AMassBattleNetworkGameModeBase

设置 bReplicates=true 并把 GameStateClass 指到 AMassBattleNetworkGameStateBase。DefaultPawnClass 有意不在此设置——在 BP GameMode 中配置为携带 UMassBattleAgentComponent 的 ACharacter 派生 Pawn。

Sets bReplicates=true and points GameStateClass at AMassBattleNetworkGameStateBase. DefaultPawnClass is intentionally unset here — configure it in the BP GameMode to an ACharacter-based pawn carrying UMassBattleAgentComponent.

先调用父类,随后当 bEnableNetworkSync 开启时提前置 MB->bNetworkedMode=true。InitGame 先于一切 Actor BeginPlay 执行——关卡摆放 Actor 的 BeginPlay(组件按会话标志路由 + 模板工厂按 bNetworkedMode 盖 FNetworkTag)不再与 BeginPlay 时的置位竞态;BeginPlay 中的置位保留(幂等)。

Calls the parent first, then — when bEnableNetworkSync is on — flips MB->bNetworkedMode=true early. InitGame runs BEFORE any actor BeginPlay, so level-placed actors' BeginPlay (component routing on the session flag + template-factory FNetworkTag stamping off bNetworkedMode) never races the flag; the BeginPlay-time set is kept idempotent.

初始化:先把插件自有的模糊测试 tag MB.FuzzTask 注册进客户端命令 allow 集(服务器来源但走标准准入归一化);把 DefaultPawnClass 同步到 GameState 的 PlayerCharacterClass(供客户端/快照访问角色类);当 bEnableNetworkSync 开启时——置 MB->bNetworkedMode=true(幂等)、确保 TCP 传输(EnsureSnapshotTcpTransport)、绑定 OnTickCompleted → OnSimTickCompleted、创建 UMassBattleNetworkRelay 并按 GetNetMode() 初始化为 DedicatedServer 或 ListenServer 模式;随后输出命令闸门自检日志([CMD_WHITELIST]:生效模式与 allow 集内容;Enforce 且 allow 集为空时输出 Warning——所有客户端命令都将被拒绝)。关闭时无网络初始化。

Initialization: first allow-lists the plugin's own fuzz tag MB.FuzzTask (server-originated but flowing through standard admission normalization); syncs DefaultPawnClass into the GameState's PlayerCharacterClass (for clients/snapshot); when bEnableNetworkSync is on — sets MB->bNetworkedMode=true (idempotent), ensures the TCP transport (EnsureSnapshotTcpTransport), binds OnTickCompleted → OnSimTickCompleted, and creates UMassBattleNetworkRelay initialized to DedicatedServer or ListenServer mode per GetNetMode(); then dumps the client command gate self-check ([CMD_WHITELIST]: effective mode and allow-set contents; Enforce with an empty allow set logs a Warning — every client command will be rejected). No network init when disabled.

玩家登录处理:分配确定性 PlayerSlot(AllocatePlayerSlot),初始化 ack 追踪(ClientAckedTicks 置当前 tick,ClientMaxObservedLag 置 0)。远程后加入者(非 ListenServer 本地控制器)标记 bForceSnapshotNextTick=true 并加入 PendingLateJoinClients,在确定性 post-command 边界做快照优先接入;监听服务器主机共享服务端模拟,从不后加入。

Player login handling: allocates a deterministic PlayerSlot (AllocatePlayerSlot) and initializes ack tracking (ClientAckedTicks to the current tick, ClientMaxObservedLag to 0). Remote late-joiners (not the listen-server local controller) get bForceSnapshotNextTick=true and join PendingLateJoinClients for snapshot-first onboarding at the deterministic post-command boundary; the listen-server host shares the server sim and never late-joins.

玩家登出:经 possess 链捕获退出 pawn 拥有的每个 agent UID(1:N),按 UID 广播 Player.Leave 锁步指令(所有端确定性销毁各 agent);清理每玩家追踪(CleanupPlayerState);调用引擎登出后释放确定性槽位(FreePlayerSlot)。

Player logout: captures every agent UID owned by the exiting pawn (1:N) via the possess chain, broadcasts one Player.Leave lockstep command per UID (all peers destroy each agent deterministically); clears per-player tracking (CleanupPlayerState); after the engine logout, frees the deterministic slot (FreePlayerSlot).

