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/retention_test.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"github.com/openteam/server/internal/store"
+	"github.com/shopspring/decimal"
 )
 
 // 保留期只删过期明细，余额流水一条不动；未开启（0）时什么都不做。
@@ -18,13 +19,13 @@ func TestRetentionPurgesOldLogsOnly(t *testing.T) {
 	fresh := time.Now().UTC()
 	for i := 0; i < 5; i++ {
 		db.Create(&store.UsageLog{TraceID: fmt.Sprintf("old-%d", i), UserID: u, ModelName: "m1",
-			Cost: 0.01, Status: store.UsageStatusSuccess, CreatedAt: old})
+			Cost: decimal.NewFromFloat(0.01), Status: store.UsageStatusSuccess, CreatedAt: old})
 	}
 	for i := 0; i < 3; i++ {
 		db.Create(&store.UsageLog{TraceID: fmt.Sprintf("new-%d", i), UserID: u, ModelName: "m1",
-			Cost: 0.01, Status: store.UsageStatusSuccess, CreatedAt: fresh})
+			Cost: decimal.NewFromFloat(0.01), Status: store.UsageStatusSuccess, CreatedAt: fresh})
 	}
-	db.Create(&store.BalanceLog{UserID: u, Change: -0.05, BalanceAfter: 99.95,
+	db.Create(&store.BalanceLog{UserID: u, Change: decimal.NewFromFloat(-0.05), BalanceAfter: decimal.NewFromFloat(99.95),
 		Type: store.BalanceTypeUsage, RefID: "usage:old-0"})
 
 	n := purgeOlderThan(context.Background(), db, 30)
@@ -47,7 +48,7 @@ func TestRetentionDisabled(t *testing.T) {
 	db := openTestDB(t)
 	u := seedUser(t, db, 100)
 	db.Create(&store.UsageLog{TraceID: "x", UserID: u, ModelName: "m1",
-		Cost: 0.01, Status: store.UsageStatusSuccess,
+		Cost: decimal.NewFromFloat(0.01), Status: store.UsageStatusSuccess,
 		CreatedAt: time.Now().UTC().AddDate(0, 0, -400)})
 
 	stop := StartRetention(context.Background(), db, 0)
@@ -67,7 +68,7 @@ func TestRetentionBatches(t *testing.T) {
 	logs := make([]store.UsageLog, 0, retentionBatch+50)
 	for i := 0; i < retentionBatch+50; i++ {
 		logs = append(logs, store.UsageLog{TraceID: fmt.Sprintf("b-%d", i), UserID: u,
-			ModelName: "m1", Cost: 0.001, Status: store.UsageStatusSuccess, CreatedAt: old})
+			ModelName: "m1", Cost: decimal.NewFromFloat(0.001), Status: store.UsageStatusSuccess, CreatedAt: old})
 	}
 	if err := db.CreateInBatches(logs, 500).Error; err != nil {
 		t.Fatal(err)

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
 }
 

server/internal/usage/usage_test.go

@@ -7,6 +7,7 @@ import (
 	"time"
 
 	"github.com/openteam/server/internal/store"
+	"github.com/shopspring/decimal"
 	"gorm.io/gorm"
 )
 
@@ -75,11 +76,13 @@ func TestBillingDeductsBalanceAtomically(t *testing.T) {
 	m := seedModel(t, db)
 
 	rec := NewRecorder(db)
+	cost := decimal.NewFromFloat(0.15).Add(decimal.NewFromFloat(0.60)).
+		Mul(decimal.NewFromInt(1000)).Div(decimal.NewFromInt(1e6)).Round(8)
 	rec.Record(&store.UsageLog{
 		TraceID: "bill", UserID: u, ModelID: m, ModelName: "m1",
 		InputTokens: 1000, OutputTokens: 1000,
-		InputPrice: 0.15, OutputPrice: 0.60,
-		Cost:      0.15*1000/1e6 + 0.60*1000/1e6,
+		InputPrice: decimal.NewFromFloat(0.15), OutputPrice: decimal.NewFromFloat(0.60),
+		Cost:      cost,
 		Status:    store.UsageStatusSuccess,
 		LatencyMS: 5,
 		CreatedAt: time.Now().UTC(),
@@ -90,15 +93,15 @@ func TestBillingDeductsBalanceAtomically(t *testing.T) {
 	if err := db.First(&user, u).Error; err != nil {
 		t.Fatal(err)
 	}
-	want := 1 - (0.15*1000/1e6 + 0.60*1000/1e6)
-	if diff := user.Balance - want; diff > 1e-9 || diff < -1e-9 {
-		t.Fatalf("余额应扣到 %.10f，实际 %.10f", want, user.Balance)
+	want := decimal.NewFromInt(1).Sub(cost)
+	if !user.Balance.Equal(want) {
+		t.Fatalf("余额应精确扣到 %s，实际 %s", want, user.Balance)
 	}
 	var bl store.BalanceLog
 	if err := db.Where("user_id = ?", u).First(&bl).Error; err != nil {
 		t.Fatalf("余额流水应落库: %v", err)
 	}
-	if bl.Change >= 0 {
+	if !bl.Change.IsNegative() {
 		t.Fatalf("流水应为负数扣费，实际 %v", bl.Change)
 	}
 }
@@ -124,7 +127,7 @@ func openTestDB(t *testing.T) *gorm.DB {
 func seedUser(t *testing.T, db *gorm.DB, balance float64) uint64 {
 	t.Helper()
 	u := store.User{Username: "u" + fmt.Sprint(time.Now().UnixNano()), Email: "u@x.com",
-		PasswordHash: "x", Role: store.RoleUser, Balance: balance, Status: store.UserStatusActive}
+		PasswordHash: "x", Role: store.RoleUser, Balance: decimal.NewFromFloat(balance), Status: store.UserStatusActive}
 	if err := db.Create(&u).Error; err != nil {
 		t.Fatal(err)
 	}
@@ -133,7 +136,7 @@ func seedUser(t *testing.T, db *gorm.DB, balance float64) uint64 {
 
