0.1.4

中文 | English

0.1.4 是一次性能版本：对「打开大图」与「编辑流畅度」两条主路径做系统性提速——所有含文字的节点改由插件自绘渲染，跳过引擎成本为「0.13ms/字符」的逐字符文本测宽（含长文本的大图打开从数十秒降到秒级）；打开过程改为分片渲染，界面全程不冻结；另以五处引擎自有补丁收敛编辑与渲染帧成本（一次编辑 47.6 → 5.3ms、空渲染落地 28.1 → 3.0ms）。同时向内收敛两项工程基建：引擎补丁体系（vendored 源入仓 + 可复现重打包）与社区目录校验对照。无破坏性变更（.mindmap.md 格式、命令 ID 均未变），新增一个默认开启的设置项。

性能

• 所有含文字的节点改用自绘渲染（新设置「文本节点快速渲染」，默认开启）：打开大图的成本主体是引擎的逐字符文本测宽（每字符一次离屏测量，0.13ms/字符），字符越多越慢——含长文本段落的大图尤其显著。现改为全部含文字节点由插件自绘（跳过测宽，尺寸走内容寻址缓存），实测 5000 节点：含长文本节点约 59 秒 → 约 1.4 秒，纯短文本 8.6 秒 → 约 1.4 秒。边界：空文本 / 含图 / 整行注释等节点不接管；双击节点改为打开编辑弹窗（引擎内联编辑框对自绘节点不可用）；该设置仅对之后打开的文件生效（切换时不重建当前大图，避免卡顿）；关闭即恢复引擎原生文本渲染与原位编辑
• 打开大图时界面不再冻结：整树渲染改为分片执行（每个节点一个宏任务，浏览器可在其间响应输入与绘制），主线程单块阻塞从约 1.1 秒降到每块约 0.2 毫秒——打开全程可交互；总时长代价经同批对照仅约 +2.4%（5000 节点 1137 → 1164ms）
• 打开成本的三处继续收敛（5000 节点实测）：① 拖宽手柄的全局监听器由「每个节点 3 个」改为拖拽会话内惰性注册（15000 → 0，打开 -13%；同时修复「节点销毁后监听器永久持有」的内存泄漏）；② 布局的兄弟子树平移消除二次复杂度（实测 293 万次节点位移 → O(n)，打开 -6%）；③ 自绘尺寸按内容寻址缓存 + 首帧前预测量（一遍内容构建、重复打开与首次打开都免重复测宽）
• 编辑与渲染帧成本（500 节点实测）：一次编辑 47.6 → 5.3ms（-89%）、空渲染落地 28.1 → 3.0ms（-89%）——三处引擎修复：兄弟节点查找 O(n) → O(1)（原先每节点每轮把全部兄弟拷进新数组线性查找）、数据快照比较去掉 JSON 序列化、SVG 属性同值写短路（不再触发无谓的样式重算与重绘）

工程

• 引擎自有补丁体系：fix.3 引擎源码入仓（vendor/upstream/）＋ 9 处自有补丁（vendor/patches/ 与 upstream 内），经 npm run build:vendor 可复现重打包，产物 sha256 由契约测试钉住——升级与审计链路完全可查（清单见 vendor/BUILD.md）
• 测试：全量 49 文件 / 1578 例；npm run verify:visual 现有 26+ 断言探针，并新增三组 opt-in 测量通道（打开基准 --bench-open、编辑帧分解 --perf --perf-ops、布局计数 --bench-open + BENCH_VARIANT=count-layout）——本版全部性能数据均由这些通道的真实墙钟差分测得
• 社区目录校验对照：新增 npm run lint:scanner（用社区目录 scanner 的固定版本依赖集与官方规则面本地预检），接入 CI 与发布门禁；scanner 工作流、官方模板、开发者指南三条对照线均 0 缺口

设置与交互变化

• 新增设置「文本节点快速渲染」（默认开启，可关闭）——见上方性能段；变更仅对之后打开的文件生效
• 自绘节点的文本编辑统一走编辑弹窗（双击 / 右键「编辑文本」/ F2），宽度拖拽手柄只在自绘节点上显示

