fix(fork): make the fallback refresh await its own queued validation

Sources/SharedLiveAgentIndex.swift

@@ -663,6 +663,26 @@ final class SharedLiveAgentIndex {
             _ = await applyPendingForkValidations(
                 pendingRequestIDsToRemoveOnCancellation: pendingRequestIDsOwnedByRequest
             )
+            // The pass above may find this caller's request already claimed by
+            // another drainer: the unguarded tail restart in
+            // `applyPendingForkValidations` can spawn a detached refresh that
+            // wins the race against a contention waiter it just resumed, and
+            // that waiter then returns to an empty queue. Returning here would
+            // break this method's contract -- the queued validation must be
+            // applied before it returns -- so callers could read stale fork
+            // availability.
+            //
+            // This is a symptom fix, not the root cause. The root cause is that
+            // the tail restart in `applyPendingForkValidations` lacks the
+            // `!resumedWaiters` guard its in-loop sibling has, so it can resume
+            // a waiter and then immediately race it. Guarding it there is the
+            // real repair, but the obvious form can strand a pending request
+            // when the resumed waiter's task is cancelled right after resuming,
+            // so it needs its own change. The live-index branch below has the
+            // same hole when `didReload` is true -- `reload()` runs
+            // `applyPendingForkValidations` internally, so the same steal can
+            // happen and that path returns without waiting.
+            await waitForForkValidationRequestCompletions(pendingRequestIDsOwnedByRequest)
             return
         }
         let reloadResult = await reloadIfLiveAgentProcessFingerprintChanged(