1.2.5.2

v1.2.5.2 — 修复：动力传感器扳手旋转（同轴滚转 · 点上/下保倾偏航）· 贴地放置修正 · 倒置朝下时屏幕读数翻正 · 行走时视角摇晃导致 HUD 虚像晃动 · 切换游戏语言后 HUD 文字变乱线 · 节点分类重构 · 视口裁剪 / 菜单命中 · 数控齿轮箱轴面常在 / 扳手回归官方 · 变速器过载后输出恢复 · 输出指令正反转与负行程 · 编码器清零改节点体引脚 · PID/PID_POWER 积分死区 · 频道名旧草稿回写（#10）· 去掉整图保存覆盖（#17）· 封装子图节点数据同步 · 封装子图展开状态跨玩家同步 · 参数输入框实时同步 · 封装子图顶栏合并 · 占用封装禁删 · 键位系统扩容（删除连线 + 像素编辑器动作）· 统一键鼠序列绑定 · 旧键位配置迁移 · 删节点撤销/重做恢复连线（pinId + 占用检测）· 变速器代理态显示（值盒「(代理)」+ 滚轮拒收） / Fix: Kinetic Gauge Wrench Rotation (Shaft Roll · Tilt-Preserving Yaw) · Floor Placement · Inverted Mount Text Upright · View Bobbing Wobble & Garbled HUD Text After a Language Switch · Node Category Refactor · Viewport Cull / Menu Hit · CNC Gearbox Shaft Faces Always Present / Wrench Back to Official · Transmission Output Recovers After Overload · Output-Command Forward-Reverse & Negative Travel · Encoder Reset on the Node Body · PID/PID_POWER Integral Deadband · channel-name stale-draft write-back (#10) · drop whole-graph save overwrite (#17) · sub-graph node data sync · sub-graph expansion state sync across players · live param-box sync · sub-graph top-bar merge · occupied-encapsulation delete guard · display-mode top-bar merge · display lock follows selection · smooth remote display-layout drags · key-binding system gains delete-wire and pixel-editor actions · unified sequence bindings (legacy key-binding configs migrate) · undo/redo node-delete restores wires (pinId + occupancy check) · transmission proxy display (value-box "(proxy)" + wheel refused while node-driven)

🔧 动力传感器扳手 / Kinetic Gauge Wrench

Change / 变更	Description / 说明
🧰 行为变更：三语义	① 轴端面（点击轴 ∥ 旋转轴）→ 同轴四态按角点序（右上→右下→左下→左上）滚转 90°，轴不动——对官方的有意偏离（官方端面点击会翻轴断连）；② 点上/下 → 沿当前倾侧偏航 90°，屏幕保持朝上/朝下不翻面（第二处偏离）；③ 其余面 → 与 Create `IWrenchable` 默认逐字相同（绕所点面的轴转；facing 同轴面翻 `axis_along_first`）。环表提纯在 `KineticGaugeStates.nextInShaftRoll`/`nextInYaw`（纯函数），`KineticGaugePlacementTest` 钉死每步绕轴 90°、`getClockWise` 同向、4 步闭合 / ① shaft-end: roll through the four same-shaft states in corner order, 90° per step, shaft fixed (deliberate deviation — official pivots the shaft); ② Y-face: yaw staying on the current tilt (second deviation); ③ other faces verbatim Create's `IWrenchable` default. Rings are pure functions in `KineticGaugeStates`, pinned for 90° steps, getClockWise sense and 4-cycle closure.
🐛 Y 轴滚转环序 (bug 修复)	竖置轴的滚转环曾写成 NW→SW→NE→SE——第 2/4 步是 180° 对角跳（「互不重合」钉子查不出顺序错）。现改为 getClockWise 环（NW→NE→SE→SW），并由环序钉子（每步法线绕轴恰转 90°）锁死 / The vertical-shaft roll ring hopped 180° on steps 2 and 4; now a getClockWise cycle, locked by ring-order pins.
🐛 横置点上/下偏航 (bug 修复)	① 竖置 facing 点上/下以前只在两个 `along_first` 态来回 → 现沿朝下环保倾偏航（倒置屏不翻回朝上）；② 四个横置朝向里有两个屏幕朝下反了 → 重排 blockstate：W/N/E/S 全部 x=0 上仰讲台（up-west/north/east/south），偏航四步屏幕都朝玩家。`KineticGaugePlacementTest` 钉死 / Vertical-facing Y-click used to toggle two states; horizontal yaw had two inverted screens. Table now gives all four horizontals an up-tilted lectern pose.
🐛 同轴四态旋转重合 (bug 修复)	`up+false` 与 `west+false` 曾同为 `x=0,y=0`，`north+true` 与 `up+true` 同为 `x=0,y=90` —— 扳手绕轴循环里连着两步长得一样（实测「右上→右上→左下→右下」）。现同一旋转轴的 4 个状态落到 4 个互不重合的 `(model,x,y)`（轴 Z：`(0,0)/(0,180)/(180,180)/(180,0)`；轴 X：`y∈{90,270}×x∈{0,180}`），与 `Direction.getClockWise` 同向；`KineticGaugeStatesTest.eachShaftAxisHasFourDistinctRotations` 钉死 / Two states used to share `x=0,y=0` (and another pair `x=0,y=90`), so a wrench roll around the shaft stuttered. Each shaft axis now maps its 4 states to 4 distinct `(model,x,y)` triples, matching `getClockWise`; pinned by `eachShaftAxisHasFourDistinctRotations`.
🧹 清理	`wrenchAction` 分类纯函数与中版的刚体反查（`rigidSideTarget`/`isDisplayFace`/`isSteepLook`）删除；环表提纯为 `KineticGaugeStates` 纯函数并补环序钉子 / The classifier and the interim rigid-lookup helpers are gone; the rings are pure functions with order pins.

🧭 贴地/贴顶放置 / Floor & Ceiling Placement

Change / 变更 Description / 说明

🧭 放置重排 facing 恒为水平（贴墙=点击面，贴地/贴顶=水平正对玩家，不再因陡视变 up/down——2 态换不出 4 偏航角）；横置/竖置由 axis_along_first 承担：平视=横置讲台，陡视（nearestLooking 轴竖直）=竖直轴平板；有结构轴时优先对齐（RotatedPillarKineticBlock.getPreferredAxis）。俯仰判定用 nearest-looking 而非 getXRot——Sable 会把 orderedByNearest mixin 到子世界局部系，世界俯仰角在旋转结构上会错（2026-09 实机） / Facing is always horizontal on placement; lectern vs vertical-shaft is the steep-look gate on axis_along_first; structure axis preferred when present; nearest-looking instead of world pitch for Sable-rotated structures.

Change / 变更	Description / 说明
🧭 放置重排	`facing` 恒为水平（贴墙=点击面，贴地/贴顶=水平正对玩家，不再因陡视变 up/down——2 态换不出 4 偏航角）；横置/竖置由 `axis_along_first` 承担：平视=横置讲台，陡视（nearestLooking 轴竖直）=竖直轴平板；有结构轴时优先对齐（`RotatedPillarKineticBlock.getPreferredAxis`）。俯仰判定用 nearest-looking 而非 `getXRot`——Sable 会把 orderedByNearest mixin 到子世界局部系，世界俯仰角在旋转结构上会错（2026-09 实机） / Facing is always horizontal on placement; lectern vs vertical-shaft is the steep-look gate on `axis_along_first`; structure axis preferred when present; nearest-looking instead of world pitch for Sable-rotated structures.

🖥️ 倒置屏幕读数 / Inverted Mount Readout

Fix / 修复 Description / 说明

🙃 倒置/朝下读数翻正 (bug 修复) facing=DOWN 走 blockstate x:180，蓝屏正确朝下，但动态读数随整机一起颠倒（倒置挂墙/朝下时上下翻转）。渲染入口改为 KineticGaugeStates.displayPanel：x:180 时把字形右/上在面板平面内再转 180°（法线不动、右手系保持），从屏幕外侧看文字恢复正立；12 个状态的世界空间文字上方向 Y>0 由 KineticGaugePlacementTest 钉死 / facing=DOWN applies blockstate x:180, so the blue face correctly points down but the dynamic readout flipped with the whole unit. Rendering now goes through KineticGaugeStates.displayPanel, which spins the glyph right/up another 180° in the panel plane on x:180 (normal untouched, still right-handed) so text reads upright from outside. All 12 states' world-space text-up Y>0 is pinned in KineticGaugePlacementTest.

Fix / 修复	Description / 说明
🙃 倒置/朝下读数翻正 (bug 修复)	`facing=DOWN` 走 blockstate `x:180`，蓝屏正确朝下，但动态读数随整机一起颠倒（倒置挂墙/朝下时上下翻转）。渲染入口改为 `KineticGaugeStates.displayPanel`：`x:180` 时把字形右/上在面板平面内再转 180°（法线不动、右手系保持），从屏幕外侧看文字恢复正立；12 个状态的世界空间文字上方向 Y>0 由 `KineticGaugePlacementTest` 钉死 / `facing=DOWN` applies blockstate `x:180`, so the blue face correctly points down but the dynamic readout flipped with the whole unit. Rendering now goes through `KineticGaugeStates.displayPanel`, which spins the glyph right/up another 180° in the panel plane on `x:180` (normal untouched, still right-handed) so text reads upright from outside. All 12 states' world-space text-up Y>0 is pinned in `KineticGaugePlacementTest`.

