azcosmos: cross-region retry policy for connection errors and 408s

sdk/data/azcosmos/CHANGELOG.md

@@ -8,6 +8,8 @@
 
 ### Bugs Fixed
 
+* Connection-error retry policy now attempts up to 3 retries against the current region before failing over, and performs at most one cross-region failover per call. Cross-region failover for writes only occurs when the error proves the request never reached the service (DNS, dial, TLS handshake, `ECONNREFUSED`, etc.); writes on ambiguous transport failures (e.g. `ECONNRESET`, `EOF`, transport-level timeouts) no longer fail over to another region, avoiding potential duplicate writes. Reads still fail over for any transport error. Caller-set context deadlines or cancellations short-circuit the policy without consuming the caller's budget with retries. See [PR 26858](https://github.com/Azure/azure-sdk-for-go/pull/26858).
+* HTTP `408 Request Timeout` responses are now handled by the Cosmos client retry policy: reads are retried exactly once against another region, and writes are returned to the caller immediately to avoid potential duplicates. See [PR 26858](https://github.com/Azure/azure-sdk-for-go/pull/26858).
 * Fixed excessive `GetDatabaseAccount` HTTP calls when using preferred regions, and stopped data-plane retries from trailing into the customer-supplied (default) endpoint once account topology is populated. See [PR 26815](https://github.com/Azure/azure-sdk-for-go/pull/26815).
 * Partition key range cache now serves concurrent callers from a single in-flight refresh per container, and the cached routing map remains readable while a refresh is in progress. The refresh runs on a detached background `context.Background()` so a caller's cancellation no longer aborts the shared fetch for other waiters; each caller continues to honor its own context deadline. See [PR 26855](https://github.com/Azure/azure-sdk-for-go/pull/26855).
 * Partition key range cache change-feed pagination is now resilient to mid-drain throttling. 429 responses are retried indefinitely (with capped linear backoff + jitter) since the service is explicitly asking the client to slow down, and the pages already accumulated are preserved instead of restarting the drain from page 1 on the next refresh. See [PR 26855](https://github.com/Azure/azure-sdk-for-go/pull/26855).

sdk/data/azcosmos/cosmos_client.go

@@ -198,8 +198,17 @@ func newClient(authPolicy policy.Policy, gem *globalEndpointManager, options *Cl
 				&globalEndpointManagerPolicy{gem: gem},
 			},
 			PerRetry: []policy.Policy{
-				authPolicy,
+				// clientRetryPolicy is listed before authPolicy so that
+				// its internal retry loop (continue inside Do) naturally
+				// re-invokes authPolicy via req.Next() on each attempt.
+				// This refreshes the Authorization header from the
+				// credential cache on every internal retry — important
+				// for short-lived AAD/MSI tokens that may expire across
+				// the cumulative same-region + cross-region retry
+				// window. Cosmos master/resource keys are unaffected
+				// because they are static.
 				&clientRetryPolicy{gem: gem},
+				authPolicy,
 			},
 			Tracing: azruntime.TracingOptions{
 				Namespace: "Microsoft.DocumentDB",