fix: WebSocket reconnect stability — eliminate kick storms and race conditions

[Jerry-Xin]

Problem

Multiple WebSocket reconnect bugs caused persistent kick storms (~950 kicked/hour across 15 bots) when connecting to WuKongIM. The root cause was a cascade of race conditions between the transport layer (socket.ts) and the application layer (channel.ts).

Fixes #135, #139

Root Cause Analysis

9 confirmed issues found via Codex + Claude Code cross-review:

Critical (direct cause of kick loops)

1. Stale socket events — open/message/error handlers lacked if (this.ws !== ws) return guard, allowing old WS DISCONNECT packets to corrupt new connection state
2. Zombie WebSocket connections — CONNACK rejection and DISCONNECT handlers didn't call ws.close(), leaving TCP connections alive
3. Heartbeat timer not stopped on disconnect — HTTP heartbeat timer continued running after WS disconnect, triggering competing reconnections with stale credentials
4. Ping timeout dual reconnect — Ping timeout called onDisconnected + scheduleReconnect() but not onError, so channel layer didn't know to stop heartbeat timer

Race condition fixes

5. Synchronous disconnect+connect — ws.close() is async; new WS opened before old TCP FIN completed, causing server to see 2 connections briefly
6. Token refresh before disconnect — registerBot(forceRefresh) called while old WS still alive during await
7. Heartbeat failure zero backoff — Immediate disconnect()+connect() bypassed exponential backoff
8. Heartbeat failure counter not reset — consecutiveHeartbeatFailures preserved across reconnects
9. Dual reconnect ownership — Both socket.ts and channel.ts independently drove reconnects

Changes

socket.ts

• Add stale guards (if (this.ws !== ws) return) to open, message, error handlers
• Add ws.close(); this.ws = null in CONNACK rejection and DISCONNECT handlers
• Add disconnectAndWait(timeoutMs=2000) — waits for WS close event or terminates on timeout
• Make stopReconnectTimer() public for cross-layer coordination
• Add stable timer, rapid disconnect detection, exponential backoff with jitter

channel.ts

• Clear heartbeat timer in onDisconnected; restart in onConnected
• Replace all disconnect()+connect() with await disconnectAndWait() + connect()
• Reorder token refresh: disconnect first, then registerBot(forceRefresh)
• Add 3-5s backoff delay on heartbeat failure reconnect
• Reset consecutiveHeartbeatFailures in onConnected
• Call stopReconnectTimer() before connect() in cooldown/heartbeat paths
• Deduplicate cooldown reconnect timer to prevent self-kick storms
• Add [accountId] to all log messages for per-bot diagnostics

Tests

• 20 new tests covering all 9 fixes in reconnect-fixes.test.ts
• All 284 tests pass (264 existing + 20 new)

Verification

Deployed to 3 nodes with 30+ bots total:

• Before: ~950 kicked/hour sustained, 170K+ total kicks
• After: 0 kicked across all nodes (stable for 15+ minutes)

[yujiawei]

LGTM. 9 个 WS 重连竞态修复覆盖完整，stable timer + rapid disconnect 检测设计合理，20 新测试 + 生产验证 (950 kicked/h → 0). CI ✅