UMassBattleHashGridSubsystem

USubsystem 生命周期入口,负责把配置值落成缓存:三套网格的逆单元尺寸(*InvCellSizeCache)、块维度(*BlockDimensionsCache)(每个维度 Max(1, 配置),单元尺寸下限 0.001cm)与块大小乘积(checkf 校验 int32 溢出);簇维度强制 Z=1(checkf 断言)。随后 RandomStream.Initialize(12345)、InitPressureField()、ClearAllCells()。

USubsystem lifecycle entry; bakes config values into caches: inverse cell sizes (*InvCellSizeCache) for the three grids, block dimensions (*BlockDimensionsCache) (each axis Max(1, config), cell size floor 0.001 cm), and the block size product (checkf for int32 overflow); cluster dimensions are forced to Z=1 (checkf). Then RandomStream.Initialize(12345), InitPressureField(), and ClearAllCells().

USubsystem 生命周期出口:先 ClearAllCells() 释放三套网格与簇索引,再调用父类实现。

USubsystem lifecycle exit: calls ClearAllCells() to release the three grids and cluster indices, then the parent implementation.

每帧 tick 入口,当前仅做一件事:FlushCommandBuffer() 排空本帧并行注册压入的命令(排序 + 串行执行 + 建块)。由 UMassBattleSubsystem 的模拟 tick 调用。

Per-frame tick entry; currently does exactly one thing: FlushCommandBuffer() drains this frame's parallel registrations (sort + serial execution + block creation). Called from UMassBattleSubsystem's simulation tick.

静态访问器:经 GEngine->GetWorldFromContextObject(LogAndReturnNull 模式)解析世界再取子系统;解析失败返回 nullptr。

Static accessor: resolves the world via GEngine->GetWorldFromContextObject (LogAndReturnNull mode) and returns its subsystem; nullptr on resolution failure.

静态访问器的引用版本:失败即 checkf 断言("Failed to get UMassBattleHashGridSubsystem from world context.")——用于子系统必须存在的路径。

Reference version of the static accessor: fails via checkf ("Failed to get UMassBattleHashGridSubsystem from world context.") — for paths where the subsystem must exist.

惰性访问并缓存 UMassAPISubsystem:命中缓存直接返回;否则从当前世界取子系统并写入 CurrentWorld / MassAPISubsystem 缓存。

Lazily accesses and caches UMassAPISubsystem: returns the cache on hit; otherwise fetches it from the current world and fills the CurrentWorld / MassAPISubsystem caches.

惰性访问并缓存 UMassBattleSubsystem,行为与 GetMassAPISubsystem 对称。

Lazily accesses and caches UMassBattleSubsystem; symmetric to GetMassAPISubsystem.

命令缓冲的确定性排空(由 ExecuteTick 与 MassBattleGridUpdateProcessor 调用)。流程:① CommandBuffer.Merge() 合并各线程槽;② 空则返回;③ bDeterministic 开启时按 (Coord X→Y→Z → CommandType → UniqueID;Segment 用 GroupID → Index) 排序;④ 收集缺失块,按固定顺序单线程建块(MakeShared),Agent/Obstacle 同时把簇坐标幂等加入 *OccupiedClusters;⑤ 串行(禁用 MBParallelFor)逐条 ExecuteRegistrationCommand——并行调度会抵消排序,导致 Cell.Agents 追加顺序逐运行变化;⑥ Reset() 清空收集器。

Deterministic drain of the command buffer (called by ExecuteTick and MassBattleGridUpdateProcessor). Flow: ① CommandBuffer.Merge() merges per-thread slots; ② returns when empty; ③ when bDeterministic is on, sorts by (Coord X→Y→Z → CommandType → UniqueID; Segments by GroupID → Index); ④ collects missing blocks and creates them single-threaded in fixed order (MakeShared), idempotently adding cluster coords to *OccupiedClusters for Agent/Obstacle; ⑤ executes commands serially (no MBParallelFor) via ExecuteRegistrationCommand — parallel dispatch would negate the sort and make Cell.Agents append order vary per run; ⑥ Reset() clears the gatherer.

单条命令的串行执行(仅 FlushCommandBuffer 调用)。按 CommandType 分派到三套网格:设置占用位、维护块级 composition(锁内 HasAll 检查消除 TOCTOU;首 agent 赋值交集,非首走写前检查 + AND)、写 CellLocation、锁内维护单元级 composition 并追加数据。RegisterAgent 分支额外用 CAS 原子最大值维护 MaxAgentNeighborExtent(max(AvoidRadiusX, ColliderHeight*0.5 + ColliderRadius*ColliderMult),ColliderMult 只放大半径)。

Serial execution of a single command (called only by FlushCommandBuffer). Dispatches on CommandType into the three grids: sets the occupancy bit, maintains block-level composition (HasAll checked inside the lock to close the TOCTOU race; first agent assigns the intersection, later ones write-before-check + AND), writes CellLocation, and appends the payload under the cell lock with cell-level composition maintenance. The RegisterAgent branch additionally maintains MaxAgentNeighborExtent via CAS atomic-max (max(AvoidRadiusX, ColliderHeight*0.5 + ColliderRadius*ColliderMult); ColliderMult inflates the radius only).

agent 注册的公开入口:转发到 RegisterCell_Internal(AgentGrid, RegisterAgent, ...)。并行段(如 MassBattleGridUpdateProcessor 的 MBForEachEntityChunk)直接调用,块存在时走加锁快路径,块缺失时压命令缓冲。

Public entry for agent registration: forwards to RegisterCell_Internal(AgentGrid, RegisterAgent, ...). Called directly from parallel sections (e.g. MassBattleGridUpdateProcessor's MBForEachEntityChunk); takes the locked fast path when the block exists, else queues a command.

障碍注册的公开入口,Capsule / Segment / Box 各一个重载,分别转发到对应 RegisterCell_Internal(ObstacleGrid, ...)。行为与 RegisterAgentCell 相同(快路径 / 命令回退),静态障碍(GridData.bStatic)走 StaticOccupiedCells 位与 Static* 数组。

Public entries for obstacle registration, one overload per Capsule / Segment / Box, each forwarding to the matching RegisterCell_Internal(ObstacleGrid, ...). Behavior matches RegisterAgentCell (fast path / command fallback); static obstacles (GridData.bStatic) use the StaticOccupiedCells bits and Static* arrays.

