[lujiajing1126]

Fixes #13141

When a previously READY EDS cluster gets an update with no usable endpoints (e.g. an empty ClusterLoadAssignment, or all endpoints unhealthy), CdsLoadBalancer2 called delegate.handleNameResolutionError(). The priority/MultiChildLoadBalancer subtree ignores that while READY, so the stale picker kept selecting endpoints the control plane had removed. Before ClusterResolverLoadBalancer was merged into CdsLoadBalancer2 (ebb9420), this case published TRANSIENT_FAILURE and shut down the child policy.

Changes:

• Use fail() for the "No usable endpoint" case, so the cluster goes to TRANSIENT_FAILURE with an error picker and the child policy is shut down.
• In fail(), replace the shut-down GracefulSwitchLoadBalancer with a new one, the same way shutdown() does. Without this, a later non-empty update switches to the same priority factory, and GracefulSwitchLoadBalancer hands it to the child policy that was already shut down. This also applies to the existing fail() callers.

Ambient xDS errors are stored separately by XdsDependencyManager and don't reach this branch, so only the authoritative "no usable endpoints" case changes. In-flight RPCs aren't affected, since subchannel shutdown is graceful.

Added ClusterResolverLoadBalancerTest.handleEdsResource_allEndpointsRemovedAfterReady_failsAndRecovers. It covers three steps: READY, then an empty EDS update, which should give TF plus the error picker and shut down the child; then endpoints coming back, which should create a new child with the new addresses.

Note: I wasn't able to run the Gradle build locally (network restrictions in my environment), so I'm relying on CI for the test run. I did validate the fail() change for the empty-endpoint branch end to end against 1.85.1 with a fake ADS server: stale picks went from 20/20 to 0/20, the RPCs failed fast with UNAVAILABLE, and the cluster recovered once endpoints came back.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

No active review issues remain for the fix addressing stale READY pickers after usable endpoints are removed.

No open issues 🎉

What changed in this PR

Fixes #13141 by preventing xDS clusters from retaining a stale READY picker after an authoritative EDS update removes all usable endpoints.

Changes:

• Fail the cluster and shut down its child policy when no usable endpoints remain.
• Recreate the graceful-switch delegate after failure so later endpoint updates can recover.
• Add a regression test covering READY → empty endpoints → recovery with new addresses.

File	Description
xds/​src/​test/​java/​io/​grpc/​xds/​ClusterResolverLoadBalancerTest.java	Tests failure, child shutdown, and recovery after endpoints return.
xds/​src/​main/​java/​io/​grpc/​xds/​CdsLoadBalancer2.java	Uses fail() for empty endpoints and resets the delegate for recovery.

_{Review effort: Balanced}