fix(infra): use Kapsule built-in pool-name label for node selection

Kapsule doesn't allow node-labels in kubelet_args. Instead, use the
auto-assigned k8s.scaleway.com/pool-name label that every node gets.

- Revert kubelet_args from Terraform (not supported)
- Switch runner node_selector: pool→k8s.scaleway.com/pool-name
- Update CI pipeline KUBERNETES_NODE_SELECTOR to match
- Helm upgraded both runners (CPU + GPU)

No more manual kubectl label after node scale-up.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.gitlab-ci.yml

@@ -7,7 +7,7 @@
 # Runner routing (by tags):
 #   kapsule,rust,docker → CPU runner → ci-compile-cpu pool (POP2-32C-128G, 32 vCPU)
 #   kapsule,gpu         → GPU runner → ci-compile (L4) or ci-training (L40S)
-# Override nodeSelector: KUBERNETES_NODE_SELECTOR_pool: "pool=<name>" (k=v format)
+# Override nodeSelector: KUBERNETES_NODE_SELECTOR_k8s.scaleway.com/pool-name: "<name>" (k=v format)
 # Pools: ci-compile-cpu (POP2-32C-128G), ci-compile (L4), ci-training (L40S), services, gitlab
 
 # Ensure pipeline is always created — individual job rules handle filtering.
@@ -438,7 +438,7 @@ build-training:
   tags:
     - kapsule
     - gpu
-  # Runner default node_selector: pool=ci-compile (L4-1-24G)
+  # Runner default node_selector: k8s.scaleway.com/pool-name=ci-compile (L4-1-24G)
   variables:
     # nvidia runtime provided by runner default (runtime_class_name = "nvidia")
     # RL hyperopt is CPU-bound (parallel PSO trials). Use full L4 (8 vCPU).
@@ -472,8 +472,8 @@ build-training:
   variables:
     # nvidia runtime provided by runner default (runtime_class_name = "nvidia")
     # Route to L40S pool (48GB VRAM for large supervised models)
-    # Overrides runner default pool=ci-compile
-    KUBERNETES_NODE_SELECTOR_pool: "pool=ci-training"
+    # Overrides runner default to L40S pool
+    KUBERNETES_NODE_SELECTOR_k8s.scaleway.com/pool-name: "ci-training"
     # Supervised is GPU-bound — minimal CPU. Low request allows 3-4 concurrent jobs.
     KUBERNETES_CPU_REQUEST: "1000m"
     KUBERNETES_CPU_LIMIT: "2000m"

infra/modules/kapsule/main.tf

@@ -51,10 +51,6 @@ resource "scaleway_k8s_pool" "services" {
   public_ip_disabled = var.public_ip_disabled
   region             = var.region
 
-  kubelet_args = {
-    "node-labels" = "pool=services"
-  }
-
   lifecycle {
     ignore_changes = [size]
   }
@@ -73,13 +69,10 @@ resource "scaleway_k8s_pool" "gitlab" {
   autohealing        = true
   public_ip_disabled = var.public_ip_disabled
   region             = var.region
-
-  kubelet_args = {
-    "node-labels" = "pool=gitlab"
-  }
 }
 
-# CI compile pool (L4 — 24GB VRAM, CUDA CC 89, for cargo check/test/build)
+# CI compile pool (L4 — 24GB VRAM, CUDA CC 89, for RL training + hyperopt)
+# Nodes auto-labeled: k8s.scaleway.com/pool-name=ci-compile
 resource "scaleway_k8s_pool" "ci_compile" {
   count       = var.enable_ci_compile_pool ? 1 : 0
   cluster_id  = scaleway_k8s_cluster.foxhunt.id
@@ -93,16 +86,13 @@ resource "scaleway_k8s_pool" "ci_compile" {
   public_ip_disabled = var.public_ip_disabled
   region             = var.region
 
-  kubelet_args = {
-    "node-labels" = "pool=ci-compile"
-  }
-
   lifecycle {
     ignore_changes = [size]
   }
 }
 
-# CI training pool (L40S — 48GB VRAM, for hyperopt + training jobs)
+# CI training pool (L40S — 48GB VRAM, for supervised hyperopt + training)
+# Nodes auto-labeled: k8s.scaleway.com/pool-name=ci-training
 resource "scaleway_k8s_pool" "ci_training" {
   count       = var.enable_ci_training_pool ? 1 : 0
   cluster_id  = scaleway_k8s_cluster.foxhunt.id
@@ -116,16 +106,13 @@ resource "scaleway_k8s_pool" "ci_training" {
   public_ip_disabled = var.public_ip_disabled
   region             = var.region
 
-  kubelet_args = {
-    "node-labels" = "pool=ci-training"
-  }
-
   lifecycle {
     ignore_changes = [size]
   }
 }
 
-# CPU compile pool (POP2-HC — high CPU, no GPU, for cargo check/build)
+# CPU compile pool (POP2 — high CPU, no GPU, for cargo check/build)
+# Nodes auto-labeled: k8s.scaleway.com/pool-name=ci-compile-cpu
 resource "scaleway_k8s_pool" "ci_compile_cpu" {
   count       = var.enable_ci_compile_cpu_pool ? 1 : 0
   cluster_id  = scaleway_k8s_cluster.foxhunt.id
@@ -139,16 +126,13 @@ resource "scaleway_k8s_pool" "ci_compile_cpu" {
   public_ip_disabled = var.public_ip_disabled
   region             = var.region
 
-  kubelet_args = {
-    "node-labels" = "pool=ci-compile-cpu"
-  }
-
   lifecycle {
     ignore_changes = [size]
   }
 }
 
 # CPU pool for remote development (DevPod — autoscales to zero)
+# Nodes auto-labeled: k8s.scaleway.com/pool-name=gpu-dev
 resource "scaleway_k8s_pool" "dev" {
   count                = var.enable_dev_pool ? 1 : 0
   cluster_id           = scaleway_k8s_cluster.foxhunt.id
@@ -163,12 +147,7 @@ resource "scaleway_k8s_pool" "dev" {
   wait_for_pool_ready  = false
   region               = var.region
 
-  kubelet_args = {
-    "node-labels" = "pool=gpu-dev"
-  }
-
   lifecycle {
     ignore_changes = [size]
   }
 }
-