战利品注册的公开入口:转发到 RegisterCell_Internal(LootGrid, RegisterLoot, ...)。

Public entry for loot registration: forwards to RegisterCell_Internal(LootGrid, RegisterLoot, ...).

agent 注册的核心实现。块存在时:占用位 Lock→置位→Unlock(已置位跳过);块级 composition 锁内 HasAll 检查 + 追加,首 agent(Composition.IsEmpty() 判断)直接赋值交集;单元级同样锁内维护并写 CellLocation + Agents.Add。块缺失时:构造 FGridCommandBufferCommand(携带 AgentComposition)压入 CommandBuffer.GetThreadBuffer()。

Core implementation of agent registration. When the block exists: occupancy bit is Lock→set→Unlock (skipped when already set); block-level composition is HasAll-checked inside the lock before append, first agent (detected via Composition.IsEmpty()) assigns the intersection directly; cell-level composition is maintained under the cell lock together with CellLocation write and Agents.Add. When the block is missing: builds an FGridCommandBufferCommand (carrying AgentComposition) into CommandBuffer.GetThreadBuffer().

障碍与战利品注册的核心实现,结构同 agent 版:块存在时置位(障碍按 bStatic 选动态/静态位数组)并在单元锁内追加到对应数组;块缺失时构造对应载荷的命令压入缓冲。

Core implementations for obstacle and loot registration, structurally identical to the agent version: when the block exists, set the bit (obstacles pick the dynamic/static bit array by bStatic) and append to the matching array under the cell lock; when the block is missing, queue a command carrying the matching payload.

三套网格的占用单元全量复位(并行,每块一个任务):遍历占用位逐单元 Reset()、占用位数组 Init(false)、块级 composition 复位。最后把 MaxAgentNeighborExtent 清零,使 gather 的剪枝截止值只由本帧注册重建。

Full occupancy reset for all three grids (parallel, one task per block): iterates the occupancy bits, Reset()s each cell, Init(false)s the bit array, resets block compositions. Finally zeroes MaxAgentNeighborExtent so the gather's prune cutoff is rebuilt from this frame's registrations only.

仅 agent 网格的占用复位(每帧由 MassBattleGridUpdateProcessor 在重新注册所有 agent 前调用;resync 路径由 MassBattleNetworkSubsystem 调用)。同时清零 MaxAgentNeighborExtent——注释标明这是清理 gather 剪枝运行最大值"正确的钩子"。

Occupancy reset for the agent grid only (called every frame by MassBattleGridUpdateProcessor before re-registering all agents; the resync path in MassBattleNetworkSubsystem also calls it). Also zeroes MaxAgentNeighborExtent — the comment marks this as the right hook for clearing the gather's running-max prune.

障碍网格的动态占用复位。注意:只清动态侧——StaticOccupiedCells 位数组与单元静态数组原样保留,静态障碍跨帧存活。

Dynamic occupancy reset for the obstacle grid. Note: only the dynamic side is cleared — the StaticOccupiedCells bits and the static arrays survive untouched, so static obstacles persist across frames.

清空障碍网格静态半区(resync 拆除用;并行,每块一个任务):对静态占用位置位的单元调用 StaticReset(),再把 StaticOccupiedCells 整体 Init(false)。逐帧重置从不清静态半区——resync 经 DestroyQuery 直接销毁实体(不走 Obstacle.Destroy),不清则孤儿静态条目与重建实体(bRegistered=false,新生语义)同格重复注册同 UID → [RVO_DESYNC] 排序去重报错;静态内容全部可由实体重建,清空即自愈。双半区皆空的块由随后的 ShrinkObstacleGrid 收割。由 MassBattleNetworkSubsystem 的 resync 恢复路径调用(:2114)。

Wipes the obstacle grid's STATIC half (resync teardown; parallel, one task per block): calls StaticReset() on cells with a static occupancy bit set, then re-Init(false)s the whole StaticOccupiedCells array. The per-frame reset never touches the static half — resync destroys networked entities directly (DestroyQuery, no Obstacle.Destroy), so without this wipe the orphaned static entries would meet the rebuilt entities' fresh registration (bRegistered=false by design) as duplicate UIDs in the same cell → [RVO_DESYNC] duplicate-sort errors; everything static is rebuildable from entities, so wiping is self-healing. Blocks left empty on both halves are reaped by the subsequent ShrinkObstacleGrid. Called from the resync restore path of MassBattleNetworkSubsystem (:2114).

战利品网格的占用复位,行为与 agent 版一致(无 composition 部分)。

Occupancy reset for the loot grid, same behavior as the agent version (minus the composition part).

按 UID 从全部块的静态胶囊/静态盒数组中移除障碍(注意:不处理静态样条段,那是 RemoveStaticSegmentObstacleByGroupID 的职责)。单元三个静态数组全空时把对应 StaticOccupiedCells 位清 false。由 MassBattleGridUpdateProcessor 在静态盒停用时调用。

Removes an obstacle by UID from every block's static capsule/box arrays (note: static spline segments are not touched — that is RemoveStaticSegmentObstacleByGroupID's job). When a cell's three static arrays are all empty, its StaticOccupiedCells bit is cleared. Called by MassBattleGridUpdateProcessor when a static box is deactivated.

按 GroupID 从全部块的静态样条段数组中移除段障碍;单元静态数组全空时清占用位。

Removes segment obstacles by GroupID from every block's static segment array; clears the occupancy bit when the cell's static arrays are all empty.

完全清空三套网格(AgentGrid / ObstacleGrid / LootGrid 的 Empty())与两个簇索引集合(AgentOccupiedClusters / ObstacleOccupiedClusters)。Initialize 与 Deinitialize 各调用一次。

Fully empties the three grids (AgentGrid / ObstacleGrid / LootGrid Empty()) and the two cluster sets (AgentOccupiedClusters / ObstacleOccupiedClusters). Called once each from Initialize and Deinitialize.

