The first debt repaid: reconciler.go, cited, against the packing you already watched

Artifact — the pod→endpoint mapping in staging/src/k8s.io/endpointslice/reconciler.go cited as file:line, verified live on the commit your clone is at, and matched against the four changes recorded in exercise 5. This is the first of P6's three owed citations.

Rests on — exercise 5, whose --max-endpoints-per-slice=5 is still in place and whose observations are the evidence this reading is checked against; and exercise 14 for the consumer's side of the same data.

Topology — pair, continued. Reading on forge, one confirmation run on the cluster.

Read — 28 KB, and the most readable substantial piece of networking code in the tree, which is why it is the one place in this phase where reading a whole file is the right call. Four things to come out with, each as a file:line:

Find	The question it settles
where a *v1.Pod becomes a discovery.Endpoint	which pod fields become which endpoint fields, and where Ready/Serving/Terminating are computed rather than copied
the function that decides which existing slice a new endpoint goes into	your change A: one slice touched, two untouched
whatever the code calls the "leave it partly full" decision	your change B: a slice left below the limit rather than repacked
where the limit itself is read	that the flag you set is read here and not in the controller

Verify the citations rather than trusting them, to P2's standard — these are staging-repo paths that move between releases, and the phase's own capstone says a copied line number rots fast:

ssh zain@10.10.10.125
cd ~/src/kubernetes && git log -1 --format='%H %ci'
kubectl version -o json | jq -r '.serverVersion.gitVersion'
git describe --tags --abbrev=0

If the clone and the cluster are not the same minor version, say so in the write-up beside every line number. That is not pedantry: reconciler.go moved into staging under KEP-3685 precisely so that other projects could vendor it, so it moves.

Do — score the four changes against the code. For each of A through D from exercise 5, name the function that produced the outcome you recorded, and mark the row explained or unexplained:

kubectl create ns pack2
kubectl -n pack2 create deployment idle --image=registry.k8s.io/pause:3.10 --replicas=13
kubectl -n pack2 expose deployment idle --port=80
kubectl -n pack2 get endpointslice -o json | jq -c '.items[]|{n:.metadata.name,count:(.endpoints|length)}'
kubectl -n kube-system logs -l component=kube-controller-manager --tail=40 | grep -i endpointslice

Then the one observation the reading predicts and exercise 5 did not produce: delete a slice by hand and watch the reconciler notice that its own output is missing.

kubectl -n pack2 delete endpointslice <one of the three>
sleep 3; kubectl -n pack2 get endpointslice -o json | jq -c '.items[]|{n:.metadata.name,count:(.endpoints|length)}'

Expect — a new slice with a new name, not the old one recreated, and the surviving slices untouched. Expect the total endpoint count to return to thirteen. That is level-triggered reconciliation on an object whose name is generated, and it is worth pausing on: the reconciler does not remember what it created, it compares what exists to what should exist — the same shape as P4's controllers and the same shape as your own CHECK in exercise 8.

Expect one of the four rows to be harder to explain than the others. Whichever it is, the honest write-up says unexplained with the counts attached rather than reaching for a plausible function name — that distinction is what makes the citation checkable.

Write down — the four file:line citations with the commit they were read at, the four changes marked explained or not, and the sentence P6 asked for: the EndpointSlice reconciler removes the endpoint — now with a line number under it. One debt down, two to go (exercise 16 and exercise 18).

Footprint note — thirteen pause pods again, under 50 MiB, on the worker. Note that they are getting their addresses from your CNI plugin now, which is a quiet piece of evidence for exercise 10 that costs nothing to collect.

Teardown — clean up, and put the control-plane flag back, because everything from exercise 16 onward should see a default cluster:

kubectl delete ns pack2 --wait=true
ssh zain@10.10.10.130 'sudo cp /root/kcm.yaml.orig /etc/kubernetes/manifests/kube-controller-manager.yaml'
sleep 20
kubectl -n kube-system get pod -l component=kube-controller-manager \
  -o jsonpath='{.items[0].spec.containers[0].command}' | tr ',' '\n' | grep -c max-endpoints    # must be 0

The topology stays.