🔌 编码器清零改节点体引脚 / Encoder Reset on the Node Body

Change / 变更 Description / 说明

🔌 行为变更 ENCODER 的清零从编辑区参数改为节点体输入引脚（电平触发语义不变：拉高清零、持续拉高保持归零；未接线 = 0）。旧连线 pinId 本就是索引 "0"，V5→V6 迁移显式钉到 0 号输入并清掉过期 reset 参数 / The ENCODER reset moves from an edit-area parameter to a body input pin (level-triggered semantics unchanged; unwired = 0). Old wires already carry pinId "0"; the V5→V6 migration pins them to input 0 and drops the stale reset param.

Change / 变更	Description / 说明
🔌 行为变更	ENCODER 的清零从编辑区参数改为节点体输入引脚（电平触发语义不变：拉高清零、持续拉高保持归零；未接线 = 0）。旧连线 pinId 本就是索引 `"0"`，V5→V6 迁移显式钉到 0 号输入并清掉过期 `reset` 参数 / The ENCODER reset moves from an edit-area parameter to a body input pin (level-triggered semantics unchanged; unwired = 0). Old wires already carry pinId `"0"`; the V5→V6 migration pins them to input 0 and drops the stale `reset` param.

🎛️ PID / PID_POWER 积分死区 / Integral Deadband

Change / 变更	Description / 说明
🎛️ 积分死区 deadband	新增可编辑/可连线参数 `deadband`（默认 0.001）：`

🎛️ 输出指令正/反转与负行程 / Output-Command Forward-Reverse & Negative Travel

Change / 变更	Description / 说明
🐛 负数恒无效 (bug 修复)	`MotionQuota.of` 用 `Math.max(0f, amount)` 把负行程打成 0 配额 → 负数 ROTATE/MOVE 当帧完成、形同假值。现行程量取 `Math.abs`，符号只表示方向 / Negative ROTATE/MOVE quotas were clamped to 0 and completed instantly. Travel is now `Math.abs`; the sign only encodes direction.
🔄 负行程 = 反转	ROTATE/MOVE 入栈快照的负数值 → 执行期间输出面 `getRotationSpeedModifier = -1`（官方 Gearshift 反转语义），编码器积分/瞬时速度同步取输出符号；WAIT / 空闲 / CLUTCH 常接合不反转 / A snapshotted negative ROTATE/MOVE runs the output face at modifier -1 (official Gearshift reverse) and the encoder books the output sign; WAIT / idle / standing CLUTCH never reverse.
🔘 正/反转按钮	位移 / 旋转 / 目标转速 / 转速控制编辑区新增「正转/反转」独立 `rev` 开关（不改数值 EditBox——数值可被引脚覆盖）：MOVE/ROTATE 在入栈时对快照值取反；TX_OUT/SPEED_CTRL 在求值结果上取反（与 SPEED_CTRL 的 dir 引脚 XOR）/ Independent `rev` toggle on Move/Rotate/Target RPM/Speed Control that does not edit the value EditBox (the value is wire-overridable). MOVE/ROTATE negate the snapshotted value at enqueue; TX_OUT/SPEED_CTRL negate the eval result (XOR with SPEED_CTRL's dir pin).

🔧 数控齿轮箱扳手回归官方 / CNC Gearbox Wrench Back to Official

Change / 变更 Description / 说明

🧹 行为变更 扳手点轴端面「翻转输入端」与 action bar 字幕（input face: …）已移除——输入/输出面由放置感知 + autoSenseInputFace 自动识别（官方从动件同款）。任意面扳手走 IWrenchable 默认旋转；disengageForFlip / resyncKineticsAfterFlip 一并删除 / Wrench-on-end-face "flip input end" and its action-bar subtitle are gone — input/output faces are auto-sensed (placement + autoSenseInputFace), like a vanilla driven member. Wrenching any face is the official IWrenchable rotate; disengageForFlip / resyncKineticsAfterFlip deleted with it.

Change / 变更	Description / 说明
🧹 行为变更	扳手点轴端面「翻转输入端」与 action bar 字幕（`input face: …`）已移除——输入/输出面由放置感知 + `autoSenseInputFace` 自动识别（官方从动件同款）。任意面扳手走 `IWrenchable` 默认旋转；`disengageForFlip` / `resyncKineticsAfterFlip` 一并删除 / Wrench-on-end-face "flip input end" and its action-bar subtitle are gone — input/output faces are auto-sensed (placement + `autoSenseInputFace`), like a vanilla driven member. Wrenching any face is the official `IWrenchable` rotate; `disengageForFlip` / `resyncKineticsAfterFlip` deleted with it.

🎛️ 变速器过载后输出不恢复 / Transmission Output Stays Dead After Overload

Fix / 修复 Description / 说明

🐛 过载拆输出 (bug 修复) 输出端负载超容 → 整网 overStressed，官方 getSpeed() 恒 0。cleanOrphanedKineticState 用 getSpeed() 判源死活，把仍在转的源打成失速，每 tick detach 拆掉输出侧子树（「过载后自动断开应力」）。输入端应力恢复后 getSpeed() 变回非 0，预检早退认为健康，不会接回已拆下游；只有改转速触发 updateTargetRotation 全量拆建才恢复（「改转速才可以刷新输出」）。现源健康/幻影速判定改用 getTheoreticalSpeed()（= speed 字段，官方 validateKinetics 同款）/ Output load past capacity trips network-wide overStressed, and official getSpeed() is hard-zeroed. The orphaned-state pre-check used getSpeed() for source health, so a live source looked dead and every tick detached, tearing down the output subtree. Once capacity recovered the pre-check saw a spinning source and early-returned — never re-attaching what it tore down; only a target change's full rebuild did. Health/phantom checks now use getTheoreticalSpeed() (the raw field, matching official validateKinetics).

Fix / 修复	Description / 说明
🐛 过载拆输出 (bug 修复)	输出端负载超容 → 整网 `overStressed`，官方 `getSpeed()` 恒 0。`cleanOrphanedKineticState` 用 `getSpeed()` 判源死活，把仍在转的源打成失速，每 tick detach 拆掉输出侧子树（「过载后自动断开应力」）。输入端应力恢复后 `getSpeed()` 变回非 0，预检早退认为健康，不会接回已拆下游；只有改转速触发 `updateTargetRotation` 全量拆建才恢复（「改转速才可以刷新输出」）。现源健康/幻影速判定改用 `getTheoreticalSpeed()`（= `speed` 字段，官方 `validateKinetics` 同款）/ Output load past capacity trips network-wide `overStressed`, and official `getSpeed()` is hard-zeroed. The orphaned-state pre-check used `getSpeed()` for source health, so a live source looked dead and every tick detached, tearing down the output subtree. Once capacity recovered the pre-check saw a spinning source and early-returned — never re-attaching what it tore down; only a target change's full rebuild did. Health/phantom checks now use `getTheoreticalSpeed()` (the raw field, matching official `validateKinetics`).

⚙️ 数控齿轮箱输出端轴面常在 / CNC Gearbox Output Shaft Face Always Present

Fix / 修复 Description / 说明

🐛 放轴不吸附 (bug 修复) 无指令/空闲自动分离时 hasShaftTowards 把输出端轴面整段否掉——Create 的放置吸附（RotatedPillarKineticBlock.getPreferredAxis）与传动耦合（RotationPropagator 要求双方轴面）都认它，贴上去的传动轴不吸、不连。现改为官方 AbstractEncasedShaftBlock 同款：两端轴面恒在（只看轴向）；离合隔离改走官方 SplitShaftBlockEntity.getRotationSpeedModifier（输出面分离时 0）+ 既有 detach/attachKinetics。BE 基类升为 SplitShaftBlockEntity / While idle the clutch used to return false from hasShaftTowards on the output face, which Create's placement snap and axis coupling both consult — shafts placed against the block would not attach. Now axis-only like official AbstractEncasedShaftBlock; isolation moves to official SplitShaftBlockEntity.getRotationSpeedModifier (0 on the output face while disengaged) plus the existing detach/attachKinetics. The BE base class is now SplitShaftBlockEntity.

Fix / 修复	Description / 说明
🐛 放轴不吸附 (bug 修复)	无指令/空闲自动分离时 `hasShaftTowards` 把输出端轴面整段否掉——Create 的放置吸附（`RotatedPillarKineticBlock.getPreferredAxis`）与传动耦合（`RotationPropagator` 要求双方轴面）都认它，贴上去的传动轴不吸、不连。现改为官方 `AbstractEncasedShaftBlock` 同款：两端轴面恒在（只看轴向）；离合隔离改走官方 `SplitShaftBlockEntity.getRotationSpeedModifier`（输出面分离时 0）+ 既有 detach/attachKinetics。BE 基类升为 `SplitShaftBlockEntity` / While idle the clutch used to return false from `hasShaftTowards` on the output face, which Create's placement snap and axis coupling both consult — shafts placed against the block would not attach. Now axis-only like official `AbstractEncasedShaftBlock`; isolation moves to official `SplitShaftBlockEntity.getRotationSpeedModifier` (0 on the output face while disengaged) plus the existing detach/attachKinetics. The BE base class is now `SplitShaftBlockEntity`.

