs.Status)
 			_, _ = w.Write([]byte(tc.Inputs.Body))
@@ -201,6 +211,10 @@ type decisionMisconfigFile struct {
 // error names the field rather than the URL.
 func TestQueryDecisionReplicasMisconfigured(t *testing.T) {
 	cases.Run(t, "testdata/helpers/decision-misconfigured", func(t *testing.T, tc decisionMisconfigFile) {
+		// An empty substring matches every error.
+		if tc.Expected.ErrorContains == "" {
+			t.Fatal("errorContains is required")
+		}
 		_, err := queryDecisionReplicas(tc.Inputs.ServerAddress, "", tc.Inputs.ServiceID, "", nil)
 		if err == nil || !strings.Contains(err.Error(), tc.Expected.ErrorContains) {
 			t.Errorf("error = %v, want it to contain %q", err, tc.Expected.ErrorContains)
diff --git a/internal/controller/autoscaling/llmscaler_casefile_test.go b/internal/controller/autoscaling/llmscaler_casefile_test.go
index ffcccc8..65c7d22 100644
--- a/internal/controller/autoscaling/llmscaler_casefile_test.go
+++ b/internal/controller/autoscaling/llmscaler_casefile_test.go
@@ -322,6 +322,10 @@ func runCustomAdmissionFile(c customAdmissionFile) {
 	if c.Inputs.PinProvider && c.Inputs.MetricProvider != autoscalingv1alpha1.MetricProviderPrometheus {
 		Fail(fmt.Sprintf("metricProvider %q, want %q", c.Inputs.MetricProvider, autoscalingv1alpha1.MetricProviderPrometheus))
 	}
+	// An empty substring matches every error.
+	if c.Expected.ErrorContains == "" {
+		Fail(c.Description + ": errorContains is required")
+	}
 	runCustomProviderAdmission(customAdmissionCase{
 		metricProvider:     c.Inputs.MetricProvider,
 		omitCustomProvider: c.Inputs.OmitCustomProvider,
diff --git a/internal/controller/autoscaling/llmscaler_controller_test.go b/internal/controller/autoscaling/llmscaler_controller_test.go
index f87395f..fc0d495 100644
--- a/internal/controller/autoscaling/llmscaler_controller_test.go
+++ b/internal/controller/autoscaling/llmscaler_controller_test.go
@@ -20,6 +20,7 @@ import (
 	. "github.com/onsi/ginkgo/v2"
 
 	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+	"github.com/modelsphere/llm-operator/internal/testutil/cases/ginkgocases"
 )
 
 // The reconcile specs are one YAML file each under testdata/controller.
@@ -37,19 +38,19 @@ var _ = Describe("LLMScaler Controller", func() {
 		})
 
 		DescribeTable("metric scaling", runScalerCase,
-			cases.Entries("testdata/controller/metric-scaling", scalerDesc))
+			ginkgocases.Entries("testdata/controller/metric-scaling", scalerDesc))
 		DescribeTable("rollout guard (G8)", runScalerCase,
-			cases.Entries("testdata/controller/rollout-guard", scalerDesc))
+			ginkgocases.Entries("testdata/controller/rollout-guard", scalerDesc))
 		DescribeTable("settle guard (G7)", runScalerCase,
-			cases.Entries("testdata/controller/settle-guard", scalerDesc))
+			ginkgocases.Entries("testdata/controller/settle-guard", scalerDesc))
 		DescribeTable("stabilization and max step", runScalerCase,
-			cases.Entries("testdata/controller/stabilization-and-step", scalerDesc))
+			ginkgocases.Entries("testdata/controller/stabilization-and-step", scalerDesc))
 		DescribeTable("deferral timeout", runScalerCase,
-			cases.Entries("testdata/controller/deferral-timeout", scalerDesc))
+			ginkgocases.Entries("testdata/controller/deferral-timeout", scalerDesc))
 		DescribeTable("deferral status", runScalerCase,
-			cases.Entries("testdata/controller/deferral-status", scalerDesc))
+			ginkgocases.Entries("testdata/controller/deferral-status", scalerDesc))
 		DescribeTable("forced-write pacing", runScalerCase,
-			cases.Entries("testdata/controller/forced-write-pacing", scalerDesc))
+			ginkgocases.Entries("testdata/controller/forced-write-pacing", scalerDesc))
 
 		for _, f := range cases.MustLoad[deferDefaultFile]("testdata/controller/crd-default") {
 			tc := f.Case
@@ -69,9 +70,9 @@ var _ = Describe("LLMScaler Controller", func() {
 		})
 
