[CONTINT-4838] Fix race in TestAutoscalerController (#49900)

### What does this PR do?

Fix the sporadically failing `pkg/util/kubernetes/apiserver/controllers.TestAutoscalerController` test on `main` by removing the race condition between `Create()` and the fake client's reflector setup.

Pre-loads the first `mockedHPA` into the fake client tracker (mirroring the pattern already used by the WPA sister test), so the informer's initial `List()` surfaces it via a delta and `AddFunc` fires from there — independent of whether the Watch is registered yet. Also fixes a latent bug where a single `time.NewTimer` was reused across three `select` blocks, and removes a `autoscalersListerSynced = true` override that bypassed the cache-sync barrier.

### Motivation

Tracked by [CONTINT-4838](https://datadoghq.atlassian.net/browse/CONTINT-4838).

Datadog test telemetry on `main` over the last 30 days:

| Metric | Value |
|---|---|
| Pass | 7562 |
| Fail | 35 |
| Failure rate | ~0.46% |
| avg(duration \| pass) | 581 ms |
| p95(duration \| pass) | 636 ms |
| p99(duration \| pass) | 649 ms |
| max(duration \| pass) | 840 ms |
| avg(duration \| fail) | 10 086 ms (= timeout) |

The distribution is **bimodal** — the test either passes in well under a second or sits at exactly 10 s and times out. There is no \"slow but eventually succeeds\" regime. All recent failures (Linux x64, Windows x64, macOS arm64) hit the same line — `hpa_controller_test.go:346`, the first `select` on `hctrl.autoscalers`, `<-timeout.C` branch.

Root cause: the test's `c.HorizontalPodAutoscalers(\"nsfoo\").Create(...)` runs in parallel with the reflector's `ListAndWatch` setup. If the test goroutine wins the race — i.e. `Create()` runs before the reflector calls `WatchWithContext` and registers a watcher on the fake client tracker — then the fake client's `tracker.add` broadcasts the ADDED event to **zero** watchers and silently drops it. The lister has the object (it lives in the tracker's `objects` map), but no `AddFunc` ever fires, so `hpaQueue` stays empty, the worker never writes to `hctrl.autoscalers`, and the `select` waits the full 10 s before failing.

### Why bumping the timeout (PR #49881) does not help

PR #49881 proposes raising the timeout from 10 s to 20 s. The bimodal distribution (max-pass 840 ms, fail 10 086 ms) makes that ineffective: there is no scenario where the test needs 12-15 s and is being denied; the failure mode is \"event lost forever.\" A higher timeout simply makes each CI failure cost 20 s instead of 10 s, while the flake rate stays the same. Closing #49881 in favor of this PR is recommended.

### Describe how you validated your changes

- All 37 tests in the package still pass: `dda inv test --targets=./pkg/util/kubernetes/apiserver/controllers`.
- 500/500 iterations of `TestAutoscalerController` pass with `-race`, ~633 ms each: `dda inv test --targets=./pkg/util/kubernetes/apiserver/controllers -e '^TestAutoscalerController\$' -r -x '-count=500 -timeout=15m'`. Given the pre-fix ~0.46% failure rate, ~2-3 failures would be expected across 500 runs; observed: 0.
- The fix mirrors the WPA sister test pattern (`wpa_controller_test.go:263`), which is already structurally race-free.

### Additional Notes

The latent timer bug (single `time.NewTimer(10s)` reused across three `select` blocks) is fixed at the same time: a `time.Timer` does not re-arm after firing, so the second/third selects could see an already-expired timer and fail in unrelated phases. Each select now uses an independent `time.After(timeoutDuration)` (5 s, matching the surrounding `EventuallyWithTf` calls).

The `autoscalersListerSynced = func() bool { return true }` override in `newFakeAutoscalerController` is removed because, with the HPA pre-loaded, the real `informer.HasSynced` flips to `true` shortly after `inf.Start` and the test now exercises the production cache-sync path in `runHPA`.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

[CONTINT-4838]: https://datadoghq.atlassian.net/browse/CONTINT-4838?atlOrigin=eyJpIjoiNWRkNTljNzYxNjVmNDY3MDlhMDU5Y2ZhYzA5YTRkZjUiLCJwIjoiZ2l0aHViLWNvbS1KU1cifQ

Co-authored-by: lenaic.huard <lenaic.huard@datadoghq.com>

Author: L3n41c

pkg/util/kubernetes/apiserver/controllers/hpa_controller_test.go

@@ -20,6 +20,7 @@ import (
 	autoscalingv2 "k8s.io/api/autoscaling/v2beta1"
 	corev1 "k8s.io/api/core/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/informers"
 	"k8s.io/client-go/kubernetes"
@@ -41,8 +42,8 @@ const (
 	hpaResource      = "horizontalpodautoscalers"
 )
 