🗂️ 节点分类与方块名单 / Node Categories & Per-Block Allowances

Change / 变更	Description / 说明
🗂️ 分类 14→21 (UI 变更)	节点菜单分组重划：`input` 拆为 `input_ctrl`/`input_view`/`input_motion`/`input_pose`，`sequential` 拆为 `sequential_acc`/`sequential_state`；新增 `transmission`/`speed`/`radar`。`ROUND` 归 `math_basic`，`SQRT/LN/LOG/EXP` 移出三角，`POSE_CONVERT`/`SPLIT` 归 `input_pose`。
📋 名单机制	10 处逐节点枚举改为 `NodeCategory` + `NodeAllowance`（分类白名单 + 类内黑名单）；新增节点落入已允许分类即自动可用。
➕ 能力补齐	程序计算机 / 数控齿轮箱净增 23 项：基础运算 11（`ADD`…`ROUND`）、比较与 `OR` 6、`FORMULA`/`INTERP`、控制类 4（`PID`/`PID_POWER`/`CLAMP`/`MAP`）。传感器净增 `POSE_CONVERT`/`SPLIT`。
➖ 能力收紧	蓝图计算机失去 `STRESS`/`RPM`（对齐文档「仅动力宿主图」）；已有图中节点不删除，仅不能新建。
🧪 回归	`NodeAllowanceMigrationTest`：覆盖比对 + 允许集差异必须等于声明变更 + 例外最小化。

🖱️ 图编辑器命中与裁剪 / Graph Editor Hit & Cull

Fix / 修复	Description / 说明
🐛 添加节点菜单命中 (bug 修复)	菜单面板（含压在顶栏上的顶部）内点击归菜单，不再被顶栏改名框/设置钮误吞；菜单改为最后渲染、盖在顶栏上，与命中顺序一致。
🐛 展开节点视口裁剪 (bug 修复)	节点体滚出屏幕、编辑区仍在画面时不再整节点剔除：体/编辑区分别判交；展开高度取有/无 `EditState` 的较大值（FORMULA MLE 动态行、ACCUMULATOR 动态字段）。

🖥️ HUD 虚像与视角摇晃 / HUD Virtual Image & View Bobbing

Fix / 修复 Description / 说明

🚶 行走时虚像晃动 (bug 修复) Minecraft 把 bobHurt/bobView 写进一个 PoseStack 后乘进投影矩阵（GameRenderer.renderLevel： matrix4f.mul(posestack.last().pose())），它作用于 view space，从不进入 BER 的 poseStack。行走时 HUD 虚像内容随之晃动（关闭视频设置里的「视角摇晃」即完全消失，已实测确认）。数值诊断（ViewBobWobbleDiagTest，16 相位采样整圈步态）定位根因：深度锚定在预 bob 坐标上做，GPU 侧把 bob 乘回顶点时，bob 的视空间平移按玻璃深度（近，~3 格）做透视除 → 虚像按近物视差随玻璃晃动（0.038 NDC ≈ 1.3°），与设计文档 §9.1「无限远聚焦」矛盾。修复（用户 2026-09-19 选定方案 C「远处虚像」）：锚定改为沿 bob 后射线（MonitorClipMath.anchoredEmit，先乘 bob 再锚定，发射时预乘 viewRotInv·B⁻¹ 让 GPU 的 bob 恰好抵消）——屏幕位置保持远处画布投影（平移视差按 D+gz 除，小 30+ 倍），只剩与全世界一致的旋转分量（0.0113 NDC，共形下限 0.0123）；玩家屏幕定位遮罩的投影眼改为视觉眼 B⁻¹·0（bob 后视线束在预 bob 视空间的汇聚点）——裁剪窗口与玻璃的落差精确为 0.0000，窗口贴住玻璃；深度仍落玻璃平面，遮挡关系不变。文字字形路径（anchorTextVertex）同步修改，锚定数学迁入 MonitorClipMath 纯函数 / Minecraft folds bob into the PROJECTION matrix, acting in view space; it never reaches the BER poseStack — and the HUD virtual-image content wobbled while walking (disabling "View Bobbing" removes it entirely, confirmed). Numeric diagnosis (ViewBobWobbleDiagTest, 16-phase sampling of a full gait cycle) found the root cause: anchoring in pre-bob coords makes the GPU-side bob translation perspective-divide at the glass depth (near, ~3 blocks) → the image swayed with near parallax, glued to the glass (0.038 NDC ≈ 1.3°), contradicting design-doc §9.1 infinite focus. Fix (user-chosen option C, "far virtual image"): anchor along the post-bob ray (MonitorClipMath.anchoredEmit — apply bob first, then anchor; emit premultiplied by viewRotInv·B⁻¹ so the GPU's bob lands exactly on the desired view position) — the screen position keeps the far-canvas projection (translation parallax divides at D+gz, 30+× steadier) and only the world-rotation sway remains (0.0113 NDC, conformal floor 0.0123); the player-screen mask now projects from the visual eye B⁻¹·0 (where the bobed view rays converge) — the clip window tracks the glass with a gap of exactly 0.0000; depth still lands on the glass plane, occlusion unchanged. The text-glyph path (anchorTextVertex) is fixed the same way, and the anchoring math moved into MonitorClipMath pure functions.

Fix / 修复	Description / 说明
🚶 行走时虚像晃动 (bug 修复)	Minecraft 把 `bobHurt`/`bobView` 写进一个 PoseStack 后乘进投影矩阵（`GameRenderer.renderLevel`： `matrix4f.mul(posestack.last().pose())`），它作用于 view space，从不进入 BER 的 poseStack。行走时 HUD 虚像内容随之晃动（关闭视频设置里的「视角摇晃」即完全消失，已实测确认）。数值诊断（`ViewBobWobbleDiagTest`，16 相位采样整圈步态）定位根因：深度锚定在预 bob 坐标上做，GPU 侧把 bob 乘回顶点时，bob 的视空间平移按玻璃深度（近，~3 格）做透视除 → 虚像按近物视差随玻璃晃动（0.038 NDC ≈ 1.3°），与设计文档 §9.1「无限远聚焦」矛盾。修复（用户 2026-09-19 选定方案 C「远处虚像」）：锚定改为沿 bob 后射线（`MonitorClipMath.anchoredEmit`，先乘 bob 再锚定，发射时预乘 `viewRotInv·B⁻¹` 让 GPU 的 bob 恰好抵消）——屏幕位置保持远处画布投影（平移视差按 D+gz 除，小 30+ 倍），只剩与全世界一致的旋转分量（0.0113 NDC，共形下限 0.0123）；玩家屏幕定位遮罩的投影眼改为视觉眼 `B⁻¹·0`（bob 后视线束在预 bob 视空间的汇聚点）——裁剪窗口与玻璃的落差精确为 0.0000，窗口贴住玻璃；深度仍落玻璃平面，遮挡关系不变。文字字形路径（`anchorTextVertex`）同步修改，锚定数学迁入 `MonitorClipMath` 纯函数 / Minecraft folds bob into the PROJECTION matrix, acting in view space; it never reaches the BER poseStack — and the HUD virtual-image content wobbled while walking (disabling "View Bobbing" removes it entirely, confirmed). Numeric diagnosis (ViewBobWobbleDiagTest, 16-phase sampling of a full gait cycle) found the root cause: anchoring in pre-bob coords makes the GPU-side bob translation perspective-divide at the glass depth (near, ~3 blocks) → the image swayed with near parallax, glued to the glass (0.038 NDC ≈ 1.3°), contradicting design-doc §9.1 infinite focus. Fix (user-chosen option C, "far virtual image"): anchor along the post-bob ray (MonitorClipMath.anchoredEmit — apply bob first, then anchor; emit premultiplied by viewRotInv·B⁻¹ so the GPU's bob lands exactly on the desired view position) — the screen position keeps the far-canvas projection (translation parallax divides at D+gz, 30+× steadier) and only the world-rotation sway remains (0.0113 NDC, conformal floor 0.0123); the player-screen mask now projects from the visual eye B⁻¹·0 (where the bobed view rays converge) — the clip window tracks the glass with a gap of exactly 0.0000; depth still lands on the glass plane, occlusion unchanged. The text-glyph path (anchorTextVertex) is fixed the same way, and the anchoring math moved into MonitorClipMath pure functions.

设计决策（用户 2026-09-19 选定，钉在 ViewBobWobbleDiagTest / ViewBobAnchorTest） Design decisions (user-chosen 2026-09-19, pinned in ViewBobWobbleDiagTest / ViewBobAnchorTest)

虚像内容 = 远处物体：行走时只随全世界一起转（bob 的旋转分量，远物参照 0.0123 NDC）， bob 平移分量必须消失。开发中曾先尝试只把遮罩眼挪到 +B·0，实测否决——遮罩眼与锚定方式必须配对（沿 bob 后射线锚定 ↔ 视觉眼 B⁻¹·0；预 bob 锚定 ↔ 原点眼）， +B·0 与任何锚定都不封闭（Y 向窗口滑差 0.046 NDC），且它没触及内容摆动的根因。 The virtual image behaves as a far object: only the world-rotation sway remains; the mask eye must be PAIRED with the anchoring (post-bob anchor ↔ visual eye B⁻¹·0; pre-bob anchor ↔ the origin eye) — the earlier +B·0 attempt closed with nothing (Y window slide 0.046 NDC) and never touched the root cause.

「补偿更差 3.7 倍」的旧测量是稻草人对比：拿静态内容对比摇晃玻璃，量到的本就是玻璃自己的晃动，不是虚像的性质。 The old "compensating drifts 3.7× more" measurement was a strawman: static content measured against a bobbing glass captures the glass's own motion, not the image's.

站定（amp=0）时 B = I，路径与修复前逐字节一致。 Standing still (amp=0) keeps bob the identity: the path is unchanged.

已知范围外 / Known limitation: 同一 PoseStack 里的 bobHurt（受伤抖动）依赖 GameRenderer 私有的 hurtTime，无法从外部重建，受伤瞬间那一下抖动不在本次修复范围（持续不到 1 秒、幅度小）。 bobHurt sits in the same PoseStack but depends on GameRenderer's private hurtTime and cannot be rebuilt from outside; that brief hurt wobble is out of scope.