 		DescribeTable("decision server", runCustomReconcileFile,
-			cases.Entries("testdata/controller/custom-reconcile", customReconcileDesc))
+			ginkgocases.Entries("testdata/controller/custom-reconcile", customReconcileDesc))
 		DescribeTable("admission", runCustomAdmissionFile,
-			cases.Entries("testdata/controller/custom-admission", customAdmissionDesc))
+			ginkgocases.Entries("testdata/controller/custom-admission", customAdmissionDesc))
 
 		for _, f := range cases.MustLoad[customDefaultFile]("testdata/controller/custom-defaults") {
 			tc := f.Case
diff --git a/internal/controller/autoscaling/llmscaler_helpers_test.go b/internal/controller/autoscaling/llmscaler_helpers_test.go
index cf8f681..8f8916c 100644
--- a/internal/controller/autoscaling/llmscaler_helpers_test.go
+++ b/internal/controller/autoscaling/llmscaler_helpers_test.go
@@ -331,6 +331,10 @@ func TestQueryPrometheusScalar(t *testing.T) {
 			t.Fatalf("expected error containing %q, got value %v", tc.Expected.ErrorContains, got)
 		}
 		for _, want := range tc.Expected.ErrorContains {
+			// An empty substring matches every error.
+			if want == "" {
+				t.Fatal("errorContains entry is empty")
+			}
 			if !strings.Contains(err.Error(), want) {
 				t.Errorf("error = %q, want it to contain %q", err, want)
 			}
diff --git a/internal/controller/autoscaling/llmscaler_scenario_test.go b/internal/controller/autoscaling/llmscaler_scenario_test.go
index 554c257..df4fc37 100644
--- a/internal/controller/autoscaling/llmscaler_scenario_test.go
+++ b/internal/controller/autoscaling/llmscaler_scenario_test.go
@@ -627,7 +627,11 @@ func assertCondition(scaler *autoscalingv1alpha1.LLMScaler, exp condExpect) {
 	if exp.reason != "" {
 		Expect(cond.Reason).To(Equal(exp.reason))
 	}
+	// An empty substring matches every message, so it is rejected on its own.
 	for _, sub := range exp.messageContains {
+		if sub == "" {
+			Fail("messageContains entry is empty")
+		}
 		Expect(cond.Message).To(ContainSubstring(sub))
 	}
 	if exp.observedGenerationCurrent {
@@ -641,6 +645,12 @@ func assertEvents(run *scenarioRun, exp stepExpect) {
 	}
 	Expect(run.rec).NotTo(BeNil())
 	for _, want := range exp.events {
+		// An empty substring matches every event, so it is rejected on its own.
+		for _, sub := range want.contains {
+			if sub == "" {
+				Fail("events.contains entry is empty")
+			}
+		}
 		var got string
 		Expect(run.rec.Events).To(Receive(&got))
 		if want.prefix != "" {
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
index c60fdbe..d5db537 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-body-that-is-not-json-is-an-error.yaml
@@ -1,6 +1,7 @@
 "name": "a-body-that-is-not-json-is-an-error"
 "description": "a body that is not JSON is an error"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "<html>502 Bad Gateway</html>"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
index f570b8e..08bafea 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-count-past-int32-is-refused-rather-than-wrapped.yaml
@@ -1,6 +1,7 @@
 "name": "a-count-past-int32-is-refused-rather-than-wrapped"
 "description": "a count past int32 is refused rather than wrapped"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":4294967296}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
index 944223e..5284bd1 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-for-another-service-does-not-match.yaml
@@ -2,6 +2,7 @@
 "name": "a-decision-for-another-service-does-not-match"
 "description": "a decision for another service does not match"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"other\",\"replicas\":{\"active\":9}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
index a03b318..4dfeaec 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-decision-with-no-active-count-is-not-a-zero.yaml
@@ -1,6 +1,7 @@
 "name": "a-decision-with-no-active-count-is-not-a-zero"
 "description": "a decision with no active count is not a zero"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
index a4bb415..f886bff 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-matching-decision-resolves-to-its-active-count.yaml
@@ -1,6 +1,7 @@
 "name": "a-matching-decision-resolves-to-its-active-count"
 "description": "a matching decision resolves to its active count"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":15}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
index 18664d7..2b1ebf3 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-negative-count-is-refused.yaml
@@ -1,6 +1,7 @@
 "name": "a-negative-count-is-refused"
 "description": "a negative count is refused"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":-1}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
index ec878c0..fb9855b 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/a-payload-with-no-apiversion-is-refused.yaml
@@ -1,6 +1,7 @@
 "name": "a-payload-with-no-apiversion-is-refused"
 "description": "a payload with no apiVersion is refused"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":3}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
index de280d4..826f0d7 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-empty-decision-list-is-not-a-zero.yaml
@@ -1,6 +1,7 @@
 "name": "an-empty-decision-list-is-not-a-zero"
 "description": "an empty decision list is not a zero"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
index 49c66da..ecd39ec 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/an-unknown-schema-is-refused-not-parsed-hopefully.yaml
@@ -1,6 +1,7 @@
 "name": "an-unknown-schema-is-refused-not-parsed-hopefully"
 "description": "an unknown schema is refused, not parsed hopefully"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v2\",\"decisions\":[{\"serviceId\":\"svc\",\"replicas\":{\"active\":3}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
index bbacfff..e2c6abc 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/duplicates-are-an-error-not-a-pick-first.yaml
@@ -1,6 +1,7 @@
 "name": "duplicates-are-an-error-not-a-pick-first"
 "description": "duplicates are an error, not a pick-first"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}},{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":7}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
index 51f4233..1b1b354 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-is-reported-when-nothing-matches.yaml
@@ -1,6 +1,7 @@
 "name": "the-namespace-filter-is-reported-when-nothing-matches"
 "description": "the namespace filter is reported when nothing matches"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "namespace": "modelforge"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}}]}"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
index 8837b9c..6948158 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/the-namespace-filter-picks-between-same-named-services.yaml
@@ -1,6 +1,7 @@
 "name": "the-namespace-filter-picks-between-same-named-services"
 "description": "the namespace filter picks between same-named services"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "namespace": "modelforge"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"staging\",\"serviceId\":\"svc\",\"replicas\":{\"active\":2}},{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":7}}]}"
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml b/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
index 60ac9ac..c556e46 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-replicas/zero-is-a-decision-not-a-missing-one.yaml
@@ -1,6 +1,7 @@
 "name": "zero-is-a-decision-not-a-missing-one"
 "description": "zero is a decision, not a missing one"
 "inputs":
+  "status": 200
   "serviceId": "svc"
   "body": "{\"apiVersion\":\"llmscaling.inference.x-k8s.io/v1alpha1\",\"decisions\":[{\"namespace\":\"modelforge\",\"serviceId\":\"svc\",\"replicas\":{\"active\":0}}]}"
 "expected":
diff --git a/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml b/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
index 9ec7fe8..34d7f3a 100644
--- a/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
+++ b/internal/controller/autoscaling/testdata/helpers/decision-sample/a-verbatim-decision-server-payload-resolves-to-15.yaml
@@ -2,6 +2,7 @@
 "name": "a-verbatim-decision-server-payload-resolves-to-15"
 "description": "a verbatim decision-server payload resolves to 15"
 "inputs":
+  "status": 200
   "serviceId": "fallback-modelforge-01"
   "body": |-
     {
diff --git a/internal/controller/serving/llmservice_cases_test.go b/internal/controller/serving/llmservice_cases_test.go
index 69ccaf7..da7988c 100644
--- a/internal/controller/serving/llmservice_cases_test.go
+++ b/internal/controller/serving/llmservice_cases_test.go
@@ -57,7 +57,6 @@ type serviceStep struct {
 	Name           string            `json:"name,omitempty"`
 	Version        string            `json:"version,omitempty"`
 	Annotations    map[string]string `json:"annotations,omitempty"`
-	Labels         map[string]string `json:"labels,omitempty"`
 	Finalizers     []string          `json:"finalizers,omitempty"`
 	ChartVersion   *string           `json:"chartVersion,omitempty"`
 	Suspend        *bool             `json:"suspend,omitempty"`
@@ -214,7 +213,6 @@ type caseRun struct {
 	capturedRevision   string
 	appliedHash        string
 	observedGeneration int64
-	generation         int64
 	planted            string
 }
 
