feat: publish v1.0.0-rc.36-custom.2

Fixed-group tokens with a backup group now take precedence over the session affinity skip-retry-on-failure rule, so retryable upstream errors (503) fall through to the backup group instead of stopping on the primary channel. Non-retryable errors, explicit channel pins, and auto groups are unchanged.

Adds controller regression tests for the affinity override and a build-time Go test gate.

service/channel_affinity.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
 	"github.com/QuantumNous/new-api/model"
 	"github.com/QuantumNous/new-api/pkg/cachex"
 	"github.com/QuantumNous/new-api/relaykit/dto"
@@ -627,6 +628,17 @@ func ShouldSkipRetryAfterChannelAffinityFailure(c *gin.Context) bool {
 	if c == nil {
 		return false
 	}
+	// Explicit fixed-token fallback takes precedence over cache affinity.
+	primary := common.GetContextKeyString(c, constant.ContextKeyTokenGroup)
+	if primary != "" && primary != "auto" && common.GetContextKeyBool(c, constant.ContextKeyTokenCrossGroupRetry) {
+		if groups, ok := common.GetContextKeyType[[]string](c, constant.ContextKeyTokenAutoGroups); ok {
+			for _, group := range FilterUserTokenAutoGroups(common.GetContextKeyString(c, constant.ContextKeyUserGroup), groups) {
+				if group != primary {
+					return false
+				}
+			}
+		}
+	}
 	v, ok := c.Get(ginKeyChannelAffinitySkipRetry)
 	if ok {
 		b, ok := v.(bool)