🔤 手动字形与语言切换 / Manual Glyphs & Language Switch

Fix / 修复 Description / 说明

🔤 切语言后 HUD 文字变乱线 (bug 修复) 手动字形路径（不走 MultiBufferSource，直接读 BakedGlyph 几何/UV 生成顶点）曾把取到的 FontSet 与字形 RenderType 各缓存一份，且只判 == null（永不刷新）。主界面切换游戏语言（或任何资源重载）时，旧缓存的字形几何与 UV 仍指向已被替换的 atlas，文字画成一团乱线。修复粒度是 FontSet 本身：getFontSet 现在每次实取（它只是 this.fonts.apply(id) 一次 map 查找，成本可忽略）；字形 RenderType 缓存绑定取到的 FontSet 实例 / The manual-glyph path cached the fetched FontSet and the glyph RenderType on == null (never refreshed). A language switch / reload replaces the atlas, so the stale glyph geometry and UV drew the text as a tangle of lines. Fixed at FontSet granularity: getFontSet is now fetched every time (it is just this.fonts.apply(id), a map lookup — negligible), and the glyph RenderType cache is bound to the fetched FontSet instance.

Fix / 修复	Description / 说明
🔤 切语言后 HUD 文字变乱线 (bug 修复)	手动字形路径（不走 MultiBufferSource，直接读 `BakedGlyph` 几何/UV 生成顶点）曾把取到的 `FontSet` 与字形 `RenderType` 各缓存一份，且只判 `== null`（永不刷新）。主界面切换游戏语言（或任何资源重载）时，旧缓存的字形几何与 UV 仍指向已被替换的 atlas，文字画成一团乱线。修复粒度是 FontSet 本身：`getFontSet` 现在每次实取（它只是 `this.fonts.apply(id)` 一次 map 查找，成本可忽略）；字形 `RenderType` 缓存绑定取到的 FontSet 实例 / The manual-glyph path cached the fetched `FontSet` and the glyph `RenderType` on `== null` (never refreshed). A language switch / reload replaces the atlas, so the stale glyph geometry and UV drew the text as a tangle of lines. Fixed at FontSet granularity: `getFontSet` is now fetched every time (it is just `this.fonts.apply(id)`, a map lookup — negligible), and the glyph `RenderType` cache is bound to the fetched FontSet instance.

为什么不能按 Font 实例做缓存守卫（第一版踩的坑，已改正） Why the cache guard must not key on the Font instance (the first version's mistake) Minecraft.font 只在构造器赋值一次（this.font = this.fontManager.createFont()），注册为重载监听器的是 fontManager —— 所以语言切换时 Font 实例不变，被重建的是 FontManager 内部的 FontSet（FontManager.reload：fontSets.clear() 后 new FontSet，旧对象已 close()）。绑 Font 的守卫永远不会再次触发，等于没修。 Minecraft.font is assigned once in the constructor, and it is fontManager — not font — that is registered as the reload listener. A language switch therefore leaves the Font instance untouched and rebuilds the FontSet inside FontManager (fontSets.clear() then new FontSet; the old one is closed). A Font-bound guard never fires again. 附注：hudTextRenderType 旧注释写着"资源重载后自动更新"，实现却是 if (== null) —— 注释与实现不符，已按代码实际行为改正 / The old comment claimed "refreshed after a resource reload" while the code was if (== null); corrected.

无回归测试 / No regression test: 该修复依赖 Minecraft 的资源重载（切换语言），沙箱内无法驱动，故未钉单测 —— 需游戏内实测：主菜单切换语言 → 进世界看 HUD 文字。 This fix depends on Minecraft's resource reload (a language switch), which cannot be driven in the sandbox, so it is not pinned by a unit test — verify in game: switch the language in the main menu, then enter a world and look at the HUD text.

✨ 删节点撤销/重做恢复连线 / Node-Delete Undo-Redo Wire Restore

Change / 变更 Description / 说明

✨ 删节点撤销/重做恢复连线（含 pinId 占用检测） 删节点（X 单删 / Delete 多选，含多选节点间互连）原先只入栈 REMOVE_NODE，removeNode 剪掉的连线一去不返——Ctrl+Z 只回来节点空壳。现删除前把触及被删节点的连线（去重）逐条记为携带稳定 pinId 的 REMOVE_CONN（同批 + 同步协作者），批次逆序撤销 = 先 ADD_NODE 恢复节点、再按 pinId 接线；ADD_CONN 恢复时走 addConnectionWithPinIds：pinId 解析不到或输入脚已被占用则静默跳过（不覆盖现有线）。pinId 编入 op stringValue（协议零扩面）；删线/连线创建同步携带 pinId，普通删线撤销同样受益。测试：GraphOpTest（打包/解析）+ NodeGraphTest（解析、占用跳过、删后按 pinId 回接、按 pinId 删线）/ Deleting a node (X / Delete, including wires between multi-selected nodes) used to record only REMOVE_NODE — removeNode pruned the wires and Ctrl+Z brought back a bare node. Wires touching the doomed set are now recorded first (de-duplicated) as pinId-carrying REMOVE_CONN ops in the same undo batch and synced to collaborators; the reverse batch restores nodes first, then rewires by pinId. ADD_CONN restore goes through addConnectionWithPinIds: unresolvable pinId or occupied input pin is skipped silently (never clobbers). PinIds ride in the op's stringValue (no wire-format change); wire create/delete carry pinIds too, so ordinary wire-delete undo benefits as well.

Change / 变更	Description / 说明
✨ 删节点撤销/重做恢复连线（含 pinId 占用检测）	删节点（X 单删 / Delete 多选，含多选节点间互连）原先只入栈 `REMOVE_NODE`，`removeNode` 剪掉的连线一去不返——Ctrl+Z 只回来节点空壳。现删除前把触及被删节点的连线（去重）逐条记为携带稳定 pinId 的 `REMOVE_CONN`（同批 + 同步协作者），批次逆序撤销 = 先 `ADD_NODE` 恢复节点、再按 pinId 接线；`ADD_CONN` 恢复时走 `addConnectionWithPinIds`：pinId 解析不到或输入脚已被占用则静默跳过（不覆盖现有线）。pinId 编入 op `stringValue`（协议零扩面）；删线/连线创建同步携带 pinId，普通删线撤销同样受益。测试：`GraphOpTest`（打包/解析）+ `NodeGraphTest`（解析、占用跳过、删后按 pinId 回接、按 pinId 删线）/ Deleting a node (X / Delete, including wires between multi-selected nodes) used to record only `REMOVE_NODE` — `removeNode` pruned the wires and Ctrl+Z brought back a bare node. Wires touching the doomed set are now recorded first (de-duplicated) as pinId-carrying `REMOVE_CONN` ops in the same undo batch and synced to collaborators; the reverse batch restores nodes first, then rewires by pinId. `ADD_CONN` restore goes through `addConnectionWithPinIds`: unresolvable pinId or occupied input pin is skipped silently (never clobbers). PinIds ride in the op's `stringValue` (no wire-format change); wire create/delete carry pinIds too, so ordinary wire-delete undo benefits as well.

