@liustack/pptfast 0.6.0 → 0.8.0

package/README.md

@@ -58,6 +58,28 @@ installNodePlatform()
 const bytes = await generatePptx(ir) // Uint8Array, ready to write to a .pptx
 ```
 
+## Browser
+
+The default `@liustack/pptfast` entry is browser-safe by construction — its whole dependency closure has no `fs`, no `commander`, no Node-only API (`docs/architecture.md`'s platform seam). Which subpath to import depends on how your page loads code:
+
+- **Bundler** (Vite, webpack, esbuild, Next.js, …) — the default entry, unchanged: `import { generatePptx, validateIr } from "@liustack/pptfast"`. `react`/`react-dom`/`zod`/`jszip`/`dagre`/`pptxgenjs` stay external, resolved from your own `node_modules` and deduped against whatever else you already depend on.
+- **Bare `<script type="module">`, no build step** — `@liustack/pptfast/browser` instead. Every dependency is bundled in (react + react-dom/server + zod + jszip + dagre + pptxgenjs, ~1.7 MB raw / ~455 KB gzip), so it loads with nothing else on the page and no import map:
+
+  ```html
+  <script type="module">
+    import { validateIr, generatePptx } from "https://esm.sh/@liustack/pptfast/browser"
+    const { ok, ir, errors } = validateIr(deckJson)
+    if (!ok) throw new Error(errors.map((e) => e.message).join("\n"))
+    const bytes = await generatePptx(ir) // Uint8Array — new Blob([bytes]) to download
+  </script>
+  ```
+
+- **Embedding just the validator** — a page that checks pasted/edited IR JSON and shows errors has no reason to carry the render/export chain at all. `@liustack/pptfast/validate` exports `validateIr`, `formatIssues`/`formatWarnings`, `irJsonSchema`/`styleJsonSchema`, `listThemes`, and the IR/style zod schemas — `react`, `react-dom/server`, `pptxgenjs`, `jszip`, and `dagre` are all absent from its closure by construction, not just unused at runtime, landing at a fraction of `/browser`'s size (~730 KB raw / ~155 KB gzip).
+
+No `installPlatform()` call is needed in a browser — DOM parsing and (for `--pixels`-equivalent audits) SVG rasterization both default to the real browser globals (`DOMParser`, `OffscreenCanvas`) automatically, unlike `installNodePlatform()` (Quick start above), which is a Node-only requirement. Two honest caveats, not bugs: an image asset needs to already be a `data:` URI, or an `http(s)` URL the browser's own `fetch` can read cross-origin (the host needs permissive CORS headers — reading raw bytes to inline needs more than an `<img>` tag ever does) — `generatePptx` fetches such a URL for you at export time, but `auditDeck`'s pixel path (`{ pixels: true }`) refuses a remote `http(s)` image reference outright, the same "never a surprise network request" rule the Node/Sharp path also follows. `--pixels`-equivalent auditing needs `OffscreenCanvas`, which every evergreen browser has, while Node needs the optional `sharp` dependency instead (see Auditing below).
+
+`pptfast preview --html` (see For AI agents below) is the project's most mature "ships to a browser" artifact today: a self-contained review page a Node CLI run generates once, with zero further dependencies or network calls once it's open in a tab — distinct from importing the SDK live into your own page, but worth knowing about if "zero-dependency browser artifact" is what you actually need.
+
 ## CLI
 
 | Command | Does |
@@ -80,9 +102,9 @@ const bytes = await generatePptx(ir) // Uint8Array, ready to write to a .pptx
 
 Run `node dist/cli.js schema` for the full JSON Schema — feed it to a model before asking it to write IR. A deck (`PptxIR`) has `version` (currently `"4"`, and now the default when omitted), `filename`, an optional `narrative` (a preset id string or a partial axes object — see Narratives below), `theme` (`id` plus optional `style`/`brand` overrides), `meta`, and `assets` — all optional with sane defaults — plus a separate optional `brand` (logo placement) and a required ordered list of `slides`. Each slide has a `type` (`cover`, `chapter`, `content`, `ending`), an optional `layout` (an explicit page-layout id that always wins over auto-selection — omit it and pptfast auto-selects one, see Layout selection below), an optional `arrangement` (how a content slide's body is laid out, e.g. `two_column`, `kpi_focus`), and a list of typed `components` (`bullets`, `kpi_cards`, `image`, `chart`, …). `assets` is `{ images: { [id]: { src, alt? } } }` — components reference images by `asset_id`, so the same image can be reused across slides without duplication.