-func newClient() kubernetes.Interface {
-	client := fake.NewSimpleClientset()
+func newClient(objects ...runtime.Object) kubernetes.Interface {
+	client := fake.NewSimpleClientset(objects...)
 	client.Resources = []*metav1.APIResourceList{
 		{
 			GroupVersion: autoscalingGroup + "/v2beta1",
@@ -58,8 +59,8 @@ func newClient() kubernetes.Interface {
 	return client
 }
 
-func newFakeConfigMapStore(t *testing.T, ns, name string, metrics map[string]custommetrics.ExternalMetricValue) (custommetrics.Store, kubernetes.Interface) {
-	client := newClient()
+func newFakeConfigMapStore(t *testing.T, ns, name string, metrics map[string]custommetrics.ExternalMetricValue, objects ...runtime.Object) (custommetrics.Store, kubernetes.Interface) {
+	client := newClient(objects...)
 	store, err := custommetrics.NewConfigMapStore(client, ns, name)
 	require.NoError(t, err)
 	err = store.SetExternalMetricValues(metrics)
@@ -104,8 +105,6 @@ func newFakeAutoscalerController(t *testing.T, client kubernetes.Interface, isLe
 	)
 	autoscalerController.enableHPA(client, informerFactory)
 
-	autoscalerController.autoscalersListerSynced = func() bool { return true }
-
 	return autoscalerController, informerFactory
 }
 
@@ -287,7 +286,20 @@ func TestAutoscalerController(t *testing.T) {
 
 	penTime := (int(time.Now().Unix()) - int(maxAge.Seconds()/2)) * 1000
 	name := custommetrics.GetConfigmapName()
-	store, client := newFakeConfigMapStore(t, "nsfoo", name, nil)
+
+	// Pre-load via the fake client tracker so the reflector's initial List
+	// surfaces the HPA. Creating after Start races with Watch registration
+	// and can drop the ADDED event.
+	mockedHPA := newFakeHorizontalPodAutoscaler(
+		"hpa_1",
+		"nsfoo",
+		"1",
+		"foo",
+		map[string]string{"foo": "bar"},
+	)
+	mockedHPA.Annotations = makeAnnotations("foo", map[string]string{"foo": "bar"})
+
+	store, client := newFakeConfigMapStore(t, "nsfoo", name, nil, mockedHPA)
 	metricName := "foo"
 	ddSeries := []datadog.Series{
 		{
@@ -322,27 +334,13 @@ func TestAutoscalerController(t *testing.T) {
 	c := client.AutoscalingV2beta1()
 	require.NotNil(t, c)
 
-	mockedHPA := newFakeHorizontalPodAutoscaler(
-		"hpa_1",
-		"nsfoo",
-		"1",
-		"foo",
-		map[string]string{"foo": "bar"},
-	)
-	mockedHPA.Annotations = makeAnnotations("foo", map[string]string{"foo": "bar"})
-
-	_, err := c.HorizontalPodAutoscalers("nsfoo").Create(context.TODO(), mockedHPA, metav1.CreateOptions{})
-	require.NoError(t, err)
-
 	timeoutDuration := 5 * time.Second
 	retryPeriod := 500 * time.Millisecond
-	timeout := time.NewTimer(10 * time.Second)
-	defer timeout.Stop()
 
 	select {
 	case key := <-hctrl.autoscalers:
 		t.Logf("hctrl process key:%s", key)
-	case <-timeout.C:
+	case <-time.After(timeoutDuration):
 		require.FailNow(t, "Timeout waiting for HPAs to update")
 	}
 
@@ -403,7 +401,7 @@ func TestAutoscalerController(t *testing.T) {
 	select {
 	case key := <-hctrl.autoscalers:
 		t.Logf("hctrl process key:%s", key)
-	case <-timeout.C:
+	case <-time.After(timeoutDuration):
 		require.FailNow(t, "Timeout waiting for HPAs to update")
 	}
 	storedHPA, err = hctrl.autoscalersLister.ByNamespace(mockedHPA.Namespace).Get(mockedHPA.Name)
@@ -453,7 +451,7 @@ func TestAutoscalerController(t *testing.T) {
 	select {
 	case key := <-hctrl.autoscalers:
 		t.Logf("hctrl process key:%s", key)
-	case <-timeout.C:
+	case <-time.After(timeoutDuration):
 		require.FailNow(t, "Timeout waiting for HPAs to update")
 	}