Upgrade the Temporal Worker Controller to 1.12

[duynhne]

What

This upgrades both Worker Controller charts from 0.28.0 to 0.31.0 (controller 1.9.0 → 1.12.0). The owner picked this from the SDK and server follow-ups in #1107.

From the official notes (v1.10.0, v1.11.0, v1.12.0):

• 1.12:
  • prunes superseded inactive versions;
  • surfaces poller and gate-workflow health as conditions and events;
  • backs off failing drained-version deletes;
  • no longer treats an already-current version as a rollback;
  • adds a CA-injector Job option.
• 1.11:
  • scales draining versions back up from 0;
  • removes the cert-manager subchart.
• 1.10:
  • deletes rendered autoscalers at sunset;
  • prunes a version before deleting its Deployment;
  • is more resilient to inconsistent Temporal state;
  • adds the ClusterConnection CRD.

cert-manager: this platform always had certmanager.install: false, so the upgrade guide's migration does not apply. I dropped the now-dead install key. Our values render identically with and without it on 0.31.0. The webhook certificate (temporal-worker-controller-serving-cert → Secret webhook-server-cert, CA injected into the WorkerResourceTemplate webhook) is still issued by the platform's cert-manager.

Render diff, 0.28.0 vs 0.31.0, with our values:

• RBAC for clusterconnections and clusterconnections/finalizers;
• env WRT_HPA_MATCH_LABELS_STRIP_TEMPORAL_PREFIX=false;
• no hooks.

Verified on Kind (branch overlay)

Check	Result
HelmReleases	crds upgraded to 0.31.0 first (dependsOn), then the manager; image temporalio/temporal-worker-controller:v1.12.0
WorkerDeployments	order/order-fulfillment (current 2.10.1-fcc8) and checkout/checkout-abandon (current 0.13.1-4955): ConnectionHealthy=True, Ready=True, RolloutComplete=True
WorkerResourceTemplates and KEDA	both scaler WRTs Ready=True; both ScaledObjects Ready
Manager log	no errors in the 5 min after the upgrade
make e2e-saga GATE=kind	SG.1–SG.4 PASS (SG.4: Pinned on the Current build, no ramp)

Draft until the rollout drill runs. The G4 release (order and checkout on SDK 1.49) pins new worker builds on Kind. That gives a real Progressive rollout (10 → 50 → 100) on this controller, and lets me see the superseded version drain and get pruned. I'll add that evidence here before marking this ready.

[duynhne]

Rollout drill on controller 1.12 (from #1196's pins)

The real Progressive rollout ran on Kind with this branch's controller overlaid. The controller log:

Time (+07)	Worker	Step
14:56:42	order-fulfillment	applying ramp 10% → 2.10.2-8d45
14:57:12	order-fulfillment	applying ramp 50%
14:57:43	order-fulfillment	registering new current version 2.10.2-8d45; 2.10.1-fcc8 → Draining
14:56:49 / 14:57:19 / 14:57:50	checkout-abandon	the same three steps → current 0.13.2-f678; 0.13.1-4955 → Draining

• Status: ConnectionHealthy=True, Ready=True, RolloutComplete=True on both.
• Pollers: an ActivePollers event for both versions during the ramp.
• Gate: make e2e GATE=kind passes all rows, SG.4 (Pinned on the new Current build) included.
• KEDA: the new version's ScaledObject logged Worker Deployment Version not found once, at 07:56:42Z, before the version registered. There were no errors after that, and KedaScaledObjectErrors cleared by itself.

Still to observe: pruning of the superseded versions. The spec has sunset.scaledownDelay: 1h and deleteDelay: 0s, so the drained versions scale down about an hour after they drain, and then their Deployments are deleted. I'll add that read-back here and then mark this ready.

[duynhne]

Version pruning, observed on controller 1.12

About an hour after promotion (sunset.scaledownDelay: 1h, deleteDelay: 0s), the controller log shows:

Time (+07)	Worker	Step
16:05:29–16:06:00	order-fulfillment	deleted worker resource on version sunset (the old version's ScaledObject)
16:06:00	order-fulfillment	deleted deprecated worker deployment version 2.10.1-fcc8, deleting deployment
16:06:07	checkout-abandon	deleted deprecated worker deployment version 0.13.1-4955, deleting deployment

Afterwards:

• Each WorkerDeployment has deprecatedVersions: [].
• Only order-fulfillment-2-10-2-8d45 and checkout-abandon-0-13-2-f678 remain, each 1/1.
• One ScaledObject per worker is left, both Ready.

That completes the drill: ramp 10 → 50 → current, drain, scale-down and prune all ran on 1.12. Ready for review.