 
-A deck also carries an optional `seed` (an integer that keeps auto-selected layouts stable across revisions — see Layout selection below for how it's derived when omitted). Any slide may set a stable `id` (what spec pages and validation error messages reference it by), `placeholder: true` (a slide with no content yet — injected by `assemble` for a spec page nobody has filled in, skipped by the content-quality checks, and blocking `render` unless `--draft`), and an optional `notes` (aliases `note`/`speaker_notes`/`speakerNotes`) that exports as a native PowerPoint speaker note — content for the presenter's own view, never drawn onto the slide canvas and never counted toward any layout capacity. Field names that commonly drift between a model's output and the schema (40 synonym pairs across component types, e.g. kpi `title`→`label`, quote `content`→`text`, swot `strength`→`strengths`, bmc `partners`→`key_partners`) are silently normalized to the canonical name at validate time — `validate`/`render`/`preview` print a note listing what changed, never a hard error. That rescue is scoped to weak-model synonym drift only — it does not cover pre-v4 vocabulary. A v4-labeled document that writes `scenario` instead of `narrative`, `mode`/`delivery` instead of `strategy`/`pacing`, or the old `narrative`/`text`/`presentation` axis values hard-rejects, listing the current names/values, exactly like any other unrecognized field or value. An explicit `version: "3"` (or `"2"`) also hard-rejects, with a migration pointer — see `pptfast migrate` below, the only supported path for old-vocabulary input.
+A deck also carries an optional `seed` (an integer that keeps auto-selected layouts stable across revisions — see Layout selection below for how it's derived when omitted). Any slide may set a stable `id` (what spec pages and validation error messages reference it by), `placeholder: true` (a slide with no content yet — injected by `assemble` for a spec page nobody has filled in, skipped by the content-quality checks, and blocking `render` unless `--draft`), and an optional `notes` (aliases `note`/`speaker_notes`/`speakerNotes`) that exports as a native PowerPoint speaker note — content for the presenter's own view, never drawn onto the slide canvas and never counted toward any layout capacity. Field names that commonly drift between a model's output and the schema (53 synonym pairs across component types, e.g. kpi `title`→`label`, quote `content`→`text`, swot `strength`→`strengths`, bmc `partners`→`key_partners`) are silently normalized to the canonical name at validate time — `validate`/`render`/`preview` print a note listing what changed, never a hard error. That rescue is scoped to weak-model synonym drift only — it does not cover pre-v4 vocabulary. A v4-labeled document that writes `scenario` instead of `narrative`, `mode`/`delivery` instead of `strategy`/`pacing`, or the old `narrative`/`text`/`presentation` axis values hard-rejects, listing the current names/values, exactly like any other unrecognized field or value. An explicit `version: "3"` (or `"2"`) also hard-rejects, with a migration pointer — see `pptfast migrate` below, the only supported path for old-vocabulary input.
 
-Four component types are *full-body*: `swot` (strengths/weaknesses/opportunities/threats), `bmc` (the nine-block Business Model Canvas), `waterfall` (a running-total bridge chart), and `gantt` (dated bars on a shared numeric axis). Each fills a slide's entire content rect and must be the only component on its slide — mixing one in with anything else fails `validate` instead of silently dropping the sibling.
+Eight component types are *full-body*: `swot` (strengths/weaknesses/opportunities/threats), `bmc` (the nine-block Business Model Canvas), `waterfall` (a running-total bridge chart), `gantt` (dated bars on a shared numeric axis), `pest` (a political/economic/social/technological macro-environment scan), `five_forces` (Porter's competitive-forces hub-and-spoke), `heatmap` (a value-driven color grid), and `sankey` (a layered, quantity-proportional flow diagram — shipped as native editable vectors, not the rasterized image the type gets elsewhere). Each fills a slide's entire content rect and must be the only component on its slide — mixing one in with anything else fails `validate` instead of silently dropping the sibling.
