diff --git a/cmd/weka-capacity/plan.go b/cmd/weka-capacity/plan.go
index 11ab6258c..6ecfe68b8 100644
--- a/cmd/weka-capacity/plan.go
+++ b/cmd/weka-capacity/plan.go
@@ -555,8 +555,8 @@ func planSummary(p *capacityplanner.CapacityPlan, result *inventory.Result, desi
 	}
 
 	idle := 0
-	for _, n := range result.Inventory {
-		if _, used := newNodes[n.NodeName]; !used {
+	for i := range result.Inventory {
+		if _, used := newNodes[result.Inventory[i].NodeName]; !used {
 			idle++
 		}
 	}
diff --git a/internal/capacityplanner/autofulldrives_compute.go b/internal/capacityplanner/autofulldrives_compute.go
index 064274ba5..980bf675f 100644
--- a/internal/capacityplanner/autofulldrives_compute.go
+++ b/internal/capacityplanner/autofulldrives_compute.go
@@ -294,8 +294,8 @@ func autoKeptCompute(in *autoComputeInput) (kept []autoComputeEntry, pinned map[
 // but which of "wait" or "intervene" is right turns on this fact, and nothing else in the report carries it.
 func drainingComputeClause(in *autoComputeInput) string {
 	var nodes []string
-	for node, nc := range in.remaining {
-		if nc.HasDeletingComputeContainer {
+	for node := range in.remaining {
+		if in.remaining[node].HasDeletingComputeContainer {
 			nodes = append(nodes, node)
 		}
 	}
diff --git a/internal/capacityplanner/cores.go b/internal/capacityplanner/cores.go
index c0d9cff14..24cfbfc2d 100644
--- a/internal/capacityplanner/cores.go
+++ b/internal/capacityplanner/cores.go
@@ -106,7 +106,7 @@ func finalCores(c *ExistingContainer, growth map[string]*ContainerGrowth) int {
 // tlcDriveCoresForContainer returns the TLC-attributable core count for a container's final state. A
 // TLC-only container (qlcGiB <= 0, checked before the tlcGiB<=0 short-circuit below) attributes all of
 // assignedCores to TLC — the only reliable figure for auto-full-drives containers, which are always
-// TLC-only. Mixed containers fall back to capacity-derived TlcDriveCores(tlcGiB, cons).
+// TLC-only. Mixed containers derive the share from capacity, capped at assignedCores.
 func tlcDriveCoresForContainer(tlcGiB, qlcGiB, assignedCores int, cons *CapacityConstraints) int {
 	if qlcGiB <= 0 && assignedCores > 0 {
 		return assignedCores
@@ -114,7 +114,13 @@ func tlcDriveCoresForContainer(tlcGiB, qlcGiB, assignedCores int, cons *Capacity
 	if tlcGiB <= 0 {
 		return 0
 	}
-	return TlcDriveCores(tlcGiB, cons)
+	tlc := TlcDriveCores(tlcGiB, cons)
+	// Only an existing container can need the cap: pinCores keeps new ones at or above the derived value,
+	// so a shortfall means the assigned count predates the current TlcCapacityPerCoreGiB.
+	if assignedCores > 0 {
+		return min(tlc, assignedCores)
+	}
+	return tlc
 }
 
