MassBattleDesyncAutoDebugger.h

发散回调,由双端各自的失步检测点调用(客户端的延迟校验 / 服务端的回执哈希失配,见管线角色)。流程:① LoadConfigOnce 首次调用时从 UMassBattleSubsystem 读取 Config 属性并打日志;② 按世界名去重(函数内 static TSet<FName>)——PIE 双 World 共进程时,进程级静态闸门会挡掉另一端的转储,按世界名键控既保留「每世界每会话只转储一次」的防刷闸,又兼容 PIE(PIE 世界名按实例唯一);③ 置 Net->bSimulationPaused = true 冻结现场;④ 调 DumpWorldState 全量转储;⑤ 打 Error 日志(含转储路径)。任一指针无效或该世界已转储过则直接返回,不暂停。

Divergence callback, invoked from both peers' divergence-detection sites (the client-side deferred check / the server-side acked-hash mismatch — see Pipeline Role). Flow: ① LoadConfigOnce reads the Config properties from UMassBattleSubsystem on first call and logs them; ② deduplicates per world name (a function-local static TSet<FName>) — with two PIE worlds in one editor process a process-static gate would block the other peer's dump; keying by world name keeps the "one dump per session per world" spam guard while remaining PIE-safe (PIE world names are unique per instance); ③ sets Net->bSimulationPaused = true to freeze the scene; ④ dumps full state via DumpWorldState; ⑤ logs an Error with the dump path. Returns immediately (no pause) if either pointer is invalid or this world was already dumped.

完整世界转储。文件头 MBDMP v1 tick=… role=…;实体来自 MB->DiagUIDEnumerateForDump() 并按 UID 升序排序;每个实体先输出 FEntityFlagFragment 旗标(低/高 32 位十六进制),再对 14 个热 fragment 做反射全量转储:FMoving、FNavigating、FLocating、FRotating、FTracing、FBehaving、FAttacking、FDeterminism、FMove、FChase、FCollider、FAttack、FAvoiding。属性序列化规则:对象/弱引用/软引用/委托/接口/Class 属性一律剥离;浮点以位模式十六进制(%016llx)输出以暴露最低位差异;整数/枚举输出十进制;名称/字符串原样;数组先记 LEN(真实长度)再最多转储 32 个元素;嵌套结构递归(深度上限 6);实体句柄类(EntityHandle / MassEntityHandle / FlowFieldLayerHandle)跳过。最后 TASKS 段反射转储每个活跃的 MoveTo 任务对象(MB->GetActiveMoveToTasks())的全部属性。整串以 UTF-8(无 BOM)写盘。

Full world dump. Header MBDMP v1 tick=… role=…; entities come from MB->DiagUIDEnumerateForDump() sorted ascending by UID; each entity first emits its FEntityFlagFragment flags (low/high 32-bit hex), then a full reflection dump of 14 hot fragments: FMoving, FNavigating, FLocating, FRotating, FTracing, FBehaving, FAttacking, FDeterminism, FMove, FChase, FCollider, FAttack, FAvoiding. Property serialization rules: object/weak/soft/delegate/interface/class properties are stripped; floats print as bit-pattern hex (%016llx) to expose last-bit differences; ints/enums print decimal; names/strings verbatim; arrays record LEN (true length) then dump up to 32 elements; nested structs recurse (depth cap 6); handle-typed structs (EntityHandle / MassEntityHandle / FlowFieldLayerHandle) are skipped. A final TASKS section reflects every active MoveTo task object (MB->GetActiveMoveToTasks()). The whole string is written UTF-8 without BOM.

模糊器的确定性 LCG(线性同余发生器):FuzzerRngState = FuzzerRngState * 1664525u + 1013904223u,返回状态右移 8 位。同种子必得同序列——这是「任何发现的发散可逐位复现」的基础。

The fuzzer's deterministic LCG (linear congruential generator): FuzzerRngState = FuzzerRngState * 1664525u + 1013904223u, returning the state shifted right by 8. Same seed ⇒ same sequence — the foundation of "any found divergence is reproducible bit-for-bit".

模糊器逐 tick 推进,仅服务器端调用(命令经正常锁步通道复制到客户端)。首次调用为初始化:置 bFuzzerInitialized、以 FuzzerSeed | 1 播种 LCG、下次触发定在当前 tick + 60、打布防日志。到达触发 tick 时:① 在 [Min, Max] 区间抽下次触发间隔;② 经 DiagUIDEnumerateForDump 枚举实体(不足 4 个则本 tick 放弃);③ 在 [MinBatch, MaxBatch] 抽批量、以 FuzzerRand 无放回抽取有效实体;④ 抽动作(0=MoveTo / 1=Chase / 2=Formation)与是否可打断;⑤ 构造 FMassBattleNetCommand(CommandTag = "MB.FuzzTask",IssuedTick = T,ExecuteTick = T + 2,IntPayload = [action, canInterrupt, targetUid, n, uid1..uidN],FloatPayload = [X, Y]——MoveTo/Formation 为 [-10000, 10000) 随机坐标、Chase 为 0),经 AMassBattleNetworkGameModeBase::EnqueueAuthorityCommandForFuzz 入队。走锁步命令通道的原因:任务工厂仅可在 ExecuteNetCommand 内合法激活。

Per-tick fuzzer advance, called on the SERVER only (commands replicate to clients through the normal lockstep path). First call initializes: sets bFuzzerInitialized, seeds the LCG with FuzzerSeed | 1, schedules the first fire at current tick + 60, and logs the arming. On a fire tick: ① draws the next interval within [Min, Max]; ② enumerates entities via DiagUIDEnumerateForDump (fewer than 4 ⇒ skip this tick); ③ draws a batch size within [MinBatch, MaxBatch] and picks valid entities without replacement via FuzzerRand; ④ draws the action (0=MoveTo / 1=Chase / 2=Formation) and the interrupt flag; ⑤ builds an FMassBattleNetCommand (CommandTag = "MB.FuzzTask", IssuedTick = T, ExecuteTick = T + 2, IntPayload = [action, canInterrupt, targetUid, n, uid1..uidN], FloatPayload = [X, Y] — random coordinates in [-10000, 10000) for MoveTo/Formation, 0 for Chase) and enqueues it via AMassBattleNetworkGameModeBase::EnqueueAuthorityCommandForFuzz. The lockstep-command routing exists because task factories may only activate legally inside ExecuteNetCommand.

在锁步门内解码并激活一条 "MB.FuzzTask" 命令——由 ExecuteNetCommand 在双端调用,保证双端同构执行(任务工厂因此可合法激活)。先做载荷长度校验(IntPayload.Num() >= 4 + N,不足直接返回);经 UMassBattleSubsystem::GetPtr 取子系统,按 UID 反查并过滤有效实体(空批次返回)。分发:action 1 → UMassBattleBPTaskAgentsChaseAttack::AgentsChaseAttack(目标为 TargetUid 实体);action 0 → 单点 AgentsMoveTo(目标 (X, Y, 0));action 2 → 编队网格 AgentsMoveTo(列数 = √N,间距 150cm,自 (X, Y) 起排布)。

Decodes and activates one "MB.FuzzTask" command inside the lockstep gate — called from ExecuteNetCommand on both peers so both execute identically (task factories thereby activate legally). Validates payload length first (IntPayload.Num() >= 4 + N, early return otherwise); resolves the subsystem via UMassBattleSubsystem::GetPtr, maps UIDs back to valid entities (empty pick returns). Dispatch: action 1 → UMassBattleBPTaskAgentsChaseAttack::AgentsChaseAttack (target = the TargetUid entity); action 0 → single-point AgentsMoveTo (goal (X, Y, 0)); action 2 → formation-grid AgentsMoveTo (columns = √N, 150 cm spacing from (X, Y)).