fix(P2): 修复4个P2轻微问题

P2-01: 通配符scope安全风险 (scope_auth.go)
- 添加hasWildcardScope()函数检测通配符scope
- 添加logWildcardScopeAccess()函数记录审计日志
- 在RequireScope/RequireAllScopes/RequireAnyScope中间件中调用审计日志

P2-02: isSamePayload比较字段不完整 (audit_service.go)
- 添加ActionDetail字段比较
- 添加ResultMessage字段比较
- 添加Extensions字段比较
- 添加compareExtensions()辅助函数

P2-03: regexp.MustCompile可能panic (sanitizer.go)
- 添加compileRegex()安全编译函数替代MustCompile
- 处理编译错误，避免panic

P2-04: StrategyRoundRobin未实现 (router.go)
- 添加selectByRoundRobin()方法
- 添加roundRobinCounter原子计数器
- 使用atomic.AddUint64实现线程安全的轮询

P2-05: 错误信息泄露内部细节 - 已在MED-09中处理，跳过

gateway/internal/router/router_roundrobin_test.go

@@ -0,0 +1,51 @@
+package router
+
+import (
+	"context"
+	"testing"
+)
+
+// TestP2_04_StrategyRoundRobin_NotImplemented 验证RoundRobin策略是否真正实现
+// P2-04: StrategyRoundRobin定义了但走default分支
+func TestP2_04_StrategyRoundRobin_NotImplemented(t *testing.T) {
+	// 创建3个provider，都设置不同的延迟
+	// 如果走latency策略，延迟最低的会被持续选中
+	// 如果走RoundRobin策略，应该轮询选择
+	r := NewRouter(StrategyRoundRobin)
+	
+	prov1 := &mockProvider{name: "p1", models: []string{"gpt-4"}, healthy: true}
+	prov2 := &mockProvider{name: "p2", models: []string{"gpt-4"}, healthy: true}
+	prov3 := &mockProvider{name: "p3", models: []string{"gpt-4"}, healthy: true}
+	
+	r.RegisterProvider("p1", prov1)
+	r.RegisterProvider("p2", prov2)
+	r.RegisterProvider("p3", prov3)
+	
+	// 设置不同的延迟 - p1延迟最低
+	r.health["p1"].LatencyMs = 10
+	r.health["p2"].LatencyMs = 20
+	r.health["p3"].LatencyMs = 30
+	
+	// 选择100次，统计每个provider被选中的次数
+	counts := map[string]int{"p1": 0, "p2": 0, "p3": 0}
+	const iterations = 99  // 99能被3整除
+	
+	for i := 0; i < iterations; i++ {
+		selected, err := r.SelectProvider(context.Background(), "gpt-4")
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		counts[selected.ProviderName()]++
+	}
+	
+	t.Logf("Selection counts with different latencies: p1=%d, p2=%d, p3=%d", counts["p1"], counts["p2"], counts["p3"])
+	
+	// 如果走latency策略，p1应该几乎100%被选中
+	// 如果走RoundRobin，应该约33% each
+	
+	// 严格检查：如果p1被选中了超过50次，说明走的是latency策略而不是round_robin
+	if counts["p1"] > iterations/2 {
+		t.Errorf("RoundRobin strategy appears to NOT be implemented. p1 was selected %d/%d times (%.1f%%), which indicates latency-based selection is being used instead.", 
+			counts["p1"], iterations, float64(counts["p1"])*100/float64(iterations))
+	}
+}