 // qlcDriveCoresForContainer returns the QLC-attributable core count for a single drive container in its
diff --git a/internal/capacityplanner/cores_test.go b/internal/capacityplanner/cores_test.go
new file mode 100644
index 000000000..9ba94f719
--- /dev/null
+++ b/internal/capacityplanner/cores_test.go
@@ -0,0 +1,69 @@
+package capacityplanner
+
+import "testing"
+
+// The split must add up to the cores the container actually has. An over-count reaches totalTlcDriveCores
+// and from there both terms of RequiredComputeCores, over-stating a mixed-pool cluster's compute.
+func TestDriveCoresForContainer_SplitSumsToAssignedCores(t *testing.T) {
+	cons := testCons() // TlcCapacityPerCoreGiB=5120, QlcCapacityPerCoreGiB=51200
+
+	tests := []struct {
+		name                     string
+		tlcGiB, qlcGiB, assigned int
+		wantTlc, wantQlc         int
+	}{
+		{
+			// 10240GiB derives 2 TLC cores, above the 1 assigned.
+			name:   "mixed, assigned below capacity-derived tlc share",
+			tlcGiB: 10240, qlcGiB: 51200, assigned: 1,
+			wantTlc: 1, wantQlc: 0,
+		},
+		{
+			name:   "mixed, assigned above tlc share leaves remainder to qlc",
+			tlcGiB: 10240, qlcGiB: 51200, assigned: 5,
+			wantTlc: 2, wantQlc: 3,
+		},
+		{
+			name:   "mixed, assigned exactly equals tlc share",
+			tlcGiB: 10240, qlcGiB: 51200, assigned: 2,
+			wantTlc: 2, wantQlc: 0,
+		},
+		{
+			name:   "tlc-only attributes every assigned core to tlc",
+			tlcGiB: 10240, qlcGiB: 0, assigned: 4,
+			wantTlc: 4, wantQlc: 0,
+		},
+		{
+			name:   "qlc-only attributes every assigned core to qlc",
+			tlcGiB: 0, qlcGiB: 51200, assigned: 3,
+			wantTlc: 0, wantQlc: 3,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			gotTlc := tlcDriveCoresForContainer(tt.tlcGiB, tt.qlcGiB, tt.assigned, cons)
+			gotQlc := qlcDriveCoresForContainer(tt.tlcGiB, tt.qlcGiB, tt.assigned, cons)
+
+			if gotTlc != tt.wantTlc || gotQlc != tt.wantQlc {
+				t.Errorf("tlc/qlc = %d/%d, want %d/%d", gotTlc, gotQlc, tt.wantTlc, tt.wantQlc)
+			}
+			if got := gotTlc + gotQlc; got != tt.assigned {
+				t.Errorf("tlc+qlc = %d, want %d (the split must not invent or drop cores)", got, tt.assigned)
+			}
+		})
+	}
+}
+
+// With no assigned count to split, both sides fall back to their capacity-derived values and the sum is
+// not bound to assignedCores — 2 TLC cores for 10240GiB plus 1 QLC core for 51200GiB.
+func TestDriveCoresForContainer_UnassignedFallsBackToCapacityDerived(t *testing.T) {
+	cons := testCons()
+
+	gotTlc := tlcDriveCoresForContainer(10240, 51200, 0, cons)
+	gotQlc := qlcDriveCoresForContainer(10240, 51200, 0, cons)
+
+	if gotTlc != 2 || gotQlc != 1 {
+		t.Errorf("tlc/qlc = %d/%d, want 2/1 (capacity-derived on both sides)", gotTlc, gotQlc)
+	}
+}
diff --git a/internal/capacityplanner/inventory/collect.go b/internal/capacityplanner/inventory/collect.go
index e393f1e1f..4e89c42d0 100644
--- a/internal/capacityplanner/inventory/collect.go
+++ b/internal/capacityplanner/inventory/collect.go
@@ -937,8 +937,8 @@ func (c Collector) nodeDetailsFromLists(ctx context.Context, cluster *weka.WekaC
 
 	// Index the planner inventory by node for free-headroom + FD lookup.
 	invByNode := make(map[string]capacityplanner.NodeCapacity, len(inv))
-	for _, nc := range inv {
-		invByNode[nc.NodeName] = nc
+	for i := range inv {
+		invByNode[inv[i].NodeName] = inv[i]
 	}
 
