feat: 金额 decimal 精确计费、failover 按失败类型封顶、渠道小时桶预聚合、分层守护测试

- 计费：19 个金额字段 float64→decimal.Decimal，DB 列保持 numeric(20,8) 零迁移；
  API 金额仍为 JSON number（MarshalJSONWithoutQuotes），前端零改动
- failover：保留时间预算，新增 OT_PROXY_FAILOVER_CAPS_NET/SERVER/RATE/QUOTA
  按失败类型封顶换渠道次数（429/配额类瞬时失败时间预算形同虚设）
- 预聚合：新增 usage_channel_hourlies（schema v10），记账事务内同步 upsert，
  长窗口渠道统计走桶不再扫明细，spent_last_30d 改扫 usage_daily，v10 一次性回填幂等
- 守护：internal/arch_test.go 用 go/parser 强制分层依赖（fail-closed）

server/internal/usage/usage.go

@@ -8,6 +8,7 @@ import (
 	"time"
 
 	"github.com/openteam/server/internal/store"
+	"github.com/shopspring/decimal"
 	"gorm.io/gorm"
 	"gorm.io/gorm/clause"
 )
@@ -207,7 +208,7 @@ func (r *Recorder) bill(tx *gorm.DB, logs []*store.UsageLog) error {
 	billable := make([]*store.UsageLog, 0, len(logs))
 	refs := make([]string, 0, len(logs))
 	for _, l := range logs {
-		if (l.Status != store.UsageStatusSuccess && l.Status != store.UsageStatusCanceled) || l.Cost <= 0 {
+		if (l.Status != store.UsageStatusSuccess && l.Status != store.UsageStatusCanceled) || !l.Cost.IsPositive() {
 			continue
 		}
 		billable = append(billable, l)
@@ -250,9 +251,9 @@ func (r *Recorder) bill(tx *gorm.DB, logs []*store.UsageLog) error {
 	ledger := make([]store.BalanceLog, 0, len(billable))
 	for _, uid := range order {
 		rows := byUser[uid]
-		total := 0.0
+		total := decimal.Zero
 		for _, l := range rows {
-			total += l.Cost
+			total = total.Add(l.Cost)
 		}
 		var user store.User
 		if err := tx.First(&user, uid).Error; err != nil {
@@ -270,10 +271,10 @@ func (r *Recorder) bill(tx *gorm.DB, logs []*store.UsageLog) error {
 		// 流水里的 balance_after 是批内累计值，与逐行扣款时的语义一致
 		running := user.Balance
 		for _, l := range rows {
-			running -= l.Cost
+			running = running.Sub(l.Cost)
 			ledger = append(ledger, store.BalanceLog{
 				UserID:       uid,
-				Change:       -l.Cost,
+				Change:       l.Cost.Neg(),
 				BalanceAfter: running,
 				Type:         store.BalanceTypeUsage,
 				RefID:        usageRefID(l.TraceID),
@@ -312,7 +313,7 @@ func (r *Recorder) bill(tx *gorm.DB, logs []*store.UsageLog) error {
 		d.InputTokens += l.InputTokens
 		d.OutputTokens += l.OutputTokens
 		d.CacheReadTokens += l.CacheReadTokens
-		d.Cost += l.Cost
+		d.Cost = d.Cost.Add(l.Cost)
 	}
 	for _, k := range dailyOrder {
 		d := agg[k]
@@ -329,6 +330,72 @@ func (r *Recorder) bill(tx *gorm.DB, logs []*store.UsageLog) error {
 			return err
 		}
 	}
+
+	// 6. 渠道×小时桶：与日聚合同事务 upsert。长窗口（6h/24h/7d）的渠道统计从桶里
+	// 读取，不再扫 usage_logs（明细有保留期，桶是完整历史）。计费口径与日聚合一致：
+	// token/成本只累计可计费且未重复的行；状态计数（含 error）计全部行。
+	type chKey struct {
+		ch   uint64
+		hour string
+	}
+	chAgg := map[chKey]*store.UsageChannelHourly{}
+	chOrder := make([]chKey, 0, len(logs))
+	for _, l := range logs {
+		if l.ChannelID == 0 {
+			continue
+		}
+		k := chKey{ch: l.ChannelID, hour: l.CreatedAt.UTC().Format("2006-01-02T15")}
+		b, ok := chAgg[k]
+		if !ok {
+			b = &store.UsageChannelHourly{ChannelID: k.ch, Hour: k.hour}
+			chAgg[k] = b
+			chOrder = append(chOrder, k)
+		}
+		b.Requests++
+		switch l.Status {
+		case store.UsageStatusSuccess:
+			b.Success++
+			b.LatencySum += int64(l.LatencyMS)
+			if l.LatencyMS > b.LatencyMax {
+				b.LatencyMax = l.LatencyMS
+			}
+		case store.UsageStatusCanceled:
+			b.Canceled++
+		default:
+			b.Errors++
+		}
+		if l.Cost.IsPositive() && seen[usageRefID(l.TraceID)] {
+			b.InputTokens += l.InputTokens
+			b.OutputTokens += l.OutputTokens
+			b.CacheReadTokens += l.CacheReadTokens
+			b.Cost = b.Cost.Add(l.Cost)
+		}
+	}
+	for _, k := range chOrder {
+		b := chAgg[k]
+		// SQLite 的双参 MAX 是标量取大，PG 用 GREATEST
+		latMax := gorm.Expr("MAX(latency_max, ?)", b.LatencyMax)
+		if tx.Dialector != nil && tx.Dialector.Name() == "postgres" {
+			latMax = gorm.Expr("GREATEST(latency_max, ?)", b.LatencyMax)
+		}
+		if err := tx.Clauses(clause.OnConflict{
+			Columns: []clause.Column{{Name: "channel_id"}, {Name: "hour"}},
+			DoUpdates: clause.Assignments(map[string]any{
+				"requests":          gorm.Expr("requests + ?", b.Requests),
+				"success":           gorm.Expr("success + ?", b.Success),
+				"canceled":          gorm.Expr("canceled + ?", b.Canceled),
+				"errors":            gorm.Expr("errors + ?", b.Errors),
+				"input_tokens":      gorm.Expr("input_tokens + ?", b.InputTokens),
+				"output_tokens":     gorm.Expr("output_tokens + ?", b.OutputTokens),
+				"cache_read_tokens": gorm.Expr("cache_read_tokens + ?", b.CacheReadTokens),
+				"cost":              gorm.Expr("cost + ?", b.Cost),
+				"latency_sum":       gorm.Expr("latency_sum + ?", b.LatencySum),
+				"latency_max":       latMax,
+			}),
+		}).Create(b).Error; err != nil {
+			return err
+		}
+	}
 	return nil
 }