feat(api): add load-aware client fallback

apps/api/internal/runner/service.go

@@ -290,6 +290,9 @@ candidatesLoop:
 			break
 		}
 		decision := failoverDecisionForCandidate(runnerPolicy, candidate, candidateErr)
+		if !decision.Retry && hasLoadAvoidanceFallback(candidates, index, maxPlatforms) {
+			decision = loadAvoidanceFallbackDecision(candidateErr)
+		}
 		s.recordAttemptTrace(ctx, task.ID, attemptNo, failoverTraceEntry(decision))
 		if !decision.Retry {
 			break
@@ -792,6 +795,37 @@ func failoverTimeBudgetExceeded(start time.Time, maxDuration time.Duration) bool
 	return maxDuration > 0 && time.Since(start) >= maxDuration
 }
 
+func hasLoadAvoidanceFallback(candidates []store.RuntimeModelCandidate, index int, maxPlatforms int) bool {
+	if index < 0 || index >= len(candidates) || candidates[index].LoadAvoided {
+		return false
+	}
+	limit := len(candidates)
+	if maxPlatforms > 0 && maxPlatforms < limit {
+		limit = maxPlatforms
+	}
+	for next := index + 1; next < limit; next++ {
+		if candidates[next].LoadAvoided {
+			return true
+		}
+	}
+	return false
+}
+
+func loadAvoidanceFallbackDecision(err error) failoverDecision {
+	return failoverDecision{
+		Retry:  true,
+		Action: "next",
+		Reason: "load_avoidance_fallback",
+		Match: policyRuleMatch{
+			Source: "runtime_client_load",
+			Policy: "loadAvoidance",
+			Rule:   "fallback",
+			Value:  "loadRatio>=1",
+		},
+		Info: failureInfoFromError(err),
+	}
+}
+
 func normalizeRequest(kind string, body map[string]any) map[string]any {
 	out := cloneMap(body)
 	if kind == "responses" && out["messages"] == nil && out["input"] != nil {