服务器向全体客户端广播一条锁步指令。自动分流:序列化负载 ≥ MassBattle.AutoFloodSwitchBytes(默认 16KB)时,把负载打包压缩为 TCP 数据块(BroadcastLockstepDataBlock)并广播轻量 UDP 引用命令(载荷只含 DataBlockId/DataBlockCrc),绕过 64KB RPC bunch 上限;否则走现有可靠多播 RPC。两种路径都经 NextServerExecSequence 分配执行序号并写入服务端自身命令队列(EnqueueNetCommand),ExecuteTick = 当前 tick + K + ExtraDeferTicks。命令标签如 AgentComponent.Spawn/Obstacle.Spawn 等由 ExecuteNetCommand 解释。

Broadcasts one lockstep command from the server to all clients. Auto-flood: when the serialized payload ≥ MassBattle.AutoFloodSwitchBytes (default 16KB), the payload is packed/compressed into a TCP data block (BroadcastLockstepDataBlock) plus a lightweight UDP reference command (payload carries only DataBlockId/DataBlockCrc), bypassing the 64KB RPC bunch limit; otherwise the existing reliable multicast RPC is used. Both paths assign an execution sequence via NextServerExecSequence and enqueue into the server's own command queue (EnqueueNetCommand), with ExecuteTick = current tick + K + ExtraDeferTicks. Command tags such as AgentComponent.Spawn/Obstacle.Spawn are interpreted by ExecuteNetCommand.

定向单播:按 PlayerIndex 找到目标 AMassBattleNetworkPlayerControllerBase 并调用 ClientReceiveLockstepCommand 只发给该客户端。ExecuteTick = 当前 tick + K,同样分配执行序号并写入服务端自身队列。

Targeted unicast: locates the AMassBattleNetworkPlayerControllerBase by PlayerIndex and delivers via ClientReceiveLockstepCommand to that client only. ExecuteTick = current tick + K, with sequence assignment and server self-enqueue as well.

导出流场基底(场自身的数据结构——Grid 格子 / NavMesh 图,由 FlowFieldCanvas 紧凑量化)并锁步广播到全端,使各端本地计算路径、引擎 navmesh 查询链彻底退出锁步路径(NavMesh-A* 失步修复)。导出后服务器自导入同一份字节(KI-011:服务器与客户端必须进入同一同步基底态,否则两套 A* 回答同一查询即失步);首次基底发送时订阅 navmesh 重建委托(动态 navmesh 自动重导出)。差分同步:SyncMode 选择线上形态——Full = 每次整份基底;Incremental = 优先对缓存基准发差分(变更格子/tile + 目标集),无基准或差分不可构造时自动回退全量;Auto(默认)= 服务器择小发送(差分达到 MassBattle.FF.DeltaToFullRatio × 全量即弃用,默认 0.75)。缓存永远持有最新全量导出——同时充当差分基准与晚加入补发源;真无变动时整次跳过(无 TCP 块、无命令)。命令标签为 FF.RebuildFlowField;重发幂等(内容哈希跳过)。

Exports a flow field's substrate (the field's own data structure — Grid cells / NavMesh graph, compactly quantized by FlowFieldCanvas) and lockstep-broadcasts it to every peer so all peers compute paths locally and the engine navmesh query chain leaves the lockstep path entirely (the NavMesh-A* desync fix). After export the server self-imports the same bytes (KI-011: server and clients must enter the same synced-substrate state, otherwise two different A* implementations answer the same query and desync); the first substrate send subscribes to the navmesh rebuild delegate (dynamic-navmesh auto re-export). Delta sync: SyncMode selects the wire form — Full = entire substrate every time; Incremental = prefer a delta (changed cells/tiles + goal set) against the cached base, auto-falling back to full when no base exists or the delta can't be built; Auto (default) = the server sends whichever is smaller (a delta at or above MassBattle.FF.DeltaToFullRatio × full is rejected, default 0.75). The cache always holds the latest FULL export — it doubles as the delta base AND the late-join re-delivery source; a true no-change dispatch skips entirely (no TCP block, no command). The command tag is FF.RebuildFlowField; re-delivery is idempotent (content-hash skip).