🛠️ 图编辑器与多人协作 / Graph Editor & Collaboration

Change / 变更	Description / 说明
🐛 #10 频道名旧草稿回写 (bug 修复)	输入框聚焦时面板重建会保留改名前草稿（有意为之，不打断打字），但随后防抖把旧文本当作新改名写回服务端，把权威名顶回去。现引入 `busNameUserDirty`：只有用户敲键才标脏；程序装入（初值 / 保留草稿 / 远端刷新）一律不得提交。非 dirty 的差异在失焦/关屏/编译时丢弃并提示「已丢弃未确认的频道名草稿」。守卫收在 `GraphBusEditor.shouldCommitBusName` 一处 / Focused busBox drafts preserved across panel rebuilds used to be re-committed by the debounce, overwriting the authoritative name. Now only real keystrokes set `busNameUserDirty`; programmatic loads never commit. Stale drafts are discarded with a chat hint.
🐛 #17 整图保存覆盖 (bug 修复)	编译/保存原先是 `BlueprintSavePacket` 整图 NBT 整体替换服务端图——客户端副本一旧就把服务端回退，静默冲掉并发编辑。现改为 `GraphSaveRequestPacket`（只带坐标）：服务端在自己的权威图上执行编译复位（GATE/T_FLIPFLOP/LATCH `params[1]=params[0]`，含子图）+ 环检测 + 落盘 + 全量同步。`BlueprintSavePacket` 与孤儿 `loadGraphFromBytes` / `loadEditorTag` 链一并退役删除 / Save/compile used to replace the server graph with the client snapshot. It is now a position-only request; the server runs compile semantics on its own graph. `BlueprintSavePacket` and the orphaned `loadGraphFromBytes` / `loadEditorTag` chain are retired.
🐛 封装子图节点数据同步 (bug 修复)	子图内编辑公式时，响应器用 `host.getGraph().findNode`（主图）找节点——找不到就静默 return，SET_FORMULA 根本不发送，本地框却仍显示输入；两端因此各显示各的公式/参数。另：远端 SET_PARAM 刷新用 `fieldParamIndices.get(paramIndex)` 把字段位次当下标，`fi==-1`（busBox 槽）时 `fields.get(-1)` 抛 IOOBE 中断刷新；UI 刷新未校验 `ownerNodeId` 作用域，子图/主图同 id 互相污染。现改为 `ed.getGraph()` + `indexOf` 反查 + 同作用域才刷 UI / Sub-graph formula edits silently dropped (wrong graph lookup), remote SET_PARAM UI refresh crashed on field-slot reverse lookup, and UI refresh ignored scope.
🐛 封装子图展开状态跨玩家同步 (bug 修复)	远端 `EXPAND_NODE`/`COLLAPSE_NODE` 把图数据字段 `n.expanded`（随 NBT 持久化，也是进入子图时 init 恢复的唯一依据）一起门在 `sameScope` 里——玩家停在主图时收到的子图展开/折叠 op 被整条忽略，随后进入子图就恢复出过期状态（别人展开的显示折叠、已折叠的显示展开）。现拆数据半边（`n.expanded` 恒应用到自己子图副本）与 UI 半边（展开集合/编辑状态仍仅同作用域） / Cross-scope remote expand/collapse ops left the local sub-graph copy's `n.expanded` stale (the flag was gated behind the same-scope UI guard), so entering the sub-graph restored the wrong expansion state. The data half now always applies; the UI half stays scope-gated.
🐛 参数输入框实时同步 (bug 修复)	① 远端只回显 `ff3(float)`，与正在输入的字符串不一致；② 清空/未完成输入 `parseFloat` 失败被吞、不发包，对端停在默认值；③ `ff3` 用 `Math.round(float)`（返回 `int`），`\|v\|>2147483.647` 时饱和成 `Integer.MAX_VALUE`，大数被压成 `2147483.x`。现：SET_PARAM 携带草稿原文（含空串）驱动对端 EditBox 显示（本端聚焦时不覆盖）；空/半截输入保持上次权威值只同步草稿；参数值不再进编辑状态指纹（值一变就 `createEditState`+`ff3` 会把 `000` 刷成 `0.0`、大数刷成科学计数）；失焦才归位为 `ff3` 规范格式并同步给对端，不改权威值；`ff3` 改 double 舍入；OpExecutor 对值相同的 SET_PARAM 跳过 `bumpGeneration` / Peers showed `ff3(float)` instead of the in-progress string; empty/partial input never shipped; `ff3`'s `Math.round(float)` saturated large magnitudes. SET_PARAM now carries the raw draft (incl. clear); empty input keeps the last committed number; param values are out of the edit-state fingerprint so drafts survive rebuilds; blur normalizes to `ff3` without changing the value; OpExecutor skips `bumpGeneration` when unchanged.
🐛 封装子图顶栏合并 (bug 修复)	子图的「封装编辑模式」横条（y=2..22）与每帧都画的顶栏重叠、右上 Back 按钮与设置按钮同槽互盖，且 Back 渲染在 y=4、命中却在 `TOP_BAR_H+2`（渲染/命中脱节，点画的按钮没反应）。现两作用域合并为同一根顶栏（同底色同高度）：主图 = 名称 + 设置；子图 = 模式标识「◈ 封装编辑模式 ◈ (n/N)」（超限红色 ⚠ + 提示放得下才画）+ Back 占用设置按钮的同一槽位（渲染/命中几何一致）；子图隐藏图名框并在进入时交出焦点 / The sub-graph mode strip (y=2..22) overlapped the always-drawn top bar, the Back button collided with Settings, and Back's render (y=4) and hit region (`TOP_BAR_H+2`) diverged. Both scopes now share one top bar: main = name + settings; sub = mode indicator (count / over-limit) + Back in the settings slot with matching render/hit geometry; the name box hides in the sub-graph and releases focus on entry.
🐛 占用封装禁删 (bug 修复)	主图按 X（悬停删除）/ DEL（批量删除）可以直接删除有玩家正在其子图内编辑的封装节点——对端的编辑器被连人带图拽出来。现删除前经 `GraphPresenceTracker.encapOccupied` 判定（只认节点编辑模式 presence）：占用则跳过/拒绝并给动作栏提示「无法删除：该封装内有玩家正在编辑」。与节点软锁同为建议性 presence 守卫（服务端对 presence 是纯中继）；批量删除时占用封装跳过、其余照删 / Pressing X (hover-delete) or DEL (bulk delete) on the main graph could delete an encapsulation while players were editing inside its sub-graph, yanking their editor out. Deletion now consults `GraphPresenceTracker.encapOccupied` (node-editor-mode presence only): occupied encapsulations are refused/skipped with an action-bar hint. Same advisory presence trust level as the node soft lock (the server merely relays presence); in bulk deletes only the occupied ones are skipped.
🐛 显示布局锁跟随选择 (bug 修复)	显示布局模式的软锁只有拖动才生效、松手即解锁——但本端选择框还在，队友看到的是「无锁可抢」，双方可同时选中/抢占同一元素。现选中元素经 presence `selectedNodeId` 上报，锁判定改为拖拽 id 或选中 id 任一命中即锁（与节点模式「选中 = 软锁」语义一致）；mode 1 时节点图字段（选择/多选/连线拖拽）一律置空，节点图锁列表叠加层过滤 mode 1——显示元素 id 不再以假锁污染相同 id 的节点图节点；协作叠加层的软锁描边同步改为拖拽或选中即常驻显示（此前只画拖拽，选中占锁在队友屏幕上不可见）/ The collaboration overlay's lock outline also renders while merely selected (it previously drew only during a drag, making a selection-held lock invisible to teammates); The display-layout soft lock only applied WHILE dragging and released on mouse-up, yet the local selection box stayed up — teammates saw a grabbable element and both sides could claim the same one. The selected element now rides presence `selectedNodeId`; the lock hits on dragged OR selected id (matching the node editor's selection-lock semantics). Node-graph fields are zeroed in mode 1 and the node lock overlay filters mode 1, so a display element id can no longer phantom-lock a same-id graph node.
🧹 显示布局远端拖拽平滑 (行为变更)	节点图远端拖拽有 smoothstep 插值（remote* 字段 + 每帧推进），显示布局的 `SET_DISPLAY_LAYOUT` 却直接落地坐标——队友拖组件在别人屏幕上一跳一跳。现对齐节点模式：位置变化超阈值（归一化 0.005）时启动插值（`GraphNode` 新增 `layoutLerpT/layoutStart/layoutTarget`，OpExecutor 在 animateMoves 时置起点/目标），客户端每帧推进 `layoutX/Y` 并在插值期间强制重建元素缓存（插值不 bump 代际，缓存不强制重建会吞掉动画）；缩放/旋转不插值。监视器以 100ms 节流流式发送，多帧插值让远端观感连续 / Remote display-layout drags used to land coordinates instantly and looked jumpy on other clients, while the node graph has smoothstep interpolation. SET_DISPLAY_LAYOUT now starts the interpolation past a normalized 0.005 threshold (new `layoutLerpT/layoutStart/layoutTarget` on GraphNode, armed by OpExecutor under animateMoves), the client advances layoutX/Y per frame and force-rebuilds the element cache while animating (a lerp doesn't bump the generation); scale/rotation don't lerp. With the monitor's 100ms throttled stream, remote motion reads as continuous.
🧹 显示布局模式并入合并顶栏 (行为变更)	显示器显示布局模式原在顶栏下方自绘一条工具条（`< 编辑` / 设置 / S·R）。现并入合并顶栏（与主图/封装子图同一 chrome）：左 = 「◈ 显示布局编辑 ◈」+ 选中元素的 S/R 内联编辑，右 = 设置（面板开关）+ 返回两枚 46 宽按钮；渲染与命中几何一致。旧工具条占位行收回——显示区与图层面板整体上移，可视画布更高 / The monitor display-layout mode used to draw its own strip below the top bar (`< Graph` / Settings / S·R). It is now part of the merged top bar (same chrome as the main graph / encapsulation sub-graph): left = 「◈ Display Layout Edit ◈」+ the selected element's S/R inline editors, right = Settings (panel toggle) and Back as two 46-wide buttons with matching render/hit geometry. The old strip's reserved row is reclaimed — the display area and the layer panel move up, giving a taller canvas.