兼容性说明

• 与 0.1.3 一致：需要 Obsidian 1.13.0+，仅桌面端（isDesktopOnly: true）
• .mindmap.md 格式、既有命令 ID、快捷键与其余设置项均未变更；本版为纯性能 / 稳定性改动，无需任何迁移

---

Performance

• All text-bearing nodes are now rendered by the plugin (new setting "Fast text node rendering", on by default): the dominant cost of opening a large map was the engine's per-character text measurement (one off-screen measurement per character, 0.13 ms/char) — the more text, the slower the open, with long paragraphs being the worst case. Every text-bearing node is now drawn by the plugin (measurement skipped; sizes come from a content-addressed cache). Measured with 5,000 nodes: a map with long-text nodes went from ~59 s to ~1.4 s; short-text-only from 8.6 s to ~1.4 s. Boundaries: empty-text / image-bearing / fully-commented nodes are not taken over; double-clicking a node now opens the edit dialog (the engine's inline editor does not work on self-drawn nodes); the setting applies to maps opened afterwards (switching it does not rebuild the current large map, avoiding a stutter); turning it off restores the engine's native text rendering and inline editing
• The UI no longer freezes while a large map opens: full-tree rendering is now sliced (one macrotask per node, so the browser can respond to input and paint in between) — the worst single main-thread block drops from ~1.1 s to ~0.2 ms per slice, so the app stays responsive throughout the open; the total-time cost measured in a same-batch comparison is only about +2.4% (5,000 nodes: 1137 → 1164 ms)
• Three further reductions in open cost (measured with 5,000 nodes): ① the width-handle global listeners went from "3 per node" to lazy registration inside a drag session (15,000 → 0; -13% open time; also fixes a memory leak where destroyed nodes were kept alive by window listeners); ② the layout's sibling-subtree shifting lost its quadratic complexity (2.93M node moves → O(n); -6% open time); ③ self-drawn sizes are cached by content and pre-measured before the first frame (one content build only; no repeated measurement on first or repeat opens)
• Edit and render frame cost (measured with 500 nodes): one edit 47.6 → 5.3 ms (-89%), an empty render pass 28.1 → 3.0 ms (-89%) — three engine fixes: sibling lookup O(n) → O(1) (previously copying all siblings into a new array for a linear scan on every node every pass), data-snapshot comparison without JSON serialization, and same-value SVG attribute writes are short-circuited (no more pointless style recalculation and repaint)

Engineering

• A first-party engine patch system: the fix.3 engine source is vendored (vendor/upstream/) with 9 in-house patches (vendor/patches/ plus in-tree ones), rebuilt reproducibly via npm run build:vendor, with the artifact sha256 pinned by a contract test — upgrade and audit trails are fully inspectable (see vendor/BUILD.md)
• Tests: the full suite is 49 files / 1578 cases; npm run verify:visual now has 26+ assertion probes, plus three opt-in measurement channels (open benchmark --bench-open, edit-frame breakdown --perf --perf-ops, layout counters via BENCH_VARIANT=count-layout) — every performance figure in this release was measured through these real-wall-clock channels
• Community-catalog parity checks: new npm run lint:scanner runs the community catalog scanner's pinned dependency set and rule set locally, wired into CI and the release gate; the scanner-workflow, sample-plugin and developer-guideline audits all found zero gaps

Settings and interaction changes

• New setting "Fast text node rendering" (on by default, can be turned off) — see Performance above; the change applies to maps opened afterwards
• Text editing on self-drawn nodes goes through the edit dialog (double-click / context menu "Edit text" / F2); the width-drag handles appear only on self-drawn nodes

Compatibility notes

• Unchanged from 0.1.3: requires Obsidian 1.13.0+, desktop only (isDesktopOnly: true)
• No changes to the .mindmap.md format, existing command IDs, hotkeys or other settings; this release is performance / stability-only and needs no migration