fix(infra): rename L4 pool ci-compile → ci-rl, remove stale moved blocks

Applied terragrunt to recreate the L4 GPU pool with the correct name
`ci-rl` (was `ci-compile` due to immutable Scaleway pool names).
Updated all K8s manifests and comments to match. Removed the 3 stale
`moved` blocks from main.tf since the state renames are now applied.

Pool naming is now consistent across Terraform, Scaleway, and K8s configs:
- ci-compile-cpu (POP2-32C-128G) — CPU compilation
- ci-rl (L4-1-24G) — RL training / CUDA compile
- ci-training (L40S-1-48G) — supervised training + hyperopt

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

.gitlab-ci.yml

@@ -9,7 +9,7 @@
 #   kapsule,rust → CPU runner → ci-compile-cpu pool (POP2-32C-128G, 32 vCPU)
 #   kapsule,gpu  → GPU runner → ci-training (L40S)
 # Override nodeSelector: KUBERNETES_NODE_SELECTOR_k8s.scaleway.com/pool-name: "<name>" (k=v format)
-# Pools: ci-compile-cpu (POP2-32C-128G), ci-compile (L4 compile), ci-training (L40S training), services, gitlab
+# Pools: ci-compile-cpu (POP2-32C-128G), ci-rl (L4 GPU), ci-training (L40S training), services, gitlab
 # All compilation (services + training) runs on ci-compile-cpu. CUDA stubs in ci-builder image.
 
 # Ensure pipeline is always created — individual job rules handle filtering.

infra/k8s/monitoring/APPLY-ORDER.md

@@ -94,7 +94,7 @@
 
 4. Repeat for gitlab pool (if not autoscaling, same process).
 
-5. Autoscaling pools (ci-training, ci-compile, gpu-dev):
+5. Autoscaling pools (ci-training, ci-rl, gpu-dev):
    - No action needed — next scale-up automatically uses private IP.
 
 6. Verify Tailscale still works:

infra/k8s/monitoring/dcgm-exporter.yaml

@@ -1,5 +1,5 @@
 # NVIDIA DCGM Exporter — GPU metrics for Cockpit
-# Runs on all GPU node pools (ci-training, ci-compile, gpu-dev)
+# Runs on all GPU node pools (ci-training, ci-rl, gpu-dev)
 # Alloy auto-discovers and scrapes these pods via annotations
 ---
 apiVersion: v1

infra/k8s/training/image-prepuller.yaml

@@ -1,5 +1,5 @@
 # DaemonSet image pre-puller — keeps training images cached on GPU nodes
-# Runs on ci-compile (L4) and ci-training (L40S) pools so training jobs skip the pull.
+# Runs on ci-rl (L4) and ci-training (L40S) pools so training jobs skip the pull.
 # Init containers pull :latest tags, then the main container sleeps forever.
 ---
 apiVersion: apps/v1
@@ -27,8 +27,7 @@ spec:
                   - key: k8s.scaleway.com/pool-name
                     operator: In
                     values:
-                      - ci-compile
-                      - ci-compile  # remove after terragrunt apply renames pool
+                      - ci-rl
                       - ci-training
       tolerations:
         - key: nvidia.com/gpu

infra/modules/kapsule/main.tf

@@ -1,21 +1,3 @@
-# Terraform state renames — prevents destroy+recreate of state entries.
-# Pool name attribute changes (always-on→services, ci-build→ci-training)
-# will still force replacement since Scaleway pool names are immutable.
-moved {
-  from = scaleway_k8s_pool.always_on
-  to   = scaleway_k8s_pool.services
-}
-
-moved {
-  from = scaleway_k8s_pool.ci_build
-  to   = scaleway_k8s_pool.ci_training
-}
-
-moved {
-  from = scaleway_k8s_pool.ci_compile
-  to   = scaleway_k8s_pool.ci_rl
-}
-
 resource "scaleway_vpc_private_network" "foxhunt" {
   name   = "${var.cluster_name}-pn"
   region = var.region