⌨️ 键位系统 / Key Bindings

Change / 变更 Description / 说明

🎛️ 键位系统扩容：删除连线 + 像素编辑器 (行为变更) 键位设置系统新增 11 个可绑定动作：① 删除连线（默认 W，悬停连线按键即删；原 Tab+左键组合保留）；② 像素编辑器的画笔(B)/橡皮(E)/填充(F)/取色(I)/直线(L)/矩形(R)/抓手(H)/网格(G)/笔刷变小([)/变大(]) 全部改走共享键位系统，出厂键与旧硬编码逐项一致（唯一例外：[ / ] 的默认键取 GLFW 真值 91/93——旧硬编码的 219/221 是 Windows VK 码，经 GLFW keyPressed 永不命中，即 [ / ] 笔刷快捷键借此首次真正可用）；1..7 按轨道顺序选工具保持固定。冲突规则（等长全同必拒等）覆盖全部新动作，设置界面即时生效；键位升级为统一序列模型（键鼠交错）——每个动作只有一条触发序列，步骤既可以是按键也可以是鼠标点击，设置里直接录入任意交错队列：点键帽追加键步、点鼠标键把按钮加入队列、再次点击移出（预览行实时显示整条队列，如「Tab → 左键 → A → 右键」），「确定绑定」整条提交；引擎分纯键域/混合域独立推进——键步 keyPressed 推进、鼠标步点击推进，任一事件完成即触发，纯右键=单鼠标步、Ctrl+Z=纯键步、Tab → 左键=出厂删线组合，与框选 Tab 键序跨域并存；上下文菜单恢复键序触发（按绑定键在光标处打开，且吞掉开菜单键的字符——搜索输入框不再吃到这部分输入；吞字符标志在下一按键事件先行作废、不滞留误吞；菜单打开时引擎不喂键——模态菜单持有键盘，搜索框打字不再误触动作）；删除连线出厂即恢复老组合 Tab+左键（W 退役）——组合的键盘半边由混合域缓冲推进（Tab 按键触发框选的同时为组合入混合缓冲，事件被引擎消费），与框选键 Tab 的纯键序分域并存、互不清理；平移 = 按住抓图组合键：绑定单键后按住它移动鼠标即拖图、成为唯一触发（绑键后左键拖动停用；出厂未绑默认左键拖动；panByKey 与 mouseMoved 物理键态轮询保证松键即停、切屏/失焦不卡死）；纯鼠标与组合可共享同一按钮（出厂 PAN 左键拖动 + 删线 Tab+左键并存），同钮冲突经前缀歧义校验——等长全同对与修饰 ⊆ 前缀对必拒，修饰可区分的不等长对可共存（如 Ctrl+左键与左键→右键各自只在己方修饰态命中）；旧配置迁移——旧版保存从不写标记，配置新增 editorKeys.format=2 标记用于新表；加载时无标记不再整表回出厂：seq 键序值原样保留（历史脏数据——早期构建未初始化的鼠标槽默认 0=左键并写盘、设置里一堆「/ 左键」的根源——只在 .mouse 鼠标槽字段，读取时忽略、保存时移除并写回标记）；冲突校验：键序前缀歧义拒绝、PAN 抓图键对所有序列的键步独占 / 像素编辑器操作指南的键位文案改为动态读取当前绑定（模板 + EditorKeys.seqText，重绑后指南同步；模板必须走 I18n.get(key, args...) 带参重载——无参重载对 %s 模板先零参 format，整行渲染成 "Format error: …"），/ Conflict rules cover every new action and the settings tab applies them live; the bindings upgrade to the unified sequence model (interleaved key+mouse) - each action holds ONE trigger sequence whose steps may be KEY presses or MOUSE clicks; the settings record arbitrary interleaved queues: keycap clicks append key steps, a mouse-chip click adds the button and a second click removes it (the preview line shows the whole queue live, "Tab → left-click → A → right-click"), and Bind commits it wholesale; the engine advances two independent domains - key steps via keyPressed, mouse steps via clicks, fired by whichever event completes them (a bare right-click = a single mouse step, Ctrl+Z = a pure key sequence, Tab → left-click = the factory wire chord coexisting cross-domain with the box-select Tab sequence); the context menu regains its key trigger (opens at the cursor on the bound key, swallowing the opening key's character so the search box never eats that input - the swallow flag is voided on the next key event so it never lingers; the engine is not fed while the menu is open - the modal menu owns the keyboard, so typing in the search box can no longer mis-fire actions); delete-wire ships as the legacy Tab+left-click chord again (W retired) - a combo's keyboard half advances through the mixed-domain buffer (the Tab press fires box-select while arming the chord, the key event consumed by the engine), coexisting with the box-select Tab sequence across independent domains; pan = the hold-grab combo: bind one key, hold it and move the mouse to drag the canvas as the only trigger (button-drag panning disabled while bound; unbound by factory with left-drag as the default; panByKey plus a physical-key poll in mouseMoved guarantee release-ends-it and no stuck pan across screen switches or focus loss); a plain and a combo may share one button (factory: PAN left-drag + the wire Tab+left chord coexist), same-button clashes go through the prefix-ambiguity check - equal-length all-equal pairs and mods-contained prefixes are refused, modifier-distinguishable pairs coexist (Ctrl+left-click vs left-click → right-click each fire only in their own modifier state); legacy config migration - the old save() never wrote a marker; configs now carry an editorKeys.format=2 marker, and a marker-less table no longer resets wholesale: its seq values carry over verbatim, while the historical dirty data (early builds' uninitialized mouse slots defaulting to 0 = left button leaking into saves - the source of the screenshot's pile of "/ left-click" rows) lived only in the .mouse slot fields, which are ignored on load and removed on save with the marker written back; conflict checks: key-prefix ambiguity refused, PAN's grab key exclusive against the key steps of every sequence; the pixel editor's guide popup now reads its key text live from the current bindings (templates + EditorKeys.seqText, staying in sync after rebinds; the templates go through the parameterized I18n.get(key, args...) - the no-arg overload zero-arg-formats a %s template first and renders "Format error: ..."); The binding system gains 11 bindable actions: ① Delete Wire (default W, press on a hovered wire; the Tab+click chord stays); ② the pixel editor's Brush(B)/Eraser(E)/Fill(F)/Eyedropper(I)/Line(L)/Rect(R)/Hand(H)/Grid(G)/brush Smaller([)/Bigger()] all move to the shared binding system with factory keys identical to the old hardcodes (one exception: [ / ] default to the true GLFW codes 91/93 - the old hardcode's 219/221 were Windows VK codes that GLFW's keyPressed never delivers, so those shortcuts become functional for the first time); the 1..7 rail-order picks stay fixed. Conflict rules cover every new action and the settings tab applies them live.

Change / 变更	Description / 说明
🎛️ 键位系统扩容：删除连线 + 像素编辑器 (行为变更)	键位设置系统新增 11 个可绑定动作：① 删除连线（默认 W，悬停连线按键即删；原 Tab+左键组合保留）；② 像素编辑器的画笔(B)/橡皮(E)/填充(F)/取色(I)/直线(L)/矩形(R)/抓手(H)/网格(G)/笔刷变小([)/变大(]) 全部改走共享键位系统，出厂键与旧硬编码逐项一致（唯一例外：[ / ] 的默认键取 GLFW 真值 91/93——旧硬编码的 219/221 是 Windows VK 码，经 GLFW keyPressed 永不命中，即 [ / ] 笔刷快捷键借此首次真正可用）；1..7 按轨道顺序选工具保持固定。冲突规则（等长全同必拒等）覆盖全部新动作，设置界面即时生效；键位升级为统一序列模型（键鼠交错）——每个动作只有一条触发序列，步骤既可以是按键也可以是鼠标点击，设置里直接录入任意交错队列：点键帽追加键步、点鼠标键把按钮加入队列、再次点击移出（预览行实时显示整条队列，如「Tab → 左键 → A → 右键」），「确定绑定」整条提交；引擎分纯键域/混合域独立推进——键步 keyPressed 推进、鼠标步点击推进，任一事件完成即触发，纯右键=单鼠标步、Ctrl+Z=纯键步、Tab → 左键=出厂删线组合，与框选 Tab 键序跨域并存；上下文菜单恢复键序触发（按绑定键在光标处打开，且吞掉开菜单键的字符——搜索输入框不再吃到这部分输入；吞字符标志在下一按键事件先行作废、不滞留误吞；菜单打开时引擎不喂键——模态菜单持有键盘，搜索框打字不再误触动作）；删除连线出厂即恢复老组合 Tab+左键（W 退役）——组合的键盘半边由混合域缓冲推进（Tab 按键触发框选的同时为组合入混合缓冲，事件被引擎消费），与框选键 Tab 的纯键序分域并存、互不清理；平移 = 按住抓图组合键：绑定单键后按住它移动鼠标即拖图、成为唯一触发（绑键后左键拖动停用；出厂未绑默认左键拖动；panByKey 与 mouseMoved 物理键态轮询保证松键即停、切屏/失焦不卡死）；纯鼠标与组合可共享同一按钮（出厂 PAN 左键拖动 + 删线 Tab+左键并存），同钮冲突经前缀歧义校验——等长全同对与修饰 ⊆ 前缀对必拒，修饰可区分的不等长对可共存（如 Ctrl+左键与左键→右键各自只在己方修饰态命中）；旧配置迁移——旧版保存从不写标记，配置新增 `editorKeys.format=2` 标记用于新表；加载时无标记不再整表回出厂：seq 键序值原样保留（历史脏数据——早期构建未初始化的鼠标槽默认 0=左键并写盘、设置里一堆「/ 左键」的根源——只在 .mouse 鼠标槽字段，读取时忽略、保存时移除并写回标记）；冲突校验：键序前缀歧义拒绝、PAN 抓图键对所有序列的键步独占 / 像素编辑器操作指南的键位文案改为动态读取当前绑定（模板 + `EditorKeys.seqText`，重绑后指南同步；模板必须走 `I18n.get(key, args...)` 带参重载——无参重载对 %s 模板先零参 format，整行渲染成 "Format error: …"），/ Conflict rules cover every new action and the settings tab applies them live; the bindings upgrade to the unified sequence model (interleaved key+mouse) - each action holds ONE trigger sequence whose steps may be KEY presses or MOUSE clicks; the settings record arbitrary interleaved queues: keycap clicks append key steps, a mouse-chip click adds the button and a second click removes it (the preview line shows the whole queue live, "Tab → left-click → A → right-click"), and Bind commits it wholesale; the engine advances two independent domains - key steps via keyPressed, mouse steps via clicks, fired by whichever event completes them (a bare right-click = a single mouse step, Ctrl+Z = a pure key sequence, Tab → left-click = the factory wire chord coexisting cross-domain with the box-select Tab sequence); the context menu regains its key trigger (opens at the cursor on the bound key, swallowing the opening key's character so the search box never eats that input - the swallow flag is voided on the next key event so it never lingers; the engine is not fed while the menu is open - the modal menu owns the keyboard, so typing in the search box can no longer mis-fire actions); delete-wire ships as the legacy Tab+left-click chord again (W retired) - a combo's keyboard half advances through the mixed-domain buffer (the Tab press fires box-select while arming the chord, the key event consumed by the engine), coexisting with the box-select Tab sequence across independent domains; pan = the hold-grab combo: bind one key, hold it and move the mouse to drag the canvas as the only trigger (button-drag panning disabled while bound; unbound by factory with left-drag as the default; panByKey plus a physical-key poll in mouseMoved guarantee release-ends-it and no stuck pan across screen switches or focus loss); a plain and a combo may share one button (factory: PAN left-drag + the wire Tab+left chord coexist), same-button clashes go through the prefix-ambiguity check - equal-length all-equal pairs and mods-contained prefixes are refused, modifier-distinguishable pairs coexist (Ctrl+left-click vs left-click → right-click each fire only in their own modifier state); legacy config migration - the old save() never wrote a marker; configs now carry an `editorKeys.format=2` marker, and a marker-less table no longer resets wholesale: its seq values carry over verbatim, while the historical dirty data (early builds' uninitialized mouse slots defaulting to 0 = left button leaking into saves - the source of the screenshot's pile of "/ left-click" rows) lived only in the `.mouse` slot fields, which are ignored on load and removed on save with the marker written back; conflict checks: key-prefix ambiguity refused, PAN's grab key exclusive against the key steps of every sequence; the pixel editor's guide popup now reads its key text live from the current bindings (templates + `EditorKeys.seqText`, staying in sync after rebinds; the templates go through the parameterized `I18n.get(key, args...)` - the no-arg overload zero-arg-formats a %s template first and renders "Format error: ..."); The binding system gains 11 bindable actions: ① Delete Wire (default W, press on a hovered wire; the Tab+click chord stays); ② the pixel editor's Brush(B)/Eraser(E)/Fill(F)/Eyedropper(I)/Line(L)/Rect(R)/Hand(H)/Grid(G)/brush Smaller([)/Bigger()] all move to the shared binding system with factory keys identical to the old hardcodes (one exception: [ / ] default to the true GLFW codes 91/93 - the old hardcode's 219/221 were Windows VK codes that GLFW's keyPressed never delivers, so those shortcuts become functional for the first time); the 1..7 rail-order picks stay fixed. Conflict rules cover every new action and the settings tab applies them live.