 func seedModel(t *testing.T, db *gorm.DB) uint64 {
 	t.Helper()
-	m := store.Model{Name: "m1", InputPrice: 0.15, OutputPrice: 0.60, Enabled: true}
+	m := store.Model{Name: "m1", InputPrice: decimal.NewFromFloat(0.15), OutputPrice: decimal.NewFromFloat(0.60), Enabled: true}
 	if err := db.Create(&m).Error; err != nil {
 		t.Fatal(err)
 	}
@@ -151,7 +154,7 @@ func TestBillingRefIDImpotency(t *testing.T) {
 	rec := NewRecorder(db)
 	mk := func(trace string) *store.UsageLog {
 		return &store.UsageLog{TraceID: trace, UserID: u, ModelName: "m1",
-			Status: store.UsageStatusSuccess, Cost: 1, CreatedAt: time.Now().UTC()}
+			Status: store.UsageStatusSuccess, Cost: decimal.NewFromInt(1), CreatedAt: time.Now().UTC()}
 	}
 	rec.Record(mk("dup"))
 	rec.Record(mk("dup"))
@@ -161,7 +164,7 @@ func TestBillingRefIDImpotency(t *testing.T) {
 	if err := db.First(&user, u).Error; err != nil {
 		t.Fatal(err)
 	}
-	if user.Balance != 4 {
+	if !user.Balance.Equal(decimal.NewFromInt(4)) {
 		t.Errorf("balance = %v, want 4 (重复 trace 不重复扣款)", user.Balance)
 	}
 	var n int64
@@ -182,7 +185,7 @@ func TestBillingAggregatesBatch(t *testing.T) {
 	for i := 0; i < n; i++ {
 		rec.Record(&store.UsageLog{
 			TraceID: fmt.Sprintf("agg-%d", i), UserID: u, ModelID: m, ModelName: "m1",
-			InputTokens: 10, OutputTokens: 5, Cost: 0.01,
+			InputTokens: 10, OutputTokens: 5, Cost: decimal.NewFromFloat(0.01),
 			Status: store.UsageStatusSuccess, CreatedAt: time.Now().UTC(),
 		})
 	}
@@ -192,7 +195,7 @@ func TestBillingAggregatesBatch(t *testing.T) {
 	if err := db.First(&user, u).Error; err != nil {
 		t.Fatal(err)
 	}
-	if want := 100 - 0.01*n; user.Balance-want > 1e-9 || user.Balance-want < -1e-9 {
+	if want := decimal.NewFromInt(100).Sub(decimal.NewFromFloat(0.01).Mul(decimal.NewFromInt(n))); !user.Balance.Equal(want) {
 		t.Fatalf("余额 = %v，期望 %v", user.Balance, want)
 	}
 	var ledgers int64
@@ -208,13 +211,13 @@ func TestBillingAggregatesBatch(t *testing.T) {
 	if daily[0].Requests != n {
 		t.Fatalf("日聚合 requests = %d，期望 %d", daily[0].Requests, n)
 	}
-	if d := daily[0].Cost - 0.01*n; d > 1e-9 || d < -1e-9 {
-		t.Fatalf("日聚合 cost = %v，期望 %v", daily[0].Cost, 0.01*n)
+	if want := decimal.NewFromFloat(0.01).Mul(decimal.NewFromInt(n)); !daily[0].Cost.Equal(want) {
+		t.Fatalf("日聚合 cost = %v，期望 %v", daily[0].Cost, want)
 	}
 	// 流水里的 balance_after 应与扣款后余额一致（批内累计）
 	var last store.BalanceLog
 	db.Where("user_id = ?", u).Order("id DESC").First(&last)
-	if last.BalanceAfter-user.Balance > 1e-9 || user.Balance-last.BalanceAfter > 1e-9 {
+	if !last.BalanceAfter.Equal(user.Balance) {
 		t.Fatalf("最后一条流水 balance_after = %v，实际余额 %v", last.BalanceAfter, user.Balance)
 	}
 }
@@ -226,7 +229,7 @@ func TestRecorderShardedByUser(t *testing.T) {
 	var users []uint64
 	for i := 0; i < 3; i++ {
 		u := store.User{Username: fmt.Sprintf("shard-u%d", i), Email: fmt.Sprintf("shard%d@x.com", i),
-			PasswordHash: "x", Role: store.RoleUser, Balance: 100, Status: store.UserStatusActive}
+			PasswordHash: "x", Role: store.RoleUser, Balance: decimal.NewFromInt(100), Status: store.UserStatusActive}
 		if err := db.Create(&u).Error; err != nil {
 			t.Fatal(err)
 		}
@@ -239,7 +242,7 @@ func TestRecorderShardedByUser(t *testing.T) {
 		for _, u := range users {
 			rec.Record(&store.UsageLog{
 				TraceID: fmt.Sprintf("u%d-%d", u, i), UserID: u, ModelName: "m1",
-				Cost: 0.001, Status: store.UsageStatusSuccess, CreatedAt: time.Now().UTC(),
+				Cost: decimal.NewFromFloat(0.001), Status: store.UsageStatusSuccess, CreatedAt: time.Now().UTC(),
 			})
 		}
 	}
@@ -258,7 +261,7 @@ func TestRecorderShardedByUser(t *testing.T) {
 			t.Fatalf("用户 %d 流水 %d 条，期望 %d", u, len(logs), per)
 		}
 		for i := 1; i < len(logs); i++ {
-			if logs[i].BalanceAfter > logs[i-1].BalanceAfter {
+			if logs[i].BalanceAfter.GreaterThan(logs[i-1].BalanceAfter) {
 				t.Fatalf("用户 %d 的流水余额应单调递减（顺序错乱）：%v → %v",
 					u, logs[i-1].BalanceAfter, logs[i].BalanceAfter)
 			}
@@ -266,6 +269,53 @@ func TestRecorderShardedByUser(t *testing.T) {
 	}
 }
 
