fix(grok): 保持模型级免费额度隔离

backend/internal/service/grok_p2_test.go

@@ -93,9 +93,9 @@ func TestApplyGrokUpstreamFailure_ModelSpecificFreeUsage(t *testing.T) {
 
 	svc.handleGrokAccountUpstreamError(context.Background(), account, 400, nil, body)
 
-	require.Equal(t, 1, repo.tempUnschedCalls)
-	require.Equal(t, "grok free usage exhausted", repo.lastTempUnschedReason)
+	require.Zero(t, repo.tempUnschedCalls, "model-scoped free usage must not cool sibling models")
 	require.True(t, isGrokModelQuotaBlocked(account.ID, "grok-4.5", time.Now()))
+	require.False(t, isGrokModelQuotaBlocked(account.ID, "grok-4.3", time.Now()))
 }
 
 func TestApplyGrokUpstreamFailure_SpendingLimitMarksReauth(t *testing.T) {

backend/internal/service/grok_upstream_failure.go

@@ -480,11 +480,10 @@ func (s *OpenAIGatewayService) applyGrokUpstreamFailureDecision(
 		if decision.Model != "" && isGrokModelSpecificFreeUsage(low, decision.Model) {
 			until := time.Now().Add(decision.Cooldown)
 			markGrokModelQuotaBlock(account.ID, decision.Model, until)
-			// Account-wide cool is still applied as a safety net when the body
-			// does not isolate a model cleanly; for model-specific free we only
-			// model-block and skip the long account cool when BlockModel was set
-			// by the classifier... actually free-usage keeps BlockModel false.
-			// Prefer model block + shorter account cool via tempUnschedule below.
+			// The upstream explicitly scoped exhaustion to this model, so an
+			// account-wide cool would incorrectly take healthy sibling models out
+			// of rotation.
+			return true
 		}
 		coolUntil := time.Now().Add(decision.Cooldown)
 		markGrokFreeUsageRecovery(account.ID, coolUntil)