[临时诊断·resync 循环猎杀,排查完成后移除] 转储障碍网格盒障碍条目清单外加折叠清单哈希:遍历 ObstacleGrid 全部块,对动/静态占用单元各收集 {块坐标, 格索引, 动静, UID, 半尺寸} 条目,按 (UID, CellIdx, bStatic) 排序后逐字段 CityHash64WithSeed 折叠为顺序无关的 invHash——网格相同 ⇒ 哈希相同,双端对比即可检验恢复/FF 后网格内容是否分歧(网格状态不在锁步 hash 内);差异可经逐条目日志定位。摘要行 [OBST_GRID_INV](total/dyn/static/invHash)恒打;仅 bObstacleGridInvFullDump 开启时才逐条打(一次全量 3k+ 行 Error)。仅 game thread。双端调用方:MassBattleNetworkGameModeBase(:919,tag SRV)与 MassBattleNetworkGameStateBase(:360,tag CLI)。

[TEMP-DIAG resync-loop hunt, remove when the hunt is done] Dumps the obstacle grid's box-obstacle entry inventory plus a folded inventory hash: walks every block of ObstacleGrid, collects {block coord, cell index, dynamic/static, UID, half extents} entries from the occupied cells on each half, sorts them by (UID, CellIdx, bStatic) and folds field-by-field with CityHash64WithSeed into an order-independent invHash — identical grids ⇒ identical hash, so diffing across peers tests whether the grid CONTENT diverges after restore/FF (grid state is not part of the lockstep hash); a diff localizes the lost/extra entry via the per-entry lines. The summary line [OBST_GRID_INV] (total/dyn/static/invHash) always prints; full per-entry enumeration only when bObstacleGridInvFullDump is on (one full dump is 3k+ Error lines). Game thread only. Called on both peers by MassBattleNetworkGameModeBase (:919, tag SRV) and MassBattleNetworkGameStateBase (:360, tag CLI).

枚举全部障碍条目(动/静 × 胶囊/线段/盒三种几何)每(格注册)一次:遍历 ObstacleGrid 全部块的动/静态占用单元,经静态助手 VisitObstCell 把六数组逐条回调 Visitor(Type 编码 0=胶囊、1=线段、2=盒)。O(K + 块数),串行、只读——供只需要小障碍群体的每帧建池方,绝不该走逐 agent 引用遍历。多格注册的障碍每格出现一次,上游按 Data 指针去重。(1.19.20 源码树暂无调用方——头文件注明为每帧建池方预留。)

Enumerates EVERY obstacle entry (dynamic/static × capsule/segment/box) exactly once per (cell-registration): walks all blocks' dynamic and static occupied cells and, via the static helper VisitObstCell, hands each of the six arrays' entries to Visitor (Type encoded 0=capsule, 1=segment, 2=box). O(K + blocks), serial, read-only — for per-frame pool builders that only need the small obstacle population, never the per-agent ref walk. Multi-cell-registered obstacles appear once per registered cell; dedup upstream by Data pointer. (No callers in the 1.19.20 tree yet — the header reserves it for per-frame pool builders.)

回收空块:移除所有无占用单元的块,并同步收掉簇索引(Agent/Obstacle 版先收集候选簇,再检查簇覆盖范围内是否还有块,无则从 *OccupiedClusters 移除;Loot 无簇层)。三个公开包装器 ShrinkAgentGrid / ShrinkObstacleGrid / ShrinkLootGrid 各调用对应的 ShrinkGrid_Internal 重载,ShrinkGrids 依序调用全部三个。

Reaps empty blocks: removes all blocks without occupied cells and prunes cluster indices accordingly (the Agent/Obstacle versions collect candidate clusters first, then check whether any block remains inside the cluster's footprint and remove the cluster coord from *OccupiedClusters if not; Loot has no cluster layer). The three public wrappers ShrinkAgentGrid / ShrinkObstacleGrid / ShrinkLootGrid each call the matching ShrinkGrid_Internal overload; ShrinkGrids calls all three in order.

把每个 agent 单元内的 Agents 按 UID 升序排序(并行,每单元一个任务)。先收集全部块坐标并按 X→Y→Z 排序保证确定性遍历,再收集占用单元。头文件注明:确定性排序对 None 与 Randomize 模式的确定性查询仍是必需的(它们按收集序输出/洗牌),NearToFar/FarToNear 因候选排序自带 UID tiebreak 而自足。由 MassBattleGridUpdateProcessor 在 bDeterministic 时每帧调用,作为命令排序之外的快速路径安全网。

Sorts Agents within each cell by UID ascending (parallel, one task per cell). Block coords are first collected and sorted by X→Y→Z for deterministic traversal, then occupied cells are gathered. The header notes that sorted order is still required for deterministic None/Randomize-mode queries (they emit/shuffle candidates in collection order), while NearToFar/FarToNear are self-sufficient (candidate sort with UID tie-break). Called every frame by MassBattleGridUpdateProcessor when bDeterministic, as a safety net on top of the command sort.

障碍单元排序:胶囊/盒按 UID,样条段按 GroupID 再 Index;动态与静态六组数组各自排序;动态/静态占用位任一置位即视为占用单元。并行处理。

Obstacle cell sorting: capsules/boxes by UID, segments by GroupID then Index; all six dynamic/static arrays sorted independently; a cell is occupied when either the dynamic or static bit is set. Processed in parallel.

战利品单元排序:每单元内按 UID 升序,并行处理。保证并行注册时(命令排序之外的快速路径)的确定性收集序。

Loot cell sorting: each cell's Loots sorted by UID ascending, in parallel. Guarantees deterministic collection order for parallel registrations (the fast path outside the command sort).

压力场格网缓存重算(Initialize 调用):格宽下限 0.001 后缓存逆格宽 PressureInvCellSizeCache;每块格数 W×H(各轴下限 1,checkf 校验乘积 int32 溢出)推导逐轴块世界尺寸 PressureBlockWorldSizeCache(默认 256×200cm = 每边 51200uu)及其逆缓存。存储惰性——块每帧由 ResetPressureField 从沉积推导创建,此处不分配任何数组。

Recomputes the pressure lattice caches (called from Initialize): floors the cell size at 0.001 and caches its inverse (PressureInvCellSizeCache); derives the per-axis block world size (PressureBlockWorldSizeCache, default 256×200 cm = 51200 uu per side) and its inverse from the per-block cell count W×H (each axis floored at 1, checkf for int32 overflow on the product). Storage is lazy — blocks are created per frame by ResetPressureField from deposits; no arrays are allocated here.