@@ -275,6 +273,10 @@ func (c *caseRun) step(i int, step serviceStep) {
 	case "setHelm":
 		c.setHelm(step)
 	case "setAllowed":
+		// An omitted list is nil, and a nil allowlist permits every repository.
+		if step.Repos == nil {
+			Fail(fmt.Sprintf("%s: step %d setAllowed requires repos", c.tc.Name, i+1))
+		}
 		c.r.AllowedChartRepos = append([]string(nil), step.Repos...)
 	case "deleteRelease":
 		delete(c.h.releases, c.h.slot(c.ns, c.name))
@@ -320,12 +322,6 @@ func (c *caseRun) update(step serviceStep) {
 		}
 		maps.Copy(obj.Annotations, step.Annotations)
 	}
-	if step.Labels != nil {
-		if obj.Labels == nil {
-			obj.Labels = map[string]string{}
-		}
-		maps.Copy(obj.Labels, step.Labels)
-	}
 	if step.ChartVersion != nil {
 		obj.Spec.Chart.Version = *step.ChartVersion
 	}
@@ -509,8 +505,6 @@ func (c *caseRun) capture(step serviceStep) {
 	obj := fetch(c.ns, c.name)
 	for _, what := range step.Capture {
 		switch what {
-		case "generation":
-			c.generation = obj.Generation
 		case "observedGeneration":
 			c.observedGeneration = obj.Status.ObservedGeneration
 		case "appliedHash":
@@ -707,8 +701,6 @@ func (c *caseRun) checkGeneration(exp *expectSpec, got *servingv1alpha1.LLMServi
 		Expect(got.Status.ObservedGeneration).To(Equal(c.observedGeneration))
 	case "nonzero":
 		Expect(got.Status.ObservedGeneration).NotTo(BeZero())
-	case "generation":
-		Expect(got.Generation).To(Equal(c.generation))
 	default:
 		Fail("unsupported observedGeneration " + exp.ObservedGeneration)
 	}
@@ -769,6 +761,16 @@ func (c *caseRun) checkConditions(conds []conditionSpec, got *servingv1alpha1.LL
 		if want.Reason != "" {
 			Expect(gotCond.Reason).To(Equal(want.Reason))
 		}
+		if want.Message != "" {
+			Expect(gotCond.Message).To(Equal(want.Message))
+		}
+		switch want.ObservedGeneration {
+		case "":
+		case "current":
+			Expect(gotCond.ObservedGeneration).To(Equal(got.Generation))
+		default:
+			Fail("unsupported observedGeneration " + want.ObservedGeneration)
+		}
 	}
 }
 
