fix: align CAPI readiness and convergence retry contracts

[reoring]

Summary

• read CAPI v1beta2 readiness conditions from canonical status.conditions
• make identical-error stall detection opt-in so ordinary Kubernetes eventual convergence runs to the action timeout
• retain explicit stallTimeout support for callers with a safe progress invariant

Root cause

Kest 0.18.0 had two runtime contract mismatches:

1. useCluster expected status.v1beta2.conditions, but a served cluster.x-k8s.io/v1beta2 Cluster exposes metav1 conditions at status.conditions.
2. The new global 30-second identical-error stall default treated a stable observation as proof of no progress. GatewayFleet legitimately kept RemoteConnectionProbe and rollout conditions unchanged for more than 30 seconds while external controllers converged, so both useCluster and ordinary 300-second assertions failed early.

Stall detection cannot infer external-system progress from an unchanged error string. It is now explicit opt-in.

Verification

• focused CAPI suite: 10 pass
• focused Biome check: 5 files
• tsc --noEmit
• full bun test ts/: 251 pass

All verification ran on monstera through the serialized test runner.