+// 渠道×小时桶：状态计数含 error 行，token/成本只累计可计费行，跨小时分桶。
+func TestChannelHourlyBucketAggregated(t *testing.T) {
+	db := openTestDB(t)
+	u := seedUser(t, db, 100)
+	m := seedModel(t, db)
+
+	rec := NewRecorder(db)
+	now := time.Now().UTC()
+	hour := now.Format("2006-01-02T15")
+	prevHour := now.Add(-time.Hour).Format("2006-01-02T15")
+	mk := func(trace, status string, at time.Time, ch uint64) *store.UsageLog {
+		return &store.UsageLog{TraceID: trace, UserID: u, ModelID: m, ModelName: "m1",
+			ChannelID: ch, InputTokens: 100, OutputTokens: 50, Cost: decimal.NewFromFloat(0.02),
+			Status: status, LatencyMS: 200, CreatedAt: at}
+	}
+	rec.Record(mk("h1", store.UsageStatusSuccess, now, 7))
+	rec.Record(mk("h2", store.UsageStatusSuccess, now, 7))
+	rec.Record(mk("h3", store.UsageStatusError, now, 7))
+	rec.Record(mk("h4", store.UsageStatusSuccess, now.Add(-time.Hour), 7))
+	rec.Record(mk("h5", store.UsageStatusSuccess, now, 0)) // channel_id=0 不入桶
+	rec.Close()
+
+	var buckets []store.UsageChannelHourly
+	db.Where("channel_id = ?", 7).Order("hour DESC").Find(&buckets)
+	if len(buckets) != 2 {
+		t.Fatalf("应有两个小时桶，实际 %d", len(buckets))
+	}
+	cur, prev := buckets[0], buckets[1]
+	if cur.Hour != hour || prev.Hour != prevHour {
+		t.Fatalf("桶小时错误：%q / %q，期望 %q / %q", cur.Hour, prev.Hour, hour, prevHour)
+	}
+	if cur.Requests != 3 || cur.Success != 2 || cur.Errors != 1 {
+		t.Fatalf("当前桶计数 = %d/%d/%d，期望 3/2/1（requests/success/errors）", cur.Requests, cur.Success, cur.Errors)
+	}
+	// 只有 success 行计延迟
+	if cur.LatencySum != 400 || cur.LatencyMax != 200 {
+		t.Fatalf("延迟聚合 = sum %d max %d，期望 400/200", cur.LatencySum, cur.LatencyMax)
+	}
+	// token/成本只累计可计费行（2×success）
+	if cur.InputTokens != 200 || cur.Cost.String() != "0.04" {
+		t.Fatalf("token/成本 = %d/%s，期望 200/0.04", cur.InputTokens, cur.Cost)
+	}
+	if prev.Requests != 1 || prev.Success != 1 {
+		t.Fatalf("上一小时桶计数 = %d/%d，期望 1/1", prev.Requests, prev.Success)
+	}
+}
+
 // 分片函数：同一用户恒定同片，且覆盖所有分片。
 func TestShardOfStable(t *testing.T) {
 	rec := NewRecorder(openTestDB(t))