 	// Union of driveNodes + computeNodeList: computeNodeList already aliases driveNodes when the
@@ -1154,15 +1154,16 @@ func resolveInventoryFDValue(node *corev1.Node, fdConfig *weka.FailureDomain) (f
 func mergeRoleNodes(driveInv, computeInv []capacityplanner.NodeCapacity) (inventory []capacityplanner.NodeCapacity, computeNodes map[string]bool) {
 	inventory = append([]capacityplanner.NodeCapacity(nil), driveInv...)
 	index := make(map[string]struct{}, len(inventory))
-	for _, nc := range inventory {
-		index[nc.NodeName] = struct{}{}
+	for i := range inventory {
+		index[inventory[i].NodeName] = struct{}{}
 	}
 	computeNodes = make(map[string]bool, len(computeInv))
-	for _, nc := range computeInv {
+	for i := range computeInv {
+		nc := &computeInv[i]
 		computeNodes[nc.NodeName] = nc.IneligibleReason == ""
 		if _, ok := index[nc.NodeName]; !ok {
 			index[nc.NodeName] = struct{}{}
-			inventory = append(inventory, nc)
+			inventory = append(inventory, *nc)
 		}
 	}
 	return inventory, computeNodes
diff --git a/internal/capacityplanner/planner.go b/internal/capacityplanner/planner.go
index 39fea4552..eda6ec96b 100644
--- a/internal/capacityplanner/planner.go
+++ b/internal/capacityplanner/planner.go
@@ -516,9 +516,10 @@ func PlanCapacity(
 
 	// Working per-node headroom, sorted deterministically by FD then node.
 	states := make(map[string]*nodeState, len(inventory))
-	for _, nc := range inventory {
+	for i := range inventory {
+		nc := &inventory[i]
 		states[nc.NodeName] = &nodeState{
-			nc:                        nc,
+			nc:                        *nc,
 			tlcFree:                   nc.TlcGiB,
 			qlcFree:                   nc.QlcGiB,
 			coresFree:                 nc.AllocatableCPU, // physical CPU remaining
diff --git a/internal/controllers/wekacluster/steps_cluster_creation.go b/internal/controllers/wekacluster/steps_cluster_creation.go
index 02a4dac0c..166f0202a 100644
--- a/internal/controllers/wekacluster/steps_cluster_creation.go
+++ b/internal/controllers/wekacluster/steps_cluster_creation.go
@@ -641,15 +641,7 @@ func (r *wekaClusterReconcilerLoop) buildPlannerComputeContainers(ctx context.Co
 	logger := instrumentation.CurrentSpanLogger(ctx)
 	cluster := r.cluster
 
-	occupied := make(map[string]struct{})
-	for _, c := range existing {
-		if c.Spec.Mode == weka.WekaContainerModeCompute {
-			if n := string(c.GetNodeAffinity()); n != "" {
-				occupied[n] = struct{}{}
-			}
-		}
-	}
-
+	occupied := computeNodesOccupied(existing)
 	totalCount := len(layout)
 
 	for _, entry := range layout {
@@ -687,18 +679,26 @@ func (r *wekaClusterReconcilerLoop) buildPlannerComputeContainers(ctx context.Co
 	return built, skippedReasons
 }
 