字节感知的锁步广播核心——恒走 auto-flood TCP 块路径(字节负载为基底级体量),镜像 SendLockstepCommandToClients 的自动分流分支:Pack→压缩→分配 ServerBlockId 自存并广播块,再多播轻量 UDP 引用命令(载荷只含 DataBlockId/DataBlockCrc),ExecuteTick = 当前 tick + K + ExtraDeferTicks。上面的公开便捷函数不带字节;本核带字节。

Byte-aware lockstep broadcast core — always rides the auto-flood TCP block path (byte payloads are substrate-sized), mirroring the auto-flood branch of SendLockstepCommandToClients: Pack → compress → allocate a ServerBlockId, self-store and broadcast the block, then multicast a lightweight UDP reference command (payload carries only DataBlockId/DataBlockCrc), with ExecuteTick = current tick + K + ExtraDeferTicks. The public helpers above pass no bytes; this is the byte-capable core.

差分修复入口:已应用基底版本脱节的端(跳过差分的晚加入端、分叉链)请求该场的显式全量重发——ServerRequestFlowFieldFullResync 的服务器侧对应(由 PlayerController RPC 与 GameStateBase 的拒收链调用)。按场墙钟节流(2 秒窗口,异常端的修复风暴收敛为每窗口一次全量重发)。非锁步命令:基底修复本身带版本且幂等,无需定序。

Delta-repair entry: a peer whose applied substrate version missed a delta's base (skipped-delta late joiner, forked chain) asks for an explicit FULL re-send of that field — the server-side counterpart of ServerRequestFlowFieldFullResync (called from the PlayerController RPC and the GameStateBase rejection chain). Wall-clock throttled per field (2s window; repair storms from misbehaving peers collapse to one full re-send per window). Not a lockstep command: substrate repair is versioned + idempotent and needs no ordering.

动态 navmesh 自动重导出:服务器 navmesh(RuntimeGeneration Dynamic / Dynamic-ModifiersOnly)重建完成时,把全部已同步 NavMesh 模式场的基底从活体重导出并重播——全端(含服务器)在命令 ExecuteTick 原子换版,零发散窗口(同步图换版前保持旧基底在线)。墙钟节流(2 秒):中间重建收敛为最新导出。Grid 模式场不参与——其基底源自场景探测而非 navmesh,重建/重导出保持显式(BP)。委托在首次基底发送时绑定。

Dynamic-navmesh auto re-export: when the server navmesh (RuntimeGeneration Dynamic / Dynamic-ModifiersOnly) finishes a rebuild, every synced NavMesh-mode field's substrate is re-exported from the live mesh and re-broadcast — all peers (server included) swap atomically at the command's ExecuteTick, zero divergence window (the synced graph stays on the old substrate until the swap). Wall-clock throttled (2s): intermediate rebuilds collapse to the latest export. Grid-mode fields are excluded — their substrate derives from scene traces, not the navmesh; their rebuild/re-export stays explicit (BP). The delegate is bound on the first substrate send.

判断 Tag 是否为服务器权威命令 tag——从 MB_FOR_EACH_SERVER_AUTHORITY_TAG 注册表展开(静态集合,可安全供 CheatManager/诊断复用)。客户端来源命令携带这些 tag 恒被拒绝:AgentComponent.Spawn / SetAgentTransform / Despawn、ProjectileComponent.Spawn / Despawn、LootComponent.Spawn / Despawn、Obstacle.Spawn / SyncTransform / Destroy、FF.RebuildFlowField、Game.Pause / Resume、Player.Leave。特权意图必须走服务器 API(SendLockstepCommandToClients)。

Whether Tag is a server-authority command tag — expanded from the MB_FOR_EACH_SERVER_AUTHORITY_TAG registry (a static set, safe for CheatManager/diagnostics reuse). Client-originated commands carrying these tags are always rejected: AgentComponent.Spawn / SetAgentTransform / Despawn, ProjectileComponent.Spawn / Despawn, LootComponent.Spawn / Despawn, Obstacle.Spawn / SyncTransform / Destroy, FF.RebuildFlowField, Game.Pause / Resume, Player.Leave. Privileged intents must go through the server API (SendLockstepCommandToClients).

运行时把游戏侧客户端命令 tag 注册进 allow 集(仅服务器,留审计日志)。拒绝服务器权威 tag——其永不允许客户端发起,注册会打 Error 并跳过(放行一个就重新打开本闸门要封的洞)。优先在 BP GameMode 默认值填写 ClientAllowedCommandTags;动态 tag 用本函数。重复注册幂等。

Runtime-registers a game-side client-issued command tag into the allow set (server only, logged). Refuses server-authority tags — they can never be client-issued; registering one logs an Error and is skipped (allowing one would reopen the exact hole this gate closes). Prefer filling ClientAllowedCommandTags in BP GameMode defaults; use this for dynamic tags. Re-registration is idempotent.

