fix(orchestration): retry worker_done while the Orca runtime is briefly unreachable (#23984)

* fix(orchestration): retry worker_done while the Orca runtime is briefly unreachable

A worker reports worker_done once and ends its turn, so a few-minute app
outage silently stranded finished work at dispatched. The CLI now retries
worker_done on runtime_unavailable for about two minutes with backoff,
reusing one request id so the host's mutation ledger replays rather than
double-applies it, then prints the existing recovery command.

The contract probe no longer caches a failed status.get, which otherwise
made every retry fail without reaching the app.

Part of STA-8833.

* refactor(cli): pass the worker_done retry window in the mutation options bag

* Revert "refactor(cli): pass the worker_done retry window in the mutation options bag"

The options bag is forwarded to client.call as-is; the retry window is not a client.call option, and folding it in needed a value scan to keep the no-options call shape.

* test(cli): fold the explicit retry-request case and drop a vacuous timing assert

* fix(cli): keep worker_done recovery when the last retry fails before sending

Also skip the Unix-socket retry test on Windows and remove its temp profile.

src/cli/runtime/client.ts

@@ -245,7 +245,13 @@ export class RuntimeClient {
 
   private async ensureOrchestrationContractCompatible(timeoutMs: number): Promise<void> {
     if (!this.orchestrationContractCheck) {
-      this.orchestrationContractCheck = this.checkOrchestrationContractCompatibility(timeoutMs)
+      this.orchestrationContractCheck = this.checkOrchestrationContractCompatibility(
+        timeoutMs
+      ).catch((error: unknown) => {
+        // Why: a failed probe must not be cached, or a retry after a brief outage never reaches the app.
+        this.orchestrationContractCheck = null
+        throw error
+      })
     }
     await this.orchestrationContractCheck
   }