🎛️ 变速器代理态显示 / Transmission Proxy State

Change / 变更 Description / 说明

🎛️ 变速器代理态显示 (行为变更) 可编程变速器此前无论目标来自节点图还是滚轮值盒，值盒都显示滚轮值——节点控制时玩家盯着的是一个不生效的数字，而滚轮照改（改完立刻被图覆盖，谁说了算看不出来）。现引入代理态：图运行且图中存在 TX_OUT（存储序首个；该规则收敛到 NodeGraph.firstNodeIdOfType，代理态判定与目标取值同处发生，避免两条手写遍历漂移成「盒里写(代理)、实际听滚轮」的死守卫）。此时 ① 值盒显示「已应用目标 (代理)」——取 CscTxApplied（已同步的生效值），不取滚轮值、也不新增"节点意图值"字段；两种态都取绝对值（正负号是"以输出面为参考"的另一套语义，不混进这个格子）；② 滚轮/右键拒收输入，走 Create 自带的双边闸门 acceptsValueSettings（客户端输入 ValueSettingsInputHandler:82、服务端收包 ValueSettingsPacket:60、剪贴板读写 ValueSettingsBehaviour:57-69 三处真检查——第三处顺带堵住"用剪贴板把滚轮值灌回来"的旁路），setValueSettings 另做防御性早退；③ 悬停提示第一行改说「目标转速 · 由节点控制」（Create 硬编码的第二行「按住以编辑」删不掉，故由第一行纠正）。代理态由服务端判定并随方块实体包同步（CscTxProxy，仅当代理态或已应用值真的变化时发包，受 4 tick 拆建冷却限频，与官方 setValue → sendData 同量级），客户端只消费、不按本地图推导（本地图会被未 ACK 的编辑带偏）。图停止或图中无 TX_OUT 时完全回到手动态：盒内裸数字、滚轮生效。文案走 lang key 双语（zh (代理) / en (proxy)）。代理态下盒内数字受 4 tick 冷却与包延迟影响，可能短暂滞后于节点当前输出；闸门同样滞后 —— 手动态→代理态翻转后、包到达前，客户端仍按旧值放行片刻，玩家可能看到一瞬「可编辑」的假象（值不会落地：服务端按自己的权威值再查一次闸门，客户端先行的输入被拒）。 / The transmission's value box showed the wheel value no matter who owned the target — while the graph drove it the player stared at a number that did nothing and the wheel stayed editable (the graph overwrote it immediately, so ownership was invisible). New proxied state: while the graph runs with a TX_OUT node (first in storage order; the rule now lives in NodeGraph.firstNodeIdOfType, and the proxy flag and the target value are decided at the same site, so two hand-rolled scans cannot drift into a "box says proxied, wheel still wins" dead guard) — (1) the box reads "applied target (proxy)" from CscTxApplied (the synced effective value), never the inert wheel value and with no extra "node intent" field; both modes take the magnitude, since the sign is output-face-relative and belongs to another concept; (2) the wheel / right-click refuses input through Create's own two-sided gate acceptsValueSettings (really checked in three places: client input, the server packet handler and clipboard read/write — the third also closes the "paste the old wheel value back" bypass), with a defensive early return in setValueSettings; (3) the hover tip's first line becomes "Target speed · Node-controlled" (Create's hardcoded "hold to edit" line cannot be removed, so the first line corrects it). The state is server-decided and shipped in the block-entity packet (CscTxProxy, sent only when the flag or the applied value actually changes, cooldown-limited by the 4-tick rebuild — the same order as the official setValue -> sendData), and the client consumes it instead of deriving it from its local graph, which local un-ACKed edits skew. With the graph stopped, or no TX_OUT at all, everything returns to manual mode: bare magnitude in the box, live wheel. Wording rides bilingual lang keys (zh (代理) / en (proxy)). While proxied the number can briefly lag the node's current output because of the 4-tick cooldown and packet latency. The gate lags the same way: between the manual→proxied flip and the packet's arrival the client still lets the interaction through, so the box can look briefly editable — no value lands, because the server re-checks the gate against its own authoritative flag and rejects the client's optimistic input.

Change / 变更	Description / 说明
🎛️ 变速器代理态显示 (行为变更)	可编程变速器此前无论目标来自节点图还是滚轮值盒，值盒都显示滚轮值——节点控制时玩家盯着的是一个不生效的数字，而滚轮照改（改完立刻被图覆盖，谁说了算看不出来）。现引入代理态：图运行且图中存在 `TX_OUT`（存储序首个；该规则收敛到 `NodeGraph.firstNodeIdOfType`，代理态判定与目标取值同处发生，避免两条手写遍历漂移成「盒里写(代理)、实际听滚轮」的死守卫）。此时 ① 值盒显示「已应用目标 (代理)」——取 `CscTxApplied`（已同步的生效值），不取滚轮值、也不新增"节点意图值"字段；两种态都取绝对值（正负号是"以输出面为参考"的另一套语义，不混进这个格子）；② 滚轮/右键拒收输入，走 Create 自带的双边闸门 `acceptsValueSettings`（客户端输入 `ValueSettingsInputHandler:82`、服务端收包 `ValueSettingsPacket:60`、剪贴板读写 `ValueSettingsBehaviour:57-69` 三处真检查——第三处顺带堵住"用剪贴板把滚轮值灌回来"的旁路），`setValueSettings` 另做防御性早退；③ 悬停提示第一行改说「目标转速 · 由节点控制」（Create 硬编码的第二行「按住以编辑」删不掉，故由第一行纠正）。代理态由服务端判定并随方块实体包同步（`CscTxProxy`，仅当代理态或已应用值真的变化时发包，受 4 tick 拆建冷却限频，与官方 `setValue → sendData` 同量级），客户端只消费、不按本地图推导（本地图会被未 ACK 的编辑带偏）。图停止或图中无 `TX_OUT` 时完全回到手动态：盒内裸数字、滚轮生效。文案走 lang key 双语（zh `(代理)` / en `(proxy)`）。代理态下盒内数字受 4 tick 冷却与包延迟影响，可能短暂滞后于节点当前输出；闸门同样滞后 —— 手动态→代理态翻转后、包到达前，客户端仍按旧值放行片刻，玩家可能看到一瞬「可编辑」的假象（值不会落地：服务端按自己的权威值再查一次闸门，客户端先行的输入被拒）。 / The transmission's value box showed the wheel value no matter who owned the target — while the graph drove it the player stared at a number that did nothing and the wheel stayed editable (the graph overwrote it immediately, so ownership was invisible). New proxied state: while the graph runs with a `TX_OUT` node (first in storage order; the rule now lives in `NodeGraph.firstNodeIdOfType`, and the proxy flag and the target value are decided at the same site, so two hand-rolled scans cannot drift into a "box says proxied, wheel still wins" dead guard) — (1) the box reads "applied target (proxy)" from `CscTxApplied` (the synced effective value), never the inert wheel value and with no extra "node intent" field; both modes take the magnitude, since the sign is output-face-relative and belongs to another concept; (2) the wheel / right-click refuses input through Create's own two-sided gate `acceptsValueSettings` (really checked in three places: client input, the server packet handler and clipboard read/write — the third also closes the "paste the old wheel value back" bypass), with a defensive early return in `setValueSettings`; (3) the hover tip's first line becomes "Target speed · Node-controlled" (Create's hardcoded "hold to edit" line cannot be removed, so the first line corrects it). The state is server-decided and shipped in the block-entity packet (`CscTxProxy`, sent only when the flag or the applied value actually changes, cooldown-limited by the 4-tick rebuild — the same order as the official `setValue -> sendData`), and the client consumes it instead of deriving it from its local graph, which local un-ACKed edits skew. With the graph stopped, or no `TX_OUT` at all, everything returns to manual mode: bare magnitude in the box, live wheel. Wording rides bilingual lang keys (zh `(代理)` / en `(proxy)`). While proxied the number can briefly lag the node's current output because of the 4-tick cooldown and packet latency. The gate lags the same way: between the manual→proxied flip and the packet's arrival the client still lets the interaction through, so the box can look briefly editable — no value lands, because the server re-checks the gate against its own authoritative flag and rejects the client's optimistic input.