运行时从 allow 集移除 tag(仅服务器,留审计日志)。已调度的在途命令不受影响。

Runtime-removes a tag from the allow set (server only, logged). In-flight commands already scheduled are unaffected.

服务器权威直通入队(自动失步取证模糊器专用,由 MassBattleDesyncAutoDebugger 调用):按 ExecuteTick 盖 ServerExecSequence 后直接进入 PendingCommands——与内置服务器命令同款路径,绕过客户端准入闸门(模糊器是服务器来源)。

Server-authority direct enqueue (for the auto-forensics desync fuzzer, called by MassBattleDesyncAutoDebugger): stamps a ServerExecSequence for the command's ExecuteTick and enters PendingCommands — the same path as the built-in server commands, bypassing the client admission gate (the fuzzer is server-originated).

客户端上行命令的服务器入口(Server_SendNetCommand 的落地,也是客户端命令闸门的唯一执法点)。准入顺序:① 客户端命令闸门——非 Disabled 模式时,服务器权威 tag 或不在 ClientAllowedCommandTags 的命令记拒绝日志(有界诊断)并广播 CommandRejectedByWhitelist 网络事件,Enforce 模式直接丢弃;② 隔离中的客户端(ClientsRejectedUntilResync)丢弃并打 Warning;③ 每玩家每 tick 准入配额(MaxClientCommandsPerTick,超限丢弃——洪泛客户端在入口即被丢弃,不被 N 倍放大给所有端);④ 待广播队列硬上限(MaxPendingCommands,聚合尖峰保护);⑤ 客户端可控调度输入钳制——ExtraDeferTicks 钳到 [0, MaxExtraDeferTicks]、IssuedTick 钳到服务器 tick 周围窗口([LastKnownServerTick-1024, +256])。之后 K-buffer 调度:ExecuteTick = max(LastKnownServerTick, Cmd.IssuedTick, 最快已 ack 客户端 tick 与 LastKnownServerTick+K 的较小者) + K + ExtraDeferTicks,再分配 ServerExecSequence。带 DataBlockId 的命令走 ConvergeUplinkBlock 汇聚。

Server entry for client-uplink commands (the landing site of Server_SendNetCommand, and the single enforcement point of the client command gate). Admission order: ① client command gate — when not Disabled, commands with a server-authority tag or absent from ClientAllowedCommandTags get a bounded rejection log plus a CommandRejectedByWhitelist network event broadcast, and are dropped outright in Enforce mode; ② quarantined clients (ClientsRejectedUntilResync) are dropped with a Warning; ③ per-player per-tick admission quota (MaxClientCommandsPerTick, overflow dropped — a flooding client is dropped at entry instead of being amplified N-fold to every peer); ④ pending-broadcast queue hard cap (MaxPendingCommands, aggregate spike protection); ⑤ client-controlled schedule inputs are clamped — ExtraDeferTicks to [0, MaxExtraDeferTicks] and IssuedTick to a window around the server tick ([LastKnownServerTick-1024, +256]). Then K-buffer scheduling: ExecuteTick = max(LastKnownServerTick, Cmd.IssuedTick, min(fastest acked client tick, LastKnownServerTick+K)) + K + ExtraDeferTicks, then a ServerExecSequence is assigned. Commands carrying a DataBlockId route through ConvergeUplinkBlock.

汇聚上行命令的 TCP 块:从 Net 的已上传块缓存按 (PlayerIndex, LocalBlockId) 取块;块未到则把命令暂存(StagePendingBlockCommand),块到达回调(OnLockstepDataBlockUploaded)时再汇聚(竞态修复)。块已到时:重新计算 CRC、压缩、分配权威 ServerBlockId(AllocServerBlockId)自存并向所有客户端广播块,命令的 DataBlockId 重映射为 ServerBlockId,消费上行块后入队 PendingCommands。

Converges an uplink command's TCP block: fetches the block from Net's uploaded-block cache by (PlayerIndex, LocalBlockId); if the block is missing, the command is staged (StagePendingBlockCommand) and converged when the block arrives (via OnLockstepDataBlockUploaded) — a race fix. When the block is present: recomputes CRC, compresses, allocates the authoritative ServerBlockId (AllocServerBlockId), self-stores and broadcasts the block to all clients, remaps the command's DataBlockId to ServerBlockId, consumes the uplink block, and enqueues into PendingCommands.