每帧 game-thread 重置:从沉积推导本帧活跃块列表(纯函数——列表是沉积块坐标的去重集合,内容与顺序无关,双端一致)。空列表即清场(PressureBlocks.Empty(),采样读 0)——本帧无沉积方时不再保留上一帧扩散值推着 agent 跑。非空时:① 销毁列表外的全部块(内存跟随战场;刻意不做 Num() 门——活跃集"平移但不减员"时幽灵块会持上一帧扩散值渗入邻块、泄漏各端私有历史,resync 后致命,场不入快照且无其他重置点);② 保留块 Memzero Front(FFix64 是可平凡清零的 Q31.32 聚合,零内存即 FFix64(0));③ 缺失块零初始化创建。块创建/销毁仅此一处 game-thread 入口,并行段只读冻结 TMap;块创建序不影响任何结果(所有消费方构造上序无关)。由 MassBattleGridUpdateProcessor 每帧调用(:195 有沉积传活跃列表,:208 无沉积传空列表)。

Per-frame game-thread reset: derives this frame's active block list from deposits (a pure function — the deduped set of deposit block coords, content- and order-independent, identical on both peers). An empty list wipes the field (PressureBlocks.Empty(), samplers read 0) — with no depositors this frame, last frame's diffused values no longer push agents around. Non-empty: ① destroys every block outside the list (memory follows the battle; deliberately no Num() gate — an active set that shifts without shrinking would leave ghost blocks holding last frame's diffused values, diffusing into neighbours and leaking peer-specific history, fatal after resync since the field is snapshot-exempt with no other reset point); ② memzeroes retained Fronts (FFix64 is a trivially-zeroable Q31.32 aggregate — zeroed memory IS FFix64(0)); ③ creates missing blocks zeroed. This is the only game-thread entry point for block create/destroy; parallel passes read the frozen TMap, and block creation order cannot influence any result (every consumer is order-insensitive by construction). Called every frame by MassBattleGridUpdateProcessor (:195 with the active list when there are deposits, :208 with an empty list otherwise).

把压力沉积并行原子散射进各块 Front:MBParallelFor 逐条把 FFix64::FromFloat(D.Amount) 的原始 int64 经 FPlatformAtomics::InterlockedAdd 加到 (BlockX, BlockY) 块的 Front[LocalIdx]。FFix64 精确整数加可交换——任意应用(交错)顺序结果逐位相同,无需预排序(bDeterministic 在内也无例外),这是 gather→排序→串行 flush 模式的一个可证明确定性例外。块必须已由 ResetPressureField 预建,本遍只读冻结 TMap;LocalIdx 越界或块缺失的条目被跳过。线程本地"上块"缓存(同一个体的脚印沉积在合并缓冲中连续且同块,常见情况一次整数比较替代一次 TMap 哈希),代数计数器为进程级(单进程 PIE 的监听服务器 + 客户端世界共用 worker 池,按实例计数会跨世界别名、把沉积写进另一个世界的块)。由 MassBattleGridUpdateProcessor 在 ResetPressureField 后调用(:198)。

Parallel atomic scatter of pressure deposits into block Fronts: MBParallelFor adds the raw int64 of FFix64::FromFloat(D.Amount) to Front[LocalIdx] of the (BlockX, BlockY) block via FPlatformAtomics::InterlockedAdd. FFix64 exact integer adds commute — the result is bit-identical under any application (interleaving) order, so no pre-sort is needed (bDeterministic included); this is a provably-deterministic exception to the gather→sort→serial-flush pattern. Blocks must already exist (created by ResetPressureField) and the pass only reads the frozen TMap; entries with an out-of-range LocalIdx or a missing block are skipped. A thread-local last-block cache trades a TMap hash for one integer compare in the common case (one agent's footprint deposits are consecutive and share a block), guarded by a process-wide generation counter (single-process PIE runs the listen-server + client worlds over one worker pool; a per-instance counter would alias across worlds and deposit into the other world's blocks). Called by MassBattleGridUpdateProcessor after ResetPressureField (:198).

Jacobi 扩散:对每个活跃块的 Front<->Back 执行 PressureDiffusionIterations 次 5 点 Laplace 模板 FFix64 迭代。先一次性快照活跃块并预解析每块 4 个侧邻(并行任务完全不碰 TMap);每迭代 ParallelFor 以(块,行)展平任务写互斥 Dst 行、只读 Front(含经预解析指针读邻块接缝行/列);迭代后先全部 Swap(Front, Back)、再统一重解析接缝指针——顺序无关性只在这两个独立循环下成立(块序随合并沉积序走,resync 后双端合法地不同)。缺失侧邻读 0(活动区边缘 zero-Dirichlet,与旧全局边界同规则)。PressureDiffusionAlpha 转为 FFix64 后夹到 [0, 0.25](5 点 Laplace 稳定性要求),迭代数为 0 或 Alpha ≤ 0 时直接返回。由 MassBattleGridUpdateProcessor 每帧在沉积后调用(:200)。

Jacobi diffusion: runs PressureDiffusionIterations passes of the 5-point Laplace stencil in FFix64 over each active block's Front<->Back. Active blocks are snapshotted once and each block's 4 side-neighbours pre-resolved (parallel tasks never touch the TMap); each pass runs a ParallelFor over flattened (block, row) tasks that write a disjoint Dst row and read Front only (neighbour-block seam rows/cols through the pre-resolved pointers); after each pass ALL blocks Swap(Front, Back) first, THEN all seam pointers are re-resolved — order-independence holds only with these two separate loops (block order follows merged-deposit order, which legitimately differs between peers after resync). Missing side-neighbours read as zero (zero-Dirichlet at the active-region frontier, the same rule the legacy global boundary applied). PressureDiffusionAlpha converts to FFix64 and clamps to [0, 0.25] (5-point Laplace stability); returns early when iterations are 0 or Alpha ≤ 0. Called every frame by MassBattleGridUpdateProcessor after deposition (:200).

在世界 XY 处双线性采样压力——FFix 核心(SamplePressureFx)的 float 边界转换(直接转发后 ToFloat),供物理漂移/可视化消费方。活跃块外(含 PressureBlocks 为空)读 0。

Bilinear pressure sample at world XY — the float boundary conversion of the FFix core (SamplePressureFx; forwards directly then ToFloat), for physics-drift / viz consumers. Reads 0 outside active blocks (including when PressureBlocks is empty).

压力梯度(∂P/∂X, ∂P/∂Y):直接转发 SamplePressureGradientFx 后逐分量 ToFloat。中心差分在单元空间、缩放为每 cm 单位;活跃块外(含空场)返回零向量。由 MassBattleAgentMoveProcessor(:4243/:5981)与 MassBattleAgentMoveProcessorFP(:5960)在移动热循环中调用(FP 的确定性压力主路径走 FFix 版 SamplePressureGradientFx,:4223)。

Pressure gradient (∂P/∂X, ∂P/∂Y): forwards directly to SamplePressureGradientFx and converts per-component with ToFloat. Central differences in cell space, scaled to per-cm units; returns the zero vector outside active blocks (empty field included). Called from the move hot loop of MassBattleAgentMoveProcessor (:4243/:5981) and MassBattleAgentMoveProcessorFP (:5960); the FP variant's deterministic pressure main path goes through the FFix version SamplePressureGradientFx (:4223).

FFix 主采样器:在世界 XY 处以 FFix64 双线性采样压力。坐标经 FromFloat 一次性转换后全程定点(逆格宽、0.5 偏移到单元中心、Floor、Lerp);单块快路径——一次解析 (X0,Y0) 所在块,4 个采样点都在块内时直接索引,仅罕见的跨缝由 PressureTap 退回 TMap 查找;界外角点读 0。需要位级一致性的消费方(锁步移动)走本函数——上面的 float 对是本核心的边界转换,两者结构上不可能分叉。由 MassBattleAgentMoveProcessorFP 在移动热循环中调用(:4223)。

FFix primary sampler: bilinear FFix64 pressure sample at world XY. Coordinates convert via FromFloat once and stay fixed-point throughout (inverse cell size, 0.5 shift to cell centers, Floor, Lerp); single-block fast path — the (X0,Y0) block is resolved once, the 4 taps direct-index while they stay inside it, and only the rare seam straddle falls back to a TMap Find in PressureTap; out-of-bounds corners read 0. Consumers that need bit-exactness (lockstep movement) go through this function — the float pair above are boundary conversions of this core, so the two can never diverge structurally. Called from the move hot loop of MassBattleAgentMoveProcessorFP (:4223).

FFix 主梯度采样器:(∂P/∂X, ∂P/∂Y) 中心差分——X/Y 各在 ±1 格处执行一次与 SamplePressureFx 完全相同的 FFix 双线性序列,差分除以 2×格宽得到每 cm 单位的 FFix64 梯度(结果与旧单网格核心逐位一致)。一次种子解析中心采样点所在块,±1 格采样绝大多数同块(跨缝时 PressureTap 退回 Find)。由 MassBattleAgentMoveProcessorFP 在移动热循环中调用(:4223)。

FFix primary gradient sampler: (∂P/∂X, ∂P/∂Y) via central differences — each axis runs the exact SamplePressureFx FFix bilinear sequence at ±1 cell, and the difference divided by 2×cell size yields the per-cm FFix64 gradient (results stay bit-identical to the legacy single-grid core). One seed resolves the centre tap's block; the ±1-cell taps hit the same block in the overwhelming majority of samples (PressureTap falls back to a Find on the rare straddle). Called from the move hot loop of MassBattleAgentMoveProcessorFP (:4223).

单元占用查询(内联):转发到对应网格的 IsCellOccupied_Internal——块存在时读占用位(障碍为动态 | 静态),块缺失返回 false。

Cell occupancy queries (inline): forward to the matching grid's IsCellOccupied_Internal — reads the occupancy bit when the block exists (dynamic OR static for obstacles), false when the block is missing.

块感知的占用单元范围迭代模板:保持与朴素逐单元循环一致的全局 Z→Y→X 迭代顺序,但缓存当前块指针与块原点,把 TMap 查找从 2×N 个单元降到约 N 次块切换。障碍版同时检查动态与静态占用位。命中单元调用 Callback(Cell)。

Block-aware occupied-cell range iteration templates: preserve the same global Z→Y→X iteration order as a naive per-cell loop while caching the current block pointer and origin, reducing TMap lookups from 2×N cells to ~N block transitions. The obstacle version checks both dynamic and static occupancy bits. Callback(Cell) is invoked per occupied cell.

Agent / Obstacle / Loot 三套同构的坐标转换器(单位:位置 cm,坐标 = 单元索引空间,向下取整)。统一模式:*LocationToCoord = floor((Location - GridOrigin) * InvCellSize);*CoordToLocation = Coord * CellSize + CellSize*0.5 + GridOrigin(单元中心);*GlobalToBlockCoord = floor(Global / BlockDim);*GlobalToLocalCoord = 正模运算;*LocalCoordToIndex = (Z*DimY + Y)*DimX + X;*IndexToLocalCoord 逆运算。块→簇映射(*BlockToClusterCoord)以世界原点为 cluster (0,0) 中心:floor((Block + Dim/2) / Dim),Z 直通。

Three isomorphic converter sets for Agent / Obstacle / Loot (units: position cm, coords = cell index space, floored). Uniform pattern: *LocationToCoord = floor((Location - GridOrigin) * InvCellSize); *CoordToLocation = Coord * CellSize + CellSize*0.5 + GridOrigin (cell center); *GlobalToBlockCoord = floor(Global / BlockDim); *GlobalToLocalCoord (positive modulo); *LocalCoordToIndex = (Z*DimY + Y)*DimX + X; *IndexToLocalCoord is the inverse. The block→cluster mapping (*BlockToClusterCoord) centers the world origin at cluster (0,0): floor((Block + Dim/2) / Dim), Z passes through.

单元访问器(内联):坐标版转发 GetCellAt_Internal(块缺失返回静态空单元引用 Empty*Cell),位置版先经 *LocationToCoord 转换再查。

Cell accessors (inline): the coord versions forward to GetCellAt_Internal (missing block returns the static empty-cell reference Empty*Cell); the location versions convert via *LocationToCoord first.

球体索敌的单元收集核心:Radius < 0 返回空;取中心 ± 半径的 AABB 盒算出坐标范围;经 ExecuteCellTrace 遍历候选单元,几何测试为"中心到单元 AABB 最近点距离 ≤ 半径"。支持全部三类网格。家族包装器:8 个公开函数(SphereTraceForAgentCells ×2(带/不带 Query)、SphereTraceForObstacleCells、SphereTraceForLootCells 及各自的 Sorted 变体)全部一行转发,默认 SortMode = None、SortFrom = 零向量;Sorted 变体透传排序参数。

Core cell collection for sphere traces: returns empty when Radius < 0; derives the coordinate range from the center ± radius AABB; iterates candidate cells via ExecuteCellTrace; the geometry test is "distance from center to closest point in the cell AABB ≤ radius". Supports all three grid types. Family wrappers: 8 public functions (SphereTraceForAgentCells ×2 (with/without Query), SphereTraceForObstacleCells, SphereTraceForLootCells, plus their Sorted variants) are one-line forwards with default SortMode = None and SortFrom = zero vector; Sorted variants pass the sort parameters through.

球体扫掠的单元收集核心:Start == End(KINDA_SMALL_NUMBER 容差)退化为 SphereTraceForCells_Internal;取两端 ± 半径的 AABB 并集为范围;几何测试为"线段最近点到单元中心 ≤ TotalRadius",其中 TotalRadius = Radius + CellHalfSize.GetMax()(保守放大)。支持全部三类网格;7 个包装器一行转发。

Core cell collection for sphere sweeps: Start == End (within KINDA_SMALL_NUMBER) degenerates to SphereTraceForCells_Internal; the range is the AABB union of both endpoints ± radius; the geometry test is "closest point on the segment to the cell center ≤ TotalRadius" with TotalRadius = Radius + CellHalfSize.GetMax() (conservatively inflated). Supports all three grid types; 7 wrappers forward in one line.

OBB 盒索敌的单元收集核心:先把旋转盒的轴对齐半边长投影到世界轴(|Axis| * Range 之和)得世界 AABB;几何测试:朝向为单位四元数时退化为 AABB-vs-AABB 快速比较,否则逐单元 Intersects_AABB_OBB(15 轴 SAT)。支持全部三类网格;8 个包装器一行转发。

Core cell collection for OBB box traces: first projects the rotated box's axis-aligned half extents onto the world axes (|Axis| * Range summed) to get the world AABB; geometry test: identity orientation degenerates to a fast AABB-vs-AABB compare, otherwise per-cell Intersects_AABB_OBB (15-axis SAT). Supports all three grid types; 8 wrappers forward in one line.

OBB 盒扫掠的单元收集核心:Start == End 退化为 BoxTraceForCells_Internal;单元先按扫掠向量做闵可夫斯基膨胀(中心减半向量、半边长加半向量的绝对值),再与起点处的静态盒做 SAT 测试。仅支持 Agent/Obstacle(三元表达式把非 Agent 一律当 Obstacle,无 Loot 分支);7 个包装器一行转发。

Core cell collection for OBB box sweeps: Start == End degenerates to BoxTraceForCells_Internal; each cell is Minkowski-inflated by the sweep vector (center minus half the vector, half extents plus half the absolute vector) and tested against the static box at the start via SAT. Agent/Obstacle only (the ternary treats any non-Agent grid as Obstacle; no Loot branch); 7 wrappers forward in one line.

胶囊索敌的单元收集核心(1.19.20 新增):Radius 或 HalfHeight 为负返回空。查询轴段 = Center ± Orientation·Up × HalfHeight;粗筛包围盒为轴段 AABB 膨胀 Radius;几何测试为"轴段与膨胀 Radius 后的单元 AABB 相交"(SegmentIntersectsAABB,精确)。支持全部三类网格;8 个包装器(Agent ×2 带/不带 Query、Obstacle、Loot 及各自 Sorted 变体)一行转发,默认 SortMode = None。

Core cell collection for capsule traces (new in 1.19.20): returns empty when Radius or HalfHeight is negative. The query axis segment is Center ± Orientation·Up × HalfHeight; the broad-phase bounds are the axis segment's AABB inflated by Radius; the geometry test is "axis segment intersects the cell AABB inflated by Radius" (SegmentIntersectsAABB, exact). Supports all three grid types; 8 wrappers (Agent ×2 with/without Query, Obstacle, Loot, plus their Sorted variants) forward in one line with default SortMode = None.

胶囊扫掠的单元收集核心(1.19.20 新增):扫掠体积 = 胶囊轴线段与扫掠向量 [0, End-Start] 的闵可夫斯基和再膨胀 Radius(平行四边形 ⊕ 球)。精确单元测试:胶囊轴线段与"单元 AABB 膨胀 Radius、平移 -SweepVector/2、半边长 +|SweepVector|/2"相交(闵可夫斯基结合律把扫掠化为静态 AABB 测试);查询轴段取扫掠起点处(Start ± Up×HalfHeight),粗筛为平行四边形 4 角 AABB 膨胀 Radius。_Internal 按 GridType 分派支持全部三类网格;6 个包装器(Agent ×2 带/不带 Query、Obstacle 及各自 Sorted 变体)一行转发,无 Loot 包装器。

Core cell collection for capsule sweeps (new in 1.19.20): the swept volume is the Minkowski sum of the capsule axis segment and the sweep VECTOR [0, End-Start], inflated by Radius (parallelogram ⊕ sphere). Exact cell test: the capsule axis segment intersects the cell AABB inflated by Radius, translated by -SweepVector/2, with half extents +|SweepVector|/2 (Minkowski associativity turns the sweep into a static AABB test); the query axis segment sits at the sweep start (Start ± Up×HalfHeight), and the broad-phase is the AABB of the parallelogram's 4 corners inflated by Radius. The _Internal dispatches on GridType and supports all three grids; 6 wrappers (Agent ×2 with/without Query, Obstacle, plus their Sorted variants) forward in one line — no Loot wrapper.

扇形索敌的单元收集核心:Radius/Height < 0 或 AngleDegrees ≤ 0 返回空。方向先压平到 XY 平面(GetSafeNormal),两条边界法线绕 UpVector 旋转 ±(90° + 角/2)得到;几何测试为 2D 半径(中心到单元 AABB 最近 XY 点)+ Z 高度(单元 Z 范围与中心 ± 半高重叠)+ 半平面测试(单元 AABB 相对法线的投影半径;AngleDegrees < 360 才做角度测试)。仅支持 Agent/Obstacle;7 个包装器一行转发。

Core cell collection for sector traces: empty when Radius/Height < 0 or AngleDegrees ≤ 0. The direction is flattened to the XY plane (GetSafeNormal); two boundary normals rotate around UpVector by ±(90° + angle/2). The geometry test is: 2D radius (center to closest XY point in the cell AABB) + Z height (cell Z range overlapping center ± half height) + half-plane tests (projected radius of the cell AABB onto each normal; the angle test only runs when AngleDegrees < 360). Agent/Obstacle only; 7 wrappers forward in one line.

视锥索敌的单元收集核心:SelectionPoints 必须恰好 4 个,否则 UE_LOG 警告并返回空。由视点 + 4 个选择点构造 4 个视锥平面(点序环绕);粗筛包围盒(视点 + 选择点,按最大单元尺寸外扩)确定坐标范围;几何测试为逐平面"中心到平面距离 ≥ -投影半径"(单元 AABB 在平面法线上的投影半径)。仅支持 Agent/Obstacle;7 个包装器一行转发。

Core cell collection for view-frustum traces: SelectionPoints must be exactly 4, otherwise a UE_LOG warning fires and an empty array is returned. Four frustum planes are built from the view point and the 4 selection points (wrapped order); the broad-phase bounding box (view point + selection points, expanded by the max cell size) defines the coord range; the geometry test per plane is "center-to-plane distance ≥ -projected radius" (cell AABB projected onto the plane normal). Agent/Obstacle only; 7 wrappers forward in one line.

圆锥索敌的单元收集核心:Length ≤ 0 或 AngleDegrees ≤ 0 返回空。粗筛包围盒从原点延伸到基底中心(半径 = 长度 × tan(角));几何测试:单元盒包含原点则直接命中,否则按单元中心算"轴向投影在 [-半长, 长度+半长] 内且垂直距离 ≤ (该投影处的圆锥半径 + 半长)"。

Core cell collection for cone traces: empty when Length ≤ 0 or AngleDegrees ≤ 0. The broad-phase AABB spans from the origin to the base center (radius = length × tan(angle)); geometry test: a cell box containing the origin hits directly, otherwise the cell center's axial projection must lie within [-half-extent, length + half-extent] and its perpendicular distance must be ≤ (cone radius at that projection + half-extent).

基于分离轴定理(SAT)的 AABB 与 OBB 相交测试:测试 AABB 3 轴 + OBB 3 轴 + 9 个叉积轴(近零叉积跳过),共最多 15 轴;任一轴上投影分离即返回 false。供 BoxTraceForCells_Internal 与 BoxSweepForCells_Internal 的几何测试使用。

AABB vs OBB intersection test based on the Separating Axis Theorem (SAT): tests the 3 AABB axes + 3 OBB axes + 9 cross-product axes (near-zero cross products skipped), up to 15 axes total; projection separation on any axis returns false. Used by the geometry tests of BoxTraceForCells_Internal and BoxSweepForCells_Internal.

静态视线检查:从 Origin 到目标表面(终点 = 目标位置 - 方向 × 目标半径,与扇形索敌的障碍过滤约定一致——贴墙目标从 agent 一侧仍算可见)做 object-type trace。障碍列表为空或世界解析失败时视为可见返回 true;CheckRadius ≤ 0 用 LineTraceSingleByObjectType,否则用 SweepSingleByObjectType(球形)。返回 true 表示视线未被遮挡。抽出来供 Trace 处理器的 skip-trace 预检、玩家目标路径与 BP 代码共享。

Static line-of-sight check: an object-type trace from Origin to the target's surface (endpoint = target location - direction × TargetRadius, matching the sector trace's obstacle-filter convention — a target against a wall still counts as visible from the agent's side). An empty obstacle list or failed world resolution counts as visible (returns true); CheckRadius ≤ 0 uses LineTraceSingleByObjectType, otherwise SweepSingleByObjectType (spherical). Returns true when the line is unblocked. Extracted so the Trace processor's skip-trace pre-check, the player-target path, and BP code share one implementation.

已废弃的 BP 节点(DeprecatedFunction,提示改用 MassBattleFuncLib 对应函数)。实体级球体索敌:重置输出,KeepCount == 0 或子系统缺失时提前返回;构造 FTraceContext(IgnoreSet 忽略集、SeenHashes 按 UID 去重多格 agent、缓存的 EntityQuery、方向桶剪枝状态);几何检查用"球心到胶囊轴线最近点 ≤ 半径 + ColliderRadius";经 ExecuteTraceWithPriority 按优先级分层执行;最后 RewriteEntityLocationsFromLocating 把结果位置改写为权威 FLocating::Location,Hit = !Results.IsEmpty();编辑器下按 DrawDebugConfig 绘制调试形状。

Deprecated BP node (DeprecatedFunction; the message points to the corresponding MassBattleFuncLib function). Entity-level sphere trace: resets outputs, early-returns when KeepCount == 0 or a subsystem is missing; builds the FTraceContext (IgnoreSet, SeenHashes UID dedup for multi-cell agents, cached EntityQuery, direction-bucket prune state); the geometry check is "closest point on the capsule axis to the sphere center ≤ radius + ColliderRadius"; executes per priority level via ExecuteTraceWithPriority; finally RewriteEntityLocationsFromLocating rewrites result positions to the authoritative FLocating::Location and Hit = !Results.IsEmpty(); in-editor debug shapes follow DrawDebugConfig.

已废弃的实体级球体扫掠节点。几何检查:胶囊轴与扫掠线段的 SegmentDistToSegmentSafe 距离 ≤ 半径 + ColliderRadius;点状实体退化为线段最近点测试。其余流程与 SphereTraceForAgents 相同。

Deprecated entity-level sphere sweep node. Geometry check: SegmentDistToSegmentSafe distance between the capsule axis and the sweep segment ≤ radius + ColliderRadius; point entities degenerate to a closest-point-on-segment test. Otherwise identical flow to SphereTraceForAgents.

已废弃的实体级扇形索敌节点。几何检查:2D 半径(Radius + ColliderRadius)、角度(Angle < 360 时方向点积 ≥ cos(角/2),方向压平到 XY)、Z 高度(半高 + 胶囊半长)。其余流程同前。

Deprecated entity-level sector trace node. Geometry check: 2D radius (Radius + ColliderRadius), angle (direction dot product ≥ cos(angle/2) when Angle < 360; direction flattened to XY), and Z height (half height + capsule half length). Otherwise identical flow.

已废弃的实体级 OBB 盒索敌节点。几何检查:实体位置旋转进盒局部空间后,与盒的 AABB 钳制最近点距离 ≤ 胶囊包围球半径(有轴时取半径与半高的平方和开方)。

Deprecated entity-level OBB box trace node. Geometry check: the entity position is unrotated into box-local space, and the distance to the clamped box AABB must be ≤ the capsule's bounding-sphere radius (sqrt of radius² + half-height² when the capsule has a shaft).

已废弃的实体级 OBB 盒扫掠节点。Start == End 直接转发 BoxTraceForAgents;否则取实体到扫掠线最近点,在该点做与 BoxTraceForAgents 相同的局部空间盒测试。

Deprecated entity-level OBB box sweep node. Start == End forwards to BoxTraceForAgents; otherwise the closest point on the sweep line is found and the same local-space box test as BoxTraceForAgents runs there.

静态胶囊索敌(1.19.20 新增,非废弃;BP 经 MassBattleFuncLib::CapsuleTraceForAgents 暴露):以 Origin 为中心、轴向沿 Orientation 的 up、轴杆半高 HalfHeight、半径 Radius 的胶囊——用于竖立光柱、站桩 AoE 体积与竖向光束检测。流程与其它实体级 trace 一致:重置输出,KeepCount == 0 或子系统缺失时提前返回;构造 FTraceContext(IgnoreSet、SeenHashes 按 UID 去重、缓存的 EntityQuery、方向桶剪枝状态);FetchCells 用 CapsuleTraceForAgentCells(/Sorted);几何检查为精确的胶囊 vs 胶囊——查询轴段与 agent 碰撞体轴段的 SegmentDistToSegmentSafe 距离 ≤ Radius + ColliderRadius(ColliderHeight > 0 时 agent 轴段为 ±Up×半高,否则退化为点);经 ExecuteTraceWithPriority 分层执行,RewriteEntityLocationsFromLocating 改写结果位置,Hit = !Results.IsEmpty();编辑器下绘制调试胶囊(高 = HalfHeight×2 + Radius×2)。

Static capsule trace for agent entities (new in 1.19.20, not deprecated; exposed to Blueprint via MassBattleFuncLib::CapsuleTraceForAgents): a capsule centered at Origin, axis along Orientation's up, shaft half-height HalfHeight, radius Radius — for upright light pillars, standing AoE volumes and vertical beam checks. Same flow as the other entity-level traces: resets outputs, early-returns when KeepCount == 0 or a subsystem is missing; builds the FTraceContext (IgnoreSet, SeenHashes UID dedup, cached EntityQuery, direction-bucket prune state); FetchCells uses CapsuleTraceForAgentCells (/Sorted); the geometry check is an exact capsule-vs-capsule test — SegmentDistToSegmentSafe distance between the query axis segment and the agent collider axis segment ≤ Radius + ColliderRadius (the agent axis segment is ±Up×half-height when ColliderHeight > 0, otherwise it degenerates to a point); executes per priority level via ExecuteTraceWithPriority, rewrites result positions with RewriteEntityLocationsFromLocating, Hit = !Results.IsEmpty(); in-editor draws a debug capsule (height = HalfHeight×2 + Radius×2).

胶囊扫掠索敌(1.19.20 新增,非废弃;BP 经 MassBattleFuncLib::CapsuleSweepForAgents 暴露):Start == End(KINDA_SMALL_NUMBER 容差)零长度扫掠直接转发 CapsuleTraceForAgents。否则查询轴段取扫掠起点处(Start ± Up×HalfHeight);几何检查为精确的 agent 碰撞体胶囊 vs 扫掠体积——agent 轴段须进入查询轴与扫掠向量 [0, End-Start] 张成的平行四边形的 (Radius + ColliderRadius) 邻域(SegmentWithinDistanceOfParallelogram,即胶囊轴与扫掠段的精确闵可夫斯基和再膨胀 agent 碰撞体半径);输出位置取扫掠段上的最近接近点(与盒扫约定一致)。其余流程与 CapsuleTraceForAgents 相同;编辑器下绘制两端点胶囊(盒扫画两端盒)。

Swept-capsule trace for agent entities (new in 1.19.20, not deprecated; exposed to Blueprint via MassBattleFuncLib::CapsuleSweepForAgents): a zero-length sweep (Start == End within KINDA_SMALL_NUMBER) forwards directly to CapsuleTraceForAgents. Otherwise the query axis segment sits at the sweep start (Start ± Up×HalfHeight); the geometry check is an exact agent-collider-capsule vs swept-volume test — the agent axis segment must come within (Radius + ColliderRadius) of the parallelogram spanned by the query axis and the sweep vector [0, End-Start] (SegmentWithinDistanceOfParallelogram, i.e. the exact Minkowski sum of the capsule axis and the sweep segment inflated by the agent's collider radius); output locations use the closest approach on the sweep segment (same convention as the box sweep). Otherwise identical flow to CapsuleTraceForAgents; in-editor draws the two endpoint capsules (the box sweep draws the two endpoint boxes).

已废弃的实体级视锥索敌节点。SelectionPoints.Num() < 4 直接返回;由视点与 4 点构造 4 个视锥平面,几何检查为逐平面"平面距离 ≥ -测试半径"(测试半径 = 胶囊包围球半径)。

Deprecated entity-level view-frustum trace node. Returns immediately when SelectionPoints.Num() < 4; builds 4 frustum planes from the view point and the 4 points; the geometry check is per-plane "plane distance ≥ -test radius" (test radius = capsule bounding-sphere radius).

已废弃的实体级圆锥索敌节点。几何检查:轴向投影在 [-测试半径, 长度 + 测试半径] 内,且到轴线的垂直距离 ≤ 该投影处圆锥半径 + 测试半径。

Deprecated entity-level cone trace node. Geometry check: the axial projection must lie within [-test radius, length + test radius], and the perpendicular distance to the axis must be ≤ cone radius at that projection + test radius.

已废弃的实体数组排序节点。空数组或 SortMode == None 直接返回;Randomize:确定性模式用 RandomStream 初始种子 ^ hash(SortOrigin) 派生局部流做 Fisher-Yates 洗牌,否则 Algo::RandomShuffle;NearToFar / FarToNear:逐实体读 FLocating::Location 算距离(无效实体距离 FLT_MAX)、FDeterminism::UniqueID 作 tiebreak,排序后原地写回。

Deprecated entity array sort node. Empty array or SortMode == None returns immediately; Randomize: deterministic mode derives a local stream from RandomStream initial seed ^ hash(SortOrigin) and Fisher-Yates shuffles, otherwise Algo::RandomShuffle; NearToFar / FarToNear: per-entity distance from FLocating::Location (invalid entities get FLT_MAX) with FDeterminism::UniqueID tiebreak, sorted and written back in place.

CAS 原子最大值更新 MaxAgentNeighborExtent(relaxed 语义)。供 ExecuteRegistrationCommand 的 RegisterAgent 分支调用。

CAS atomic-max update of MaxAgentNeighborExtent (relaxed semantics). Called from the RegisterAgent branch of ExecuteRegistrationCommand.