🧪 测试 / Tests

Change / 变更 Description / 说明

🧪 测试 CompileResetAndBusNameGateTest 钉死 dirty 门闩（含工厂装入/保留/响应器执行路径）与编译复位（含子图递归）+ fieldIndexOf 反查 + ff3 大数不饱和 + 草稿文本随 op；EncapSubGraphExpandSyncTest 钉死展开状态跨作用域数据同步、「进入子图 → init 恢复」全链路（renderBg 恢复块抽为 restoreOrRebuildEditStates 作测试缝隙）+ encapOccupied 占用判定（节点编辑模式才算、显示模式/异作用域/非法 id 不算）/ Pins the dirty gate (including the factory load/preserve/responder path), compile reset (with sub-graph recursion) and field reverse-lookup, ff3 large-magnitude safety and draft text on ops; EncapSubGraphExpandSyncTest pins cross-scope expansion data sync, the full enter-sub-graph → restore chain (the renderBg restore block extracted as restoreOrRebuildEditStates) and the encapOccupied policy (node-editor mode only; display mode / other scopes / invalid ids never count)；EditorKeysPixelAndWireTest / EditorKeysSequenceTest 钉死出厂触发（含 [ / ] 的 GLFW 91/93）、键鼠交错队列、能力/同域冲突、PAN 抓图/拖动切换与旧配置迁移（legacy seq 原样保留、.mouse 槽字段弃用并写回标记）; EditorKeysPixelAndWireTest / EditorKeysSequenceTest pin the factory triggers (incl. GLFW 91/93 for [ / ]), the interleaved key+mouse queues, the capability/same-domain clashes, the PAN grab/drag switch and the legacy config migration (seq values carry over verbatim; .mouse slots dropped and the marker written back). TransmissionScrollDisplayTest 钉死值盒文本形态（手动态 = 裸绝对值；代理态 = 绝对值 + 空格 + 后缀，含空后缀、Integer 极值与「符号不外泄」），GearboxNodesEvalTest 增补 firstNodeIdOfType（存储序首个 TX_OUT、删掉首个由次个顶上、全无则 -1 = 回落手动）/ TransmissionScrollDisplayTest pins the box-text shape (bare magnitude when manual; magnitude + space + suffix when proxied, incl. the empty suffix, the Integer extremes and the no-sign-leak rule) and GearboxNodesEvalTest now covers firstNodeIdOfType (first TX_OUT in storage order, the next one taking over once the first is removed, -1 when none remain = manual fallback). 游戏内单客户端实测通过（2026-09-27）：图 running 且 TX_OUT 在控（节点要 -10、滚轮值停在 256）时，值盒显示 10 (代理)、悬停提示第一行「目标转速 · 由节点控制」、滚轮与右键均无反应；同刻存档 NBT 直读为 CscTxProxy=1 / CscTxApplied=-10 / CscTxScroll=256（region 文件逐块解压核对）。未覆盖：停图回落手动态、图 running 但图中无 TX_OUT、TX_OUT 存在但未接线（按实现应为 0 (代理)）。 / Verified in-game on a single client (2026-09-27): with the graph running and a TX_OUT in control (node asking -10 while the wheel sat at 256) the box read 10 (proxy), the hover tip's first line read "Target speed · Node-controlled", and both the wheel and right-click did nothing; the save's own NBT at that moment held CscTxProxy=1 / CscTxApplied=-10 / CscTxScroll=256 (read straight out of the decompressed region chunks). Not covered yet: stopping the graph (fallback to manual), a running graph with no TX_OUT, and a TX_OUT with nothing wired to it (expected 0 (proxy)).

Change / 变更	Description / 说明
🧪 测试	`CompileResetAndBusNameGateTest` 钉死 dirty 门闩（含工厂装入/保留/响应器执行路径）与编译复位（含子图递归）+ `fieldIndexOf` 反查 + `ff3` 大数不饱和 + 草稿文本随 op；`EncapSubGraphExpandSyncTest` 钉死展开状态跨作用域数据同步、「进入子图 → init 恢复」全链路（renderBg 恢复块抽为 `restoreOrRebuildEditStates` 作测试缝隙）+ `encapOccupied` 占用判定（节点编辑模式才算、显示模式/异作用域/非法 id 不算）/ Pins the dirty gate (including the factory load/preserve/responder path), compile reset (with sub-graph recursion) and field reverse-lookup, `ff3` large-magnitude safety and draft text on ops; `EncapSubGraphExpandSyncTest` pins cross-scope expansion data sync, the full enter-sub-graph → restore chain (the renderBg restore block extracted as `restoreOrRebuildEditStates`) and the `encapOccupied` policy (node-editor mode only; display mode / other scopes / invalid ids never count)；`EditorKeysPixelAndWireTest` / `EditorKeysSequenceTest` 钉死出厂触发（含 [ / ] 的 GLFW 91/93）、键鼠交错队列、能力/同域冲突、PAN 抓图/拖动切换与旧配置迁移（legacy seq 原样保留、.mouse 槽字段弃用并写回标记）; `EditorKeysPixelAndWireTest` / `EditorKeysSequenceTest` pin the factory triggers (incl. GLFW 91/93 for [ / ]), the interleaved key+mouse queues, the capability/same-domain clashes, the PAN grab/drag switch and the legacy config migration (seq values carry over verbatim; .mouse slots dropped and the marker written back). `TransmissionScrollDisplayTest` 钉死值盒文本形态（手动态 = 裸绝对值；代理态 = 绝对值 + 空格 + 后缀，含空后缀、Integer 极值与「符号不外泄」），`GearboxNodesEvalTest` 增补 `firstNodeIdOfType`（存储序首个 `TX_OUT`、删掉首个由次个顶上、全无则 -1 = 回落手动）/ `TransmissionScrollDisplayTest` pins the box-text shape (bare magnitude when manual; magnitude + space + suffix when proxied, incl. the empty suffix, the Integer extremes and the no-sign-leak rule) and `GearboxNodesEvalTest` now covers `firstNodeIdOfType` (first TX_OUT in storage order, the next one taking over once the first is removed, -1 when none remain = manual fallback). 游戏内单客户端实测通过（2026-09-27）：图 running 且 `TX_OUT` 在控（节点要 -10、滚轮值停在 256）时，值盒显示 `10 (代理)`、悬停提示第一行「目标转速 · 由节点控制」、滚轮与右键均无反应；同刻存档 NBT 直读为 `CscTxProxy=1` / `CscTxApplied=-10` / `CscTxScroll=256`（region 文件逐块解压核对）。未覆盖：停图回落手动态、图 running 但图中无 `TX_OUT`、`TX_OUT` 存在但未接线（按实现应为 `0 (代理)`）。 / Verified in-game on a single client (2026-09-27): with the graph running and a TX_OUT in control (node asking -10 while the wheel sat at 256) the box read `10 (proxy)`, the hover tip's first line read "Target speed · Node-controlled", and both the wheel and right-click did nothing; the save's own NBT at that moment held `CscTxProxy=1` / `CscTxApplied=-10` / `CscTxScroll=256` (read straight out of the decompressed region chunks). Not covered yet: stopping the graph (fallback to manual), a running graph with no TX_OUT, and a TX_OUT with nothing wired to it (expected `0 (proxy)`).

create-schematic-compute

Links

Tags

Creators

Details

Compatibility

Required content

Changes

🔧 动力传感器扳手 / Kinetic Gauge Wrench

🧭 贴地/贴顶放置 / Floor & Ceiling Placement

🖥️ 倒置屏幕读数 / Inverted Mount Readout

🔌 编码器清零改节点体引脚 / Encoder Reset on the Node Body

🎛️ PID / PID_POWER 积分死区 / Integral Deadband

🎛️ 输出指令正/反转与负行程 / Output-Command Forward-Reverse & Negative Travel

🔧 数控齿轮箱扳手回归官方 / CNC Gearbox Wrench Back to Official

🎛️ 变速器过载后输出不恢复 / Transmission Output Stays Dead After Overload

⚙️ 数控齿轮箱输出端轴面常在 / CNC Gearbox Output Shaft Face Always Present

🗂️ 节点分类与方块名单 / Node Categories & Per-Block Allowances

🖱️ 图编辑器命中与裁剪 / Graph Editor Hit & Cull

🖥️ HUD 虚像与视角摇晃 / HUD Virtual Image & View Bobbing

🔤 手动字形与语言切换 / Manual Glyphs & Language Switch

✨ 删节点撤销/重做恢复连线 / Node-Delete Undo-Redo Wire Restore

🛠️ 图编辑器与多人协作 / Graph Editor & Collaboration

⌨️ 键位系统 / Key Bindings

🎛️ 变速器代理态显示 / Transmission Proxy State

🧪 测试 / Tests

Optional dependencies

Developer information

Maven coordinates:

Version ID:

build.gradle: