refactor: 审查修复 — netload 消除重复 /sys 读 + 全项目 gofmt

- netload_linux.go: SampleNetLoad 聚合循环不再对每接口重复
  readIfaceSpeed (snapshot 已汇总 cur.capMbps), 每次采样省 N 次 /sys 读
- gofmt -w: ring.go/ring_engine.go/auth.go/handlers.go/handlers_logs.go/
  handlers_users.go/store.go 结构体字段对齐与注释缩进
- README.md: markdownlint 自动修复 (MD028/MD040)

审查结论: 令牌环本身即互斥协议 — OnToken(收令牌)与 StartRing(发令牌)
在同一节点上由令牌串行化, 不存在需要加锁的竞争; WatchLeader 的读为
良性读, 无需 mutex

internal/cluster/netload_linux.go

@@ -196,8 +196,9 @@ func SampleNetLoad() float64 {
 	//
 	// rx+tx counts BOTH directions — a forward proxies in AND out, so a NIC
 	// that's saturated in one direction is saturated for forwarding purposes.
+	// Capacity comes from the CURRENT snapshot's capMbps (already summed in
+	// snapshotNet) — no extra /sys reads per interface here.
 	var deltaBits float64
-	sharedCap := uint64(0)
 	for name, c := range cur.bytes {
 		p, ok := prev.bytes[name]
 		if !ok {
@@ -209,15 +210,11 @@ func SampleNetLoad() float64 {
 			continue
 		}
 		deltaBits += float64(c-p) * 8
-		// approximate this interface's capacity contribution proportional to
-		// its share of the summed speed — read back from /sys would be noisy
-		// per call, so reuse the prior snapshot's capMbps scaling.
-		sharedCap += readIfaceSpeed(name)
 	}
-	if sharedCap == 0 {
+	if cur.capMbps == 0 {
 		return 0
 	}
-	capBitsPerSec := float64(sharedCap) * 1e6
+	capBitsPerSec := float64(cur.capMbps) * 1e6
 	pct := (deltaBits / dt) / capBitsPerSec * 100
 	if pct < 0 {
 		return 0

internal/cluster/ring.go

@@ -81,9 +81,9 @@ type TopoEntry struct {
 // State is the full cluster picture, replicated on every node:
 // ring membership + pending tasks + active forward topology.
 type State struct {
-	LeaderID   string        `json:"leaderId"`
-	Nodes      []Node        `json:"nodes"`
-	Cycle      int64         `json:"cycle"`
+	LeaderID string `json:"leaderId"`
+	Nodes    []Node `json:"nodes"`
+	Cycle    int64  `json:"cycle"`
 	// RoundDelay rides the token so every node converges on the same ring
 	// cadence (serialized as nanoseconds; the leader refreshes it each round
 	// via tkDelaySince). MUST be serialized — otherwise a token round-trip

internal/cluster/ring_engine.go

@@ -162,16 +162,17 @@ func (e *Engine) loadSnapshot() Load {
 
 // OnToken is the SINGLE-ROUND token handler. Per the authoritative design
 // (plan §令牌环协议): on receiving the token a node simultaneously
-//   (a) ADOPTS the carried cluster picture — incoming state is authoritative
-//       for membership + per-node fields. This is safe because structural
-//       changes (join/remove) are NOT kept in local state waiting to survive
-//       a merge: joins ride a separate pendingJoin channel injected INTO the
-//       token here, and a self-remove writes the node out of state so the
-//       downstream merge naturally drops it.
-//   (b) APPLIES the incremental log delta, then re-attaches own fresh entries.
-//   (c) APPENDS own node info (load/version) — skipped if we self-removed.
-//   (d) INJECTS pending newcomers right after ourselves + forwards to them.
-//   (e) Paces via a parallel rhythm timer (max(ops, timer)).
+//
+//	(a) ADOPTS the carried cluster picture — incoming state is authoritative
+//	    for membership + per-node fields. This is safe because structural
+//	    changes (join/remove) are NOT kept in local state waiting to survive
+//	    a merge: joins ride a separate pendingJoin channel injected INTO the
+//	    token here, and a self-remove writes the node out of state so the
+//	    downstream merge naturally drops it.
+//	(b) APPLIES the incremental log delta, then re-attaches own fresh entries.
+//	(c) APPENDS own node info (load/version) — skipped if we self-removed.
+//	(d) INJECTS pending newcomers right after ourselves + forwards to them.
+//	(e) Paces via a parallel rhythm timer (max(ops, timer)).
 func (e *Engine) OnToken(ctx context.Context, tk *Token) (*Token, error) {
 	// Multi-token guard: only the newest leader-stamped token is kept.
 	if tk.SentAt > 0 && e.lastTokenAt > 0 && tk.SentAt < e.lastTokenAt {