@@ -791,10 +793,17 @@ func (c *caseRun) checkMessage(exp *messageExpect, got *servingv1alpha1.LLMServi
 	if exp.Equals != "" {
 		Expect(got.Status.Message).To(Equal(exp.Equals))
 	}
+	// An empty substring matches every message, so it is rejected on its own.
 	for _, sub := range exp.Contains {
+		if sub == "" {
+			Fail("message.contains entry is empty")
+		}
 		Expect(got.Status.Message).To(ContainSubstring(sub))
 	}
 	for _, sub := range exp.NotContains {
+		if sub == "" {
+			Fail("message.notContains entry is empty")
+		}
 		Expect(got.Status.Message).NotTo(ContainSubstring(sub))
 	}
 }
diff --git a/internal/controller/serving/repo_test.go b/internal/controller/serving/repo_test.go
index 63567c1..e174238 100644
--- a/internal/controller/serving/repo_test.go
+++ b/internal/controller/serving/repo_test.go
@@ -6,8 +6,11 @@ import (
 	"github.com/modelsphere/llm-operator/internal/testutil/cases"
 )
 
-// repoCase is one chartRepoAllowed input. NilAllowlist passes a nil allowlist,
-// which permits every repository. Otherwise Allowed is the allowlist.
+// repoCase is one chartRepoAllowed input.
+// allowed is the allowlist and must be set, possibly as an empty list.
+// An empty list and a nil list both permit every repository; nilAllowlist
+// is how a case passes nil explicitly. Omitting both is an error, because
+// a forgotten allowlist would otherwise permit every repository.
 type repoCase struct {
 	Name         string   `json:"name,omitempty"`
 	Description  string   `json:"description,omitempty"`
@@ -19,6 +22,9 @@ type repoCase struct {
 
 func TestChartRepoAllowed(t *testing.T) {
 	cases.Run(t, "testdata/repo", func(t *testing.T, tc repoCase) {
+		if !tc.NilAllowlist && tc.Allowed == nil {
+			t.Fatal("set allowed (an empty list allows every repository) or nilAllowlist")
+		}
 		var allowed []string
 		if !tc.NilAllowlist {
 			allowed = tc.Allowed
diff --git a/internal/testutil/cases/cases.go b/internal/testutil/cases/cases.go
index 038dd7a..b90a81a 100644
--- a/internal/testutil/cases/cases.go
+++ b/internal/testutil/cases/cases.go
@@ -16,23 +16,37 @@ limitations under the License.
 
 // Package cases loads one YAML document per test case.
 //
-// Files are discovered recursively under a directory. Only names ending in
-// ".yaml" are cases; anything else (a README, a note) is ignored. The result
-// is ordered by the slash-separated path relative to that directory, so a
-// rename or a new file changes order only by that path.
+// Files are discovered recursively under a directory. A name ending in
+// ".yaml" is a case. A name ending in ".yml" is an error, so a second
+// extension cannot hide a case or make two files share a stem. Any other
+// name (a README, a note) is ignored. The suffix check is case-sensitive,
+// matching the ".yaml" files already in the tree. The result is ordered by
+// the slash-separated path relative to that directory, so a rename or a new
+// file changes order only by that path.
 //
-// Each file is decoded with sigs.k8s.io/yaml in strict mode: an unknown field,
-// a duplicate key, or a value of the wrong type fails the load. An empty
-// directory, a missing directory, and an unreadable file fail as well.
+// Each file must contain exactly one mapping. An empty file, a comment-only
+// file, a null document (`null`, `~`, or a bare `---`), or a second document
+// fails the load. A trailing `---` is a second, empty document and fails too.
+// The mapping is decoded with sigs.k8s.io/yaml in strict mode: an unknown
+// field, a duplicate key, or a value of the wrong type fails the load.
+// Mapping keys are case-sensitive. A key must equal the json name of a field
+// (or the field name when the struct has no json name). An anonymous embedded
+// struct contributes its fields, as encoding/json does. A type with
+// UnmarshalJSON, including Duration and Kubernetes types that decode
+// themselves, is a leaf; other structs, slices, maps, and pointers are
+// walked. An empty directory, a missing directory, and an unreadable file
+// fail as well.
 //
 // A decoded case may carry Name and Description. Name, when set, must equal the
 // file stem (the base name without ".yaml"). Description is the sentence
-// runners use as the subtest or Ginkgo entry text; when it is empty, the
-// stem is used instead.
+// runners use as the subtest text; when it is empty, the stem is used
+// instead. Ginkgo table entries are built by package ginkgocases, so this
+// package does not import Ginkgo.
 //
-// Durations in a case are Go duration strings ("30s", "1h", "-2h") decoded
-// into Duration. A nil pointer is YAML null or an omitted field; a pointer
-// to zero is the string "0s".
+// Durations in a case are Go duration strings ("30s", "1h", "-2h", "0s")
+// decoded into Duration. Only a pointer to Duration may be omitted or null.
+// A non-pointer Duration rejects null and treats an omission as zero. A
+// pointer to zero is the string "0s", not a bare 0.
 //
 // Keys are decoded with YAML 1.1 rules. An unquoted n, y, yes, no, on, or off
 // is a boolean, not a string, and strict decoding then reports it as an
@@ -40,8 +54,10 @@ limitations under the License.
 package cases
 
 import (
+	"bytes"
 	"encoding/json"
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
 	"reflect"
@@ -50,7 +66,7 @@ import (
 	"testing"
 	"time"
 