锁步命令执行钩子。_Implementation 默认空实现——游戏特定命令在 BP 或派生 C++ 中覆写。网络会话中,单位任务(AgentsMoveTo、AgentsChaseAttack 等)必须在此钩子内创建,使各端确定性地构建——本地创建的任务只写入本端可见的 fragment,会导致失步。内置命令标签(AgentComponent.Spawn 等)由 AMassBattleNetworkGameStateBase::ExecuteNetCommand 统一执行,不经此钩子。

Lockstep command execution hook. The _Implementation is empty by default — game-specific commands override it in BP or derived C++. In networked sessions, agent tasks (AgentsMoveTo, AgentsChaseAttack, ...) must be created inside this hook so every peer builds them deterministically — creating them locally writes fragments only this peer can see and causes a desync. Built-in command tags (AgentComponent.Spawn etc.) are executed centrally by AMassBattleNetworkGameStateBase::ExecuteNetCommand, not through this hook.

重同步结果钩子。插件在 _Implementation 提供默认状态处理:失败时把客户端移入 ClientsRejectedUntilResync(隔离);成功时无额外动作(延迟 Agent.Spawn 已由 OnClientResyncComplete → FlushDeferredLockstepSpawnsForOwner 冲刷,快照优先保证)。BP 覆写以接入游戏特定生成状态机。

Resync outcome hook. The plugin's _Implementation provides default state handling: on failure the client is moved into ClientsRejectedUntilResync (quarantine); on success nothing extra is done (deferred Agent.Spawn is already flushed by OnClientResyncComplete → FlushDeferredLockstepSpawnsForOwner — the snapshot-first guarantee). Override in BP for game-specific spawn state machines.

游戏特定生成超时检查,在 CheckPacingTimeouts 末尾被调用。BP 实现;默认空操作。

Game-specific spawn-timeout check, called at the end of CheckPacingTimeouts. Implement in BP; no-op by default.

是否有生成正阻塞重同步。BP 实现;默认 false。被 ProcessPendingResyncs / HandleClientResyncRequest 用于在生成进行中推迟 resync。

Whether a spawn is currently blocking resync. Implement in BP; default false. Used by ProcessPendingResyncs / HandleClientResyncRequest to defer resync while spawning.

查询玩家是否处于待 resync 状态:PendingLateJoinClients 或 ClientsWithPendingResync 含该玩家即返回 true。被 ProcessPendingLockstepSpawns 用于延迟待 resync 玩家的 Agent.Spawn(快照优先保证)。

Queries whether a player is awaiting resync: true when the player is in PendingLateJoinClients or ClientsWithPendingResync. Used by ProcessPendingLockstepSpawns to defer a resync-pending player's Agent.Spawn (snapshot-first guarantee).

服务器端 AgentComponent 在 BeginPlay 自注册(AddUnique)到 PendingComponentSpawns。ProcessPendingLockstepSpawns 在确定性 tick 边界冲刷:属主就绪(IsLockstepOwnerReady)且不在 resync 中时调用 RequestLockstepSpawn。

Server-side AgentComponent self-registers (AddUnique) into PendingComponentSpawns on BeginPlay. ProcessPendingLockstepSpawns flushes at the deterministic tick boundary: RequestLockstepSpawn fires when the owner is ready (IsLockstepOwnerReady) and not awaiting resync.

ProjectileComponent 驱动的锁步生成注册(仅服务器)。无玩家门控(投射物无属主槽位),在同一 tick 边界通道冲刷。

Registers a ProjectileComponent-driven lockstep spawn (server only). No player gating (projectiles have no owner slot); flushed in the same tick-boundary pass.

LootComponent 驱动的锁步生成注册(仅服务器)。无玩家门控,同一 tick 边界通道冲刷。

Registers a LootComponent-driven lockstep spawn (server only). No player gating; flushed in the same tick-boundary pass.

为指定玩家立即冲刷其延迟的 Agent.Spawn:遍历 PendingComponentSpawns,属主匹配的组件直接 RequestLockstepSpawn,其余保留。由 OnClientResyncComplete 调用——快照优先保证:客户端拥有世界状态后才触发其生成。

Immediately flushes one player's deferred Agent.Spawn: iterates PendingComponentSpawns, firing RequestLockstepSpawn for matching owners and keeping the rest. Called from OnClientResyncComplete — the snapshot-first guarantee: a client's spawn fires only after it holds the world state.

为玩家分配确定性槽位:优先复用 FreeSlots,否则取 NextPlayerSlot++;写入双向映射,并清除 TCP transport 的 freed 守卫(ReusePlayerSlot)使该槽位重新成为合法认证目标。

Assigns a deterministic slot: reuses FreeSlots first, else NextPlayerSlot++; fills both direction maps and clears the TCP transport's freed guard (ReusePlayerSlot) so the slot is a legitimate auth target again.

释放玩家槽位:清除双向映射、槽位入 FreeSlots 复用池,并调用 transport 的 ReleasePlayerSlot 释放该槽位全部 TCP 资源(socket、认证注册、挂起快照)——防止死 socket 跨玩家登出残留污染下一次重连。

Frees a player slot: clears both maps, returns the slot to the FreeSlots pool, and calls the transport's ReleasePlayerSlot to release all TCP resources (socket, auth registration, pending snapshot) — so a dead socket cannot survive a player's logout and poison the slot's next reconnect.

客户端主动报告的失步(ServerReportDesync 落地):查服务器该 tick 的 hash,不一致则 RecordDesync 并把客户端移入 ClientsRejectedUntilResync(拒绝其输入直到 resync 完成)。

Client-reported desync (landing of ServerReportDesync): looks up the server hash for that tick; on mismatch, records the desync and quarantines the client in ClientsRejectedUntilResync (input rejected until resync completes).

处理客户端 ack(ServerAckTick 落地):先做 ack 钳制——超过服务器当前 tick 的 ack(伪造值,合法客户端只能 ack 已广播过的 tick)钳到当前 tick,每玩家首次伪造打一条 [ACK_CLAMP] Warning(作弊信号),其后静默钳制(不给日志洪水向量);推进 ClientAckedTicks、更新最大滞后;若服务器在该 tick 有 hash 且与客户端不一致——计入 ClientConsecutiveHashMismatches 连击,连续 DesyncConfirmTicks 次(阈值见 UMassBattleNetworkSubsystem)才记录失步并隔离客户端(瞬态自愈型 1-2 tick 毛刺到不了阈值);任一匹配 ack 清零连击(与客户端 ConsecutiveDesyncMismatchTicks 镜像)。监听服务器本地玩家每 tick 自 ack(共享服务端模拟故恒匹配)。

Handles a client ack (landing of ServerAckTick): first the ack clamp — an ack beyond the server's current tick (a fabricated value; a legit client can only ack ticks the server has broadcast) is clamped to the current tick, with one [ACK_CLAMP] Warning per player on first fabrication (cheat signal) and silent clamping afterwards (no log-flood vector); advances ClientAckedTicks, updates the max observed lag; when the server has a hash for that tick and it mismatches the client — the streak counter ClientConsecutiveHashMismatches increments, and the desync is recorded + the client quarantined only after DesyncConfirmTicks consecutive mismatches (threshold on UMassBattleNetworkSubsystem; a transient self-healing 1-2 tick blip never reaches it); any matching ack resets the streak (mirrors the client-side ConsecutiveDesyncMismatchTicks). The listen-server local player self-acks every tick (shares the server sim, so it always matches).

服务端心跳监控(内部 bPacedMode 门控):最后 ack 早于 PacingTimeoutTicks 的客户端判死,踢出前广播 ClientKickedPacingTimeout 网络事件。单阈值——ack 年龄即停滞。网络恢复中的客户端经 MulticastDesyncCheck 持续 ack,天然豁免。末尾委托 CheckSpawnTimeouts。由 OnSimTickCompleted 每 tick 调用。

Server-side heartbeat monitor (gated by bPacedMode internally): clients whose last ack is older than PacingTimeoutTicks are judged dead; a ClientKickedPacingTimeout network event is broadcast before kicking. Single threshold — the ack age IS the stall. Clients in network recovery keep acking via MulticastDesyncCheck and are naturally exempt. Delegates to CheckSpawnTimeouts at the end. Called every tick from OnSimTickCompleted.

