trafficrouting: avoid divide-by-zero on PromoteFull with zero-replica rollout

reconcileTrafficRouting computes the PromoteFull desired weight as
(maxWeight * canaryAvailable) / *rollout.Spec.Replicas when
dynamicStableScale is on. Rollouts that are temporarily scaled to
zero still have a spec and are promotable, so PromoteFull can run
while *rollout.Spec.Replicas is 0 - the division panics, the
recovered error is logged in a hot loop, and the rollout is stuck
in a progressing state (#3686).

Guard the division: pull spec.Replicas out, treat a nil pointer as
zero, and only compute the ratio when we actually have replicas.
With zero spec replicas there is nothing to ratio against, so
desiredWeight stays at its zero default rather than shifting
traffic to a canary that has nowhere to land.

Signed-off-by: SAY-5 <SAY-5@users.noreply.github.com>

Fixes #3686

Author: SAY-5

rollout/trafficrouting.go

@@ -236,7 +236,17 @@ func (c *rolloutContext) reconcileTrafficRouting() error {
 			// But we can only increase canary weight according to available replica counts of the canary.
 			// we will need to set the desiredWeight to 0 when the newRS is not available.
 			if c.rollout.Spec.Strategy.Canary.DynamicStableScale {
-				desiredWeight = (weightutil.MaxTrafficWeight(c.rollout) * c.newRS.Status.AvailableReplicas) / *c.rollout.Spec.Replicas
+				// A Rollout scaled to zero replicas can still be fully promoted
+				// (see #3686), so guard the division here. With no spec replicas
+				// there is nothing to ratio against; leave desiredWeight at 0 so
+				// we don't shift traffic to a canary that has nowhere to land.
+				var specReplicas int32
+				if c.rollout.Spec.Replicas != nil {
+					specReplicas = *c.rollout.Spec.Replicas
+				}
+				if specReplicas > 0 {
+					desiredWeight = (weightutil.MaxTrafficWeight(c.rollout) * c.newRS.Status.AvailableReplicas) / specReplicas
+				}
 			} else if c.rollout.Status.Canary.Weights != nil {
 				desiredWeight = c.rollout.Status.Canary.Weights.Canary.Weight
 			}