+// computeNodesOccupied indexes the nodes already hosting one of this cluster's compute containers — the
+// nodes a planner layout entry creates nothing on, since in-place growth resizes them instead.
+func computeNodesOccupied(containers []*weka.WekaContainer) map[string]struct{} {
+	occupied := make(map[string]struct{})
+	for _, c := range containers {
+		if c.Spec.Mode != weka.WekaContainerModeCompute {
+			continue
+		}
+		if n := string(c.GetNodeAffinity()); n != "" {
+			occupied[n] = struct{}{}
+		}
+	}
+	return occupied
+}
+
 // unusedComputeNodes returns the planner-reserved compute nodes that are not already hosting a compute
 // container, preserving the planner's order. New compute containers pin to these (one each) so they never
 // schedule onto a drive-pinned node lacking the post-drive hugepages to host both.
 func unusedComputeNodes(existing []*weka.WekaContainer, planned []string) []string {
-	used := make(map[string]struct{})
-	for _, c := range existing {
-		if c.Spec.Mode == weka.WekaContainerModeCompute {
-			if n := string(c.GetNodeAffinity()); n != "" {
-				used[n] = struct{}{}
-			}
-		}
-	}
+	used := computeNodesOccupied(existing)
 	free := make([]string, 0, len(planned))
 	for _, n := range planned {
 		if _, taken := used[n]; !taken {
diff --git a/internal/validation/cluster_drive_compute_core_ratio.go b/internal/validation/cluster_drive_compute_core_ratio.go
index 35ed3a3ec..8a5ae1b04 100644
--- a/internal/validation/cluster_drive_compute_core_ratio.go
+++ b/internal/validation/cluster_drive_compute_core_ratio.go
@@ -20,6 +20,9 @@ import (
 // auto-derived drive cores count too — except under clusterCapacity/auto-full-drives, where the planner
 // assigns both sides and the template's numbers are not the ones the cluster runs on. Skips cases
 // already below clusterComputeDriveCoresFloor's hard 1:1 floor, which owns those exclusively.
+//
+// At shipped defaults only exclusive full-drives (ratio 2.0) leaves room between that floor and the
+// recommendation; drive-sharing's 1.0 makes the two equal, so the advisory is inert there by design.
 type clusterDriveComputeCoreRatio struct{}
 
 func (clusterDriveComputeCoreRatio) ID() string {
diff --git a/internal/validation/cluster_drive_compute_core_ratio_test.go b/internal/validation/cluster_drive_compute_core_ratio_test.go
index 0dd7a5511..061b39c96 100644
--- a/internal/validation/cluster_drive_compute_core_ratio_test.go
+++ b/internal/validation/cluster_drive_compute_core_ratio_test.go
@@ -29,7 +29,8 @@ func setRatioConfig(t *testing.T, tlcRatio, fullDrivesRatio float64) {
 }
 
 func TestClusterDriveComputeCoreRatio(t *testing.T) {
-	// Ratio pinned to 2.0, matching the old hardcoded 1:2 behavior this test mirrors.
+	// Pinned above the shipped drive-sharing 1.0, where the advisory is inert by design, so these cases are
+	// observable at all. TestClusterDriveComputeCoreRatio_FullDrivesUsesFullDrivesRatio covers the shipped pair.
 	setRatioConfig(t, 2.0, 2.0)
 
 	v := &clusterDriveComputeCoreRatio{}
diff --git a/internal/validation/cluster_min_drives_feasibility.go b/internal/validation/cluster_min_drives_feasibility.go
index c8daf90f9..4b881fd75 100644
--- a/internal/validation/cluster_min_drives_feasibility.go
+++ b/internal/validation/cluster_min_drives_feasibility.go
@@ -36,14 +36,12 @@ func (clusterMinDrivesFeasibility) Validate(ctx context.Context, c client.Client
 
 	fldPath := field.NewPath("spec", "startIoConditions", "minNumDrives")
 
-	// Checked before the nil guard below: a nil dynamicTemplate IS auto-full-drives mode, so it takes
-	// the per-node branch rather than falling through as "nothing configured".
+	// A nil dynamicTemplate IS auto-full-drives mode (UsesAutoFullDrives returns true on a nil receiver),
+	// so it takes the per-node branch rather than falling through as "nothing configured" — which is also
+	// what lets everything below dereference Dynamic unguarded.
 	if cluster.Spec.Dynamic.UsesAutoFullDrives() {
 		return validateMinDrivesAutoFullDrives(ctx, c, cluster, minNumDrives, fldPath)
 	}
-	if cluster.Spec.Dynamic == nil {
-		return nil
-	}
 
 	driveContainers := cluster.Spec.Dynamic.DriveContainers
 	numDrives := cluster.Spec.Dynamic.NumDrives