处理客户端 resync 请求(ServerRequestResync 落地):已在派发中(ClientsWithPendingResync)则忽略重复;非 post-command 边界或生成进行中则延迟入 PendingResyncClients。到达边界后:取最新缓存快照,从命令历史填充半开区间 [PayloadTick, CurrentTick+1] 的 gap 命令(客户端 FF 期间重放,不得依赖在途投递),派发 TCP 快照(DispatchSnapshotViaTcp_Async,后台序列化/压缩),派发前加入 in-flight 守卫。

Handles a client resync request (landing of ServerRequestResync): duplicates are ignored while a dispatch is in flight (ClientsWithPendingResync); requests are deferred into PendingResyncClients outside the post-command boundary or while spawning. At the boundary: takes the latest cached snapshot, fills gap commands over the half-open range [PayloadTick, CurrentTick+1] from command history (the client replays them during FF and must not depend on in-flight delivery), dispatches via TCP (DispatchSnapshotViaTcp_Async, background serialize/compress), and adds the in-flight guard before dispatch.

在确定性 post-command 边界逐个处理延迟的 resync 请求:取队列头玩家,调 HandleClientResyncRequest。生成进行中(IsSpawnInProgress)则整批推迟。

Processes deferred resync requests one by one at the deterministic post-command boundary: takes the head player and calls HandleClientResyncRequest. Deferred wholesale while a spawn is in progress (IsSpawnInProgress).

在确定性 post-command 边界处理 late-join 客户端:队列清空后逐个解析控制器与槽位;无缓存快照则标记强制捕获并重新入队;有快照则填充 [PayloadTick, 当前 tick) 的 gap 命令,加入 LateJoinSnapshotsInFlight 后异步派发——派发失败移除 in-flight 并重新入队(下一边界重试),防止失败快照从重试中永久消失。

Handles late-join clients at the deterministic post-command boundary: after clearing the queue, resolves each controller and slot; without a cached snapshot it forces a capture and re-queues; with one, fills gap commands over [PayloadTick, current tick), adds the client to LateJoinSnapshotsInFlight, and dispatches asynchronously — on dispatch failure the in-flight marker is removed and the client re-queued (retried next boundary), so a failed late-join snapshot never vanishes from retry forever.

客户端 resync 完成的清理(ServerResyncComplete / ServerReportResyncResult 落地):ack 置当前 tick、解除隔离、移除 in-flight 守卫,并冲刷该玩家的延迟 Agent.Spawn(快照优先保证)。

Post-resync cleanup (landing of ServerResyncComplete / ServerReportResyncResult): ack to the current tick, un-quarantine, remove the in-flight guard, and flush the player's deferred Agent.Spawn (snapshot-first guarantee).

resync 结果统一收敛点(TCP 异步派发失败、ServerReportTcpSnapshotResult、ServerReportResyncResult 都汇聚于此):失败且属于 late-join in-flight 时重新入队(服务端主动重试,不依赖客户端 300-tick pacing 自救);随后经 H2 派发匹配闸——客户端上报的结果必须匹配一条在途派发记录(双方都有 SnapshotId 时还需匹配;服务器内部失败作为可信方直接通过),无效上报保持隔离但释放 ClientsWithPendingResync 传输守卫(否则错配上报会让该客户端下一次真实 resync 请求被去重死锁);通过后 OnClientResyncComplete + OnClientResynced 钩子。

Unified resync-outcome funnel (async TCP dispatch failure, ServerReportTcpSnapshotResult, and ServerReportResyncResult all converge here): on failure of a late-join in-flight dispatch the client is re-queued (the server retries proactively instead of relying on the client's 300-tick pacing self-recovery); then the H2 dispatch-match gate — a client-reported result must match an outstanding dispatch record (SnapshotId too when both sides have one; server-internal failures pass as trusted), and an invalid report holds the quarantine but releases the ClientsWithPendingResync transport guard (otherwise a mismatched report would deadlock the client's next genuine resync request in the dedup); on pass, OnClientResyncComplete + the OnClientResynced hook run.

H2 派发匹配验证三件套(由 PlayerController RPC 与静态异步派发助手调用):TryConsumePendingDispatch 是验证核心——服务器内部完成(bServerInitiated)作为可信方直接消费 PendingSnapshotDispatches 记录;客户端上报必须有在途记录,且双方 SnapshotId 都有效时须一致;RecordPendingDispatchSnapshotId 在异步派发拿到 SnapshotId 后回填记录;ValidateAndCompleteResync 是 ServerResyncComplete 的带验证入口——TCP 派发的快照会产生两份完成上报(传输层 ack 与追帧完成 RPC),后到者查无记录:未隔离玩家按幂等重复上报放行(仅清理语义),隔离中的玩家保持审计拒绝(但释放传输守卫)。

The H2 dispatch-match validation trio (called by PlayerController RPCs and the static async dispatch helper): TryConsumePendingDispatch is the validation core — server-internal completions (bServerInitiated) consume the PendingSnapshotDispatches record as trusted; client reports require an outstanding record, with SnapshotIds matching when both are valid; RecordPendingDispatchSnapshotId backfills the record once the async dispatch obtains a SnapshotId; ValidateAndCompleteResync is the validated entry of ServerResyncComplete — a TCP-dispatched snapshot produces TWO completion reports (transport ack + FF-completion RPC), and whichever lands second finds no record: an unquarantined reporter passes as an idempotent duplicate (cleanup semantics only), a quarantined one is still audit-rejected (but the transport guard is released).

专用服务器路径:接收客户端模拟 hash(ServerReportHash 落地)并转发给 NetworkRelay->ReportClientHash 参与投票。

DedicatedServer path: receives a client simulation hash (landing of ServerReportHash) and forwards it to NetworkRelay->ReportClientHash for voting.

专用服务器路径:接收压缩快照上传(ServerUploadSnapshot 落地)用于灾备。入口验证:大小须在 (0, 64MB] 且 tick 在 [LastKnownServerTick-1024, LastKnownServerTick+256] 窗口内——未来 tick 视为恶意标记直接拒绝(防 relay 的 SnapshotHistory 无界增长,ListenServer 模式 relay 自身 CurrentTick 从不推进);另加每玩家上传冷却(5 秒,审计 H6 带宽向量——relay 字节预算限制内存,冷却限制 CPU/带宽:反复的 64MB 上传消耗真实的序列化/压缩时间);通过后转发 NetworkRelay->UploadSnapshot 存储。

DedicatedServer path: receives a compressed snapshot upload (landing of ServerUploadSnapshot) for disaster recovery. Entry validation: size in (0, 64MB] and tick inside [LastKnownServerTick-1024, LastKnownServerTick+256] — a future tick is treated as a hostile marker and rejected (prevents unbounded SnapshotHistory growth, since the relay's own CurrentTick never advances in ListenServer mode); plus a per-player upload cooldown (5s, audit H6 bandwidth vector — the relay's byte budget bounds memory while the cooldown bounds CPU/bandwidth: repeated 64MB uploads cost real serialization/compression time); valid uploads forward to NetworkRelay->UploadSnapshot.

计算动态 K(毫秒)。当前实现返回静态 MinInputBufferMs——LAN 环境无需动态抖动/RTT 测量;广域网请在配置中增大 MinInputBufferMs。每次 tick 由 OnSimTickCompleted 读取并随 MulticastSyncTick 下发。

Computes the dynamic K (ms). The current implementation returns the static MinInputBufferMs — LAN play needs no dynamic jitter/RTT measurement; for WAN, raise MinInputBufferMs in config. Read each tick by OnSimTickCompleted and broadcast with MulticastSyncTick.

服务端每 tick 网络中枢(私有,由 UMassBattleNetworkSubsystem::OnTickCompleted 触发,仅权威端执行)。流程:① 网络 flush tick(按 NetworkFlushIntervalTicks 节拍)为每个带 Component→Entity 轴的组件采样变换、K-buffer 预测外推、网格编码生成 AgentComponent.SetAgentTransform 指令;② 排序并多播 PendingCommands((ExecuteTick, ServerExecSequence) 序);③ 每 tick 计算服务端 tick 间隔 EWMA 与时间膨胀并 MulticastSyncTick;④ flush tick 且开启失步检测时 MulticastDesyncCheck;⑤ post-command 边界处理 late-join 与 resync;⑥ 冲刷组件驱动锁步生成;⑦ 监听服务器本地自 ack、步调超时检查、强制 flush 连接降低调速延迟。

Per-tick server network hub (private, fired by UMassBattleNetworkSubsystem::OnTickCompleted, authority only). Flow: ① on network flush ticks (paced by NetworkFlushIntervalTicks), samples every component with a Component→Entity axis, extrapolates by K-buffer prediction, grid-encodes, and produces AgentComponent.SetAgentTransform commands; ② sorts and multicasts PendingCommands (by (ExecuteTick, ServerExecSequence)); ③ every tick computes the server tick-interval EWMA and time dilation and calls MulticastSyncTick; ④ on flush ticks with desync check enabled, MulticastDesyncCheck; ⑤ handles late-join and resync at the post-command boundary; ⑥ flushes component-driven lockstep spawns; ⑦ listen-server local self-ack, pacing timeout checks, and forced connection flush to lower pace-sync latency.