-	"github.com/onsi/ginkgo/v2"
+	yamlv3 "gopkg.in/yaml.v3"
 	"sigs.k8s.io/yaml"
 )
 
@@ -61,6 +77,7 @@ type Duration time.Duration
 func (d Duration) Std() time.Duration { return time.Duration(d) }
 
 // UnmarshalJSON accepts a JSON string in Go duration syntax.
+// Null is rejected: only a pointer to Duration can represent an omitted value.
 func (d *Duration) UnmarshalJSON(b []byte) error {
 	var s string
 	if err := json.Unmarshal(b, &s); err != nil {
@@ -110,9 +127,9 @@ func LoadFiles[T any](dir string) ([]File[T], error) {
 		if err != nil {
 			return nil, fmt.Errorf("cases: read %s: %w", rel, err)
 		}
-		var tc T
-		if err := yaml.UnmarshalStrict(body, &tc); err != nil {
-			return nil, fmt.Errorf("cases: decode %s: %w", rel, err)
+		tc, err := decodeCase[T](rel, body)
+		if err != nil {
+			return nil, err
 		}
 		stem := strings.TrimSuffix(filepath.Base(rel), ".yaml")
 		if err := checkName(tc, stem, rel); err != nil {
@@ -123,7 +140,7 @@ func LoadFiles[T any](dir string) ([]File[T], error) {
 	return out, nil
 }
 
-// MustLoad is LoadFiles for a Ginkgo container, which has no testing.T yet.
+// MustLoad is LoadFiles for a test container that has no testing.T yet.
 // A bad directory panics so the suite fails instead of running zero cases.
 func MustLoad[T any](dir string) []File[T] {
 	files, err := LoadFiles[T](dir)
@@ -149,28 +166,7 @@ func Run[T any](t *testing.T, dir string, fn func(t *testing.T, tc T)) {
 	}
 }
 
-// Entries loads dir as Ginkgo table entries, one Entry per file.
-// description supplies the entry text; a nil description, or a result of "",
-// uses the file stem. A load error panics so suite setup fails loudly.
-func Entries[T any](dir string, description func(T) string) []ginkgo.TableEntry {
-	files, err := LoadFiles[T](dir)
-	if err != nil {
-		panic(err)
-	}
-	entries := make([]ginkgo.TableEntry, len(files))
-	for i, f := range files {
-		text := f.Stem
-		if description != nil {
-			if d := description(f.Case); d != "" {
-				text = d
-			}
-		}
-		entries[i] = ginkgo.Entry(text, f.Case)
-	}
-	return entries
-}
-
-// SpecName is the subtest or entry text for a case: its Description, or stem.
+// SpecName is the subtest text for a case: its Description, or stem.
 func SpecName[T any](tc T, stem string) string {
 	if d, ok := stringField(tc, "Description"); ok && d != "" {
 		return d
@@ -178,20 +174,88 @@ func SpecName[T any](tc T, stem string) string {
 	return stem
 }
 
+func decodeCase[T any](rel string, body []byte) (T, error) {
+	var zero T
+	if err := oneMapping(body); err != nil {
+		return zero, fmt.Errorf("cases: decode %s: %w", rel, err)
+	}
+	var tc T
+	if err := yaml.UnmarshalStrict(body, &tc); err != nil {
+		return zero, fmt.Errorf("cases: decode %s: %w", rel, err)
+	}
+	var tree any
+	if err := yaml.Unmarshal(body, &tree); err != nil {
+		return zero, fmt.Errorf("cases: decode %s: %w", rel, err)
+	}
+	if err := checkKeys(tree, reflect.TypeOf(tc)); err != nil {
+		return zero, fmt.Errorf("cases: decode %s: %w", rel, err)
+	}
+	return tc, nil
+}
+
+// oneMapping requires exactly one YAML document, and that document to be a
+// non-null mapping. yaml.v3 counts documents; sigs.k8s.io/yaml then decodes
+// the bytes, so YAML 1.1 values are unchanged.
+func oneMapping(body []byte) error {
+	dec := yamlv3.NewDecoder(bytes.NewReader(body))
+	var (
+		count int
+		doc   yamlv3.Node
+	)
+	for {
+		var next yamlv3.Node
+		err := dec.Decode(&next)
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			return err
+		}
+		count++
+		doc = next
+	}
+	if count == 0 {
+		return fmt.Errorf("no document")
+	}
+	if count != 1 {
+		return fmt.Errorf("%d documents, want 1", count)
+	}
+	if doc.Kind != yamlv3.DocumentNode || len(doc.Content) != 1 {
+		return fmt.Errorf("document is empty")
+	}
+	n := doc.Content[0]
+	if n.Tag == "!!null" {
+		return fmt.Errorf("document is null")
+	}
+	if n.Kind != yamlv3.MappingNode {
+		return fmt.Errorf("document must be a mapping")
+	}
+	return nil
+}
+
 func yamlRels(dir string) ([]string, error) {
 	var rels []string
 	err := filepath.WalkDir(dir, func(path string, d os.DirEntry, err error) error {
 		if err != nil {
 			return err
 		}
-		if d.IsDir() || !strings.HasSuffix(d.Name(), ".yaml") {
+		if d.IsDir() {
 			return nil
 		}
 		rel, err := filepath.Rel(dir, path)
 		if err != nil {
 			return err
 		}
-		rels = append(rels, filepath.ToSlash(rel))
+		rel = filepath.ToSlash(rel)
+		// .yml is an error rather than a skip: one extension keeps the stem
+		// unambiguous, and a mis-named file must not vanish from the run.
+		if strings.HasSuffix(d.Name(), ".yml") {
+			return fmt.Errorf("%s: .yml files are not loaded; rename it to .yaml", rel)
+		}
+		if !strings.HasSuffix(d.Name(), ".yaml") {
+			return nil
+		}
+		rels = append(rels, rel)
 		return nil
 	})
 	if err != nil {
diff --git a/internal/testutil/cases/cases_test.go b/internal/testutil/cases/cases_test.go
index 75c9cf9..7873563 100644
--- a/internal/testutil/cases/cases_test.go
+++ b/internal/testutil/cases/cases_test.go
@@ -83,17 +83,6 @@ func TestRunUsesDescription(t *testing.T) {
 	}
 }
 
-func TestEntriesFollowsDescription(t *testing.T) {
-	entries := Entries("testdata/ok", func(s sample) string { return s.Description })
-	if len(entries) != 2 {
-		t.Fatalf("got %d entries", len(entries))
-	}
-	entries = Entries[sample]("testdata/ok", nil)
-	if len(entries) != 2 {
-		t.Fatalf("nil description: got %d entries", len(entries))
-	}
-}
-
 func TestRejectsBadDocuments(t *testing.T) {
 	checks := []struct {
 		dir  string
@@ -133,6 +122,145 @@ func TestMustLoadPanics(t *testing.T) {
 	MustLoad[sample](filepath.Join(t.TempDir(), "absent"))
 }
 
+func TestRejectsDocumentsThatAreNotOneMapping(t *testing.T) {
+	checks := []struct {
+		name string
+		body string
+		want string
+	}{
+		{"empty.yaml", "", "no document"},
+		{"comment.yaml", "# nothing to run\n", "no document"},
+		{"blank.yaml", "\n\n", "no document"},
+		{"null.yaml", "null\n", "document is null"},
+		{"tilde.yaml", "~\n", "document is null"},
+		{"sep.yaml", "---\n", "document is null"},
+		{"seq.yaml", "[]\n", "must be a mapping"},
+		{"multi.yaml", "name: multi\ninputs:\n  value: 1\nexpected:\n  value: 1\n---\nother: 2\n", "2 documents"},
+		{"later.yaml", "name: later\ninputs:\n  value: 1\nexpected:\n  value: 1\n---\nbogus: 2\n", "2 documents"},
+	}
+	for _, tc := range checks {
+		t.Run(tc.name, func(t *testing.T) {
+			dir := writeYAML(t, tc.name, tc.body)
+			_, err := Load[sample](dir)
+			if err == nil || !strings.Contains(err.Error(), tc.want) || !strings.Contains(err.Error(), tc.name) {
+				t.Fatalf("got %v, want it to name %s and contain %q", err, tc.name, tc.want)
+			}
+		})
+	}
+}
+
+func TestLeadingSeparatorIsOneDocument(t *testing.T) {
+	dir := writeYAML(t, "lead.yaml", "---\nname: lead\ninputs:\n  value: 3\nexpected:\n  value: 3\n")
+	got, err := Load[sample](dir)
+	if err != nil || len(got) != 1 || got[0].Inputs.Value != 3 {
+		t.Fatalf("got %+v, %v", got, err)
+	}
+}
+
+func TestEmptyMappingIsOneDocument(t *testing.T) {
+	dir := writeYAML(t, "empty-map.yaml", "{}\n")
+	got, err := Load[sample](dir)
+	if err != nil || len(got) != 1 {
+		t.Fatalf("got %+v, %v", got, err)
+	}
+}
+
+func TestRejectsYMLExtension(t *testing.T) {
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, "kept.yaml"), []byte("name: kept\ninputs:\n  value: 1\nexpected:\n  value: 1\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(dir, "skipped.yml"), []byte("name: skipped\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	_, err := Load[sample](dir)
+	if err == nil || !strings.Contains(err.Error(), "skipped.yml") || !strings.Contains(err.Error(), ".yml") {
+		t.Fatalf("got %v", err)
+	}
+}
+
+func TestRejectsCaseMismatchedKey(t *testing.T) {
+	body := "name: case\ninputs:\n  Value: 1\nexpected:\n  value: 1\n"
+	dir := writeYAML(t, "case.yaml", body)
+	_, err := Load[sample](dir)
+	if err == nil || !strings.Contains(err.Error(), "case.yaml") || !strings.Contains(err.Error(), "Value") {
+		t.Fatalf("got %v", err)
+	}
+}
+
+func TestRejectsCaseMismatchInSliceAndMap(t *testing.T) {
+	dir := writeYAML(t, "slice.yaml", "items:\n  - Kind: pod\n")
+	if _, err := Load[nest](dir); err == nil || !strings.Contains(err.Error(), "Kind") {
+		t.Fatalf("slice: %v", err)
+	}
+	dir = writeYAML(t, "map.yaml", "bags:\n  x:\n    N: 1\n")
+	if _, err := Load[nest](dir); err == nil || !strings.Contains(err.Error(), "N") {
+		t.Fatalf("map: %v", err)
+	}
+	// "n" is quoted: an unquoted n is a YAML 1.1 boolean, not a key.
+	dir = writeYAML(t, "ok.yaml", "items:\n  - kind: pod\nbags:\n  x:\n    \"n\": 1\n")
+	got, err := Load[nest](dir)
+	if err != nil || len(got) != 1 || got[0].Items[0].Kind != "pod" || got[0].Bags["x"].N != 1 {
+		t.Fatalf("got %+v, %v", got, err)
+	}
+}
+
+func TestNonPointerDurationRejectsNull(t *testing.T) {
+	dir := writeYAML(t, "wait.yaml", "name: wait\ninputs:\n  value: 1\n  wait: null\nexpected:\n  value: 1\n")
+	_, err := Load[sample](dir)
+	if err == nil || !strings.Contains(err.Error(), "wait.yaml") || !strings.Contains(err.Error(), "duration") {
+		t.Fatalf("got %v", err)
+	}
+	dir = writeYAML(t, "zero.yaml", "name: zero\ninputs:\n  value: 1\n  wait: \"0s\"\nexpected:\n  value: 1\n")
+	got, err := Load[sample](dir)
+	if err != nil || got[0].Inputs.Wait.Std() != 0 {
+		t.Fatalf("got %+v, %v", got, err)
+	}
+	dir = writeYAML(t, "bare.yaml", "name: bare\ninputs:\n  value: 1\n  wait: 0\nexpected:\n  value: 1\n")
+	if _, err := Load[sample](dir); err == nil {
+		t.Fatal("bare 0 decoded as a duration")
+	}
+}
+
+func TestPromotedEmbeddedFieldsAreCaseSensitive(t *testing.T) {
+	dir := writeYAML(t, "ok.yaml", "a: 1\nb: 2\n")
+	got, err := Load[outer](dir)
+	if err != nil || got[0].A != 1 || got[0].B != 2 {
+		t.Fatalf("got %+v, %v", got, err)
+	}
+	dir = writeYAML(t, "bad.yaml", "A: 1\nb: 2\n")
+	if _, err := Load[outer](dir); err == nil || !strings.Contains(err.Error(), "A") {
+		t.Fatalf("got %v", err)
+	}
+}
+
+type nest struct {
+	Items []struct {
+		Kind string `json:"kind"`
+	} `json:"items"`
+	Bags map[string]struct {
+		N int `json:"n"`
+	} `json:"bags"`
+}
+
+type inner struct {
+	A int `json:"a"`
+}
+
+type outer struct {
+	inner
+	B int `json:"b"`
+}
+
+func writeYAML(t *testing.T, name, body string) string {
+	t.Helper()
+	dir := t.TempDir()
+	if err := os.WriteFile(filepath.Join(dir, name), []byte(body), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return dir
+}
+
 func TestUnreadableFileFails(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "gone.yaml")
diff --git a/internal/testutil/cases/ginkgocases/entries.go b/internal/testutil/cases/ginkgocases/entries.go
new file mode 100644
index 0000000..a875eee
--- /dev/null
+++ b/internal/testutil/cases/ginkgocases/entries.go
@@ -0,0 +1,49 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+// Package ginkgocases builds Ginkgo table entries from case files.
+//
+// It lives apart from package cases so a test binary that only calls
+// cases.Run or cases.Load does not import Ginkgo and does not register
+// ginkgo flags.
+package ginkgocases
+
+import (
+	"github.com/onsi/ginkgo/v2"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+)
+
+// Entries loads dir as Ginkgo table entries, one Entry per file.
+// description supplies the entry text; a nil description, or a result of "",
+// uses the file stem. A load error panics so suite setup fails loudly.
+func Entries[T any](dir string, description func(T) string) []ginkgo.TableEntry {
+	files, err := cases.LoadFiles[T](dir)
+	if err != nil {
+		panic(err)
+	}
+	entries := make([]ginkgo.TableEntry, len(files))
+	for i, f := range files {
+		text := f.Stem
+		if description != nil {
+			if d := description(f.Case); d != "" {
+				text = d
+			}
+		}
+		entries[i] = ginkgo.Entry(text, f.Case)
+	}
+	return entries
+}
diff --git a/internal/testutil/cases/ginkgocases/entries_test.go b/internal/testutil/cases/ginkgocases/entries_test.go
new file mode 100644
index 0000000..9261ccf
--- /dev/null
+++ b/internal/testutil/cases/ginkgocases/entries_test.go
@@ -0,0 +1,49 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package ginkgocases
+
+import (
+	"testing"
+
+	"github.com/modelsphere/llm-operator/internal/testutil/cases"
+)
+
+// sample matches the documents under ../testdata/ok.
+type sample struct {
+	Name        string `json:"name,omitempty"`
+	Description string `json:"description,omitempty"`
+	Inputs      struct {
+		Value int             `json:"value"`
+		Wait  cases.Duration  `json:"wait,omitempty"`
+		Maybe *cases.Duration `json:"maybe,omitempty"`
+		Empty *cases.Duration `json:"empty,omitempty"`
+	} `json:"inputs"`
+	Expected struct {
+		Value int `json:"value"`
+	} `json:"expected"`
+}
+
+func TestEntriesFollowsDescription(t *testing.T) {
+	entries := Entries("../testdata/ok", func(s sample) string { return s.Description })
+	if len(entries) != 2 {
+		t.Fatalf("got %d entries", len(entries))
+	}
+	entries = Entries[sample]("../testdata/ok", nil)
+	if len(entries) != 2 {
+		t.Fatalf("nil description: got %d entries", len(entries))
+	}
+}
diff --git a/internal/testutil/cases/keys.go b/internal/testutil/cases/keys.go
new file mode 100644
index 0000000..d91e20a
--- /dev/null
+++ b/internal/testutil/cases/keys.go
@@ -0,0 +1,133 @@
+/*
+Copyright 2026.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+package cases
+
+import (
+	"encoding/json"
+	"fmt"
+	"reflect"
+	"strings"
+)
+
+// checkKeys walks a decoded YAML tree beside t and rejects a mapping key
+// that does not exactly equal a json name. encoding/json, which
+// sigs.k8s.io/yaml uses, matches those names without regard to case.
+// Types that implement json.Unmarshaler are leaves: their own decoder owns
+// the value, and Kubernetes types that decode themselves fall in that set.
+func checkKeys(tree any, t reflect.Type) error {
+	if t == nil || tree == nil {
+		return nil
+	}
+	for t.Kind() == reflect.Pointer {
+		t = t.Elem()
+	}
+	if hasUnmarshalJSON(t) {
+		return nil
+	}
+	switch t.Kind() {
+	case reflect.Struct:
+		m, ok := tree.(map[string]any)
+		if !ok {
+			return fmt.Errorf("mapping for %s has type %T", typeName(t), tree)
+		}
+		fields := jsonFields(t)
+		for k, v := range m {
+			ft, ok := fields[k]
+			if !ok {
+				return fmt.Errorf("key %q does not match a field (names are case-sensitive)", k)
+			}
+			if err := checkKeys(v, ft); err != nil {
+				return fmt.Errorf("%s: %w", k, err)
+			}
+		}
+	case reflect.Slice, reflect.Array:
+		items, ok := tree.([]any)
+		if !ok {
+			return fmt.Errorf("sequence for %s has type %T", typeName(t), tree)
+		}
+		for i, item := range items {
+			if err := checkKeys(item, t.Elem()); err != nil {
+				return fmt.Errorf("[%d]: %w", i, err)
+			}
+		}
+	case reflect.Map:
+		if t.Key().Kind() != reflect.String {
+			return nil
+		}
+		m, ok := tree.(map[string]any)
+		if !ok {
+			return fmt.Errorf("mapping for %s has type %T", typeName(t), tree)
+		}
+		for k, v := range m {
+			if err := checkKeys(v, t.Elem()); err != nil {
+				return fmt.Errorf("[%s]: %w", k, err)
+			}
+		}
+	default:
+		return nil
+	}
+	return nil
+}
+
+func jsonFields(t reflect.Type) map[string]reflect.Type {
+	out := map[string]reflect.Type{}
+	collectFields(out, t)
+	return out
+}
+
+// collectFields records the json names encoding/json would accept, including
+// fields promoted from an anonymous embedded struct.
+func collectFields(out map[string]reflect.Type, t reflect.Type) {
+	for f := range t.Fields() {
+		if f.PkgPath != "" && !f.Anonymous {
+			continue
+		}
+		tag := f.Tag.Get("json")
+		if tag == "-" {
+			continue
+		}
+		name, _, _ := strings.Cut(tag, ",")
+		// encoding/json promotes an anonymous struct when the tag names no
+		// field of its own. json:",inline" is that case: the name is empty.
+		if f.Anonymous && name == "" {
+			ft := f.Type
+			for ft.Kind() == reflect.Pointer {
+				ft = ft.Elem()
+			}
+			if ft.Kind() == reflect.Struct && !hasUnmarshalJSON(ft) {
+				collectFields(out, ft)
+				continue
+			}
+		}
+		if name == "" {
+			name = f.Name
+		}
+		out[name] = f.Type
+	}
+}
+
+func hasUnmarshalJSON(t reflect.Type) bool {
+	u := reflect.TypeFor[json.Unmarshaler]()
+	return t.Implements(u) || reflect.PointerTo(t).Implements(u)
+}
+
+func typeName(t reflect.Type) string {
+	if t.Name() != "" {
+		return t.Name()
+	}
+	return t.String()
+}