 
 The v4 IR schema is frozen as of 0.4.0 — future evolution is additive only (new optional fields, new enum members), and any breaking change ships under a new top-level `version` value with the same hard-reject-and-migration treatment v3 got. `pptfast migrate <v3-file.json> -o <out.json>` deterministically converts a v3 file to v4 (field renames only — same theme, layout selection, content budgets, and visual output) — see Deck projects below for the sibling `deck.plan.json` → `deck.spec.json` conversion.
 
@@ -180,7 +202,7 @@ pptfast audit examples/basic.json
 
 ## For AI agents
 
-The recommended loop for an agent generating a deck: read `pptfast schema` to learn the vocabulary, write an IR JSON, run `pptfast validate` and fix whatever it reports (errors carry a page number and a fixable-in-place message — the point is to close this loop without a human), then `pptfast audit` for the same kind of fixable-in-place feedback on what a *valid* deck can still get wrong at render time (overflow, low-contrast, overlap — exit code alone says whether it's clean), then `pptfast render`. `pptfast preview` gives the agent SVG files it can look at to self-check layout before committing to a render. Add `--html` to also write a self-contained `preview.html` for a human to review (keyboard nav, placeholder badges — a remote-URL image asset stays remote, the one self-containment gap). When every page is filled, that `preview.html` also overlays the same `audit` findings (per-page badges plus a findings panel, click to jump to the page) so a human reviewer sees them without a terminal — a deck with any placeholder page shows a one-line "audit skipped" notice instead. The reviewer can leave free-text per-page annotations right in `preview.html` and export them as a `revision-request.json` (a Blob download, no network or file write — preview stays read-only) for the agent to route back through `pages/*.json`. The Claude Code plugin above wraps this loop as a skill ([`skills/pptfast/SKILL.md`](./skills/pptfast/SKILL.md)). This exact loop is exercised by an internal, model-agnostic benchmark (`tests/bench/`, not published to npm) that mechanically scores how well a model follows the skill on a fixed question bank — see `tests/bench/README.md`.
+The recommended loop for an agent generating a deck: read `pptfast schema` to learn the vocabulary, write an IR JSON, run `pptfast validate` and fix whatever it reports (errors carry a page number and a fixable-in-place message — the point is to close this loop without a human), then `pptfast audit` for the same kind of fixable-in-place feedback on what a *valid* deck can still get wrong at render time (overflow, low-contrast, overlap — exit code alone says whether it's clean), then `pptfast render`. `pptfast preview` gives the agent SVG files it can look at to self-check layout before committing to a render. Add `--html` to also write a self-contained `preview.html` for a human to review (keyboard nav, placeholder badges — a remote-URL image asset stays remote, the one self-containment gap) — zero network calls and zero further dependencies once it's open in a tab, the project's most mature zero-dependency browser artifact today (see Browser above). When every page is filled, that `preview.html` also overlays the same `audit` findings (per-page badges plus a findings panel, click to jump to the page) so a human reviewer sees them without a terminal — a deck with any placeholder page shows a one-line "audit skipped" notice instead. The reviewer can leave free-text per-page annotations right in `preview.html` and export them as a `revision-request.json` (a Blob download, no network or file write — preview stays read-only) for the agent to route back through `pages/*.json`. The Claude Code plugin above wraps this loop as a skill ([`skills/pptfast/SKILL.md`](./skills/pptfast/SKILL.md)). This exact loop is exercised by an internal, model-agnostic benchmark (`tests/bench/`, not published to npm) that mechanically scores how well a model follows the skill on a fixed question bank — see `tests/bench/README.md`.
 
 ## Roadmap
 

package/README.zh-CN.md

@@ -56,6 +56,28 @@ installNodePlatform()
 const bytes = await generatePptx(ir) // Uint8Array，可直接写成 .pptx 文件
 ```
 
+## 浏览器
+
+`@liustack/pptfast` 默认入口天生浏览器安全——它的整个依赖闭包不含 `fs`、不含 `commander`、不含任何 Node-only API（见 `docs/architecture.md` 的 platform seam 一节）。具体该 import 哪个子路径，取决于页面怎么加载代码：
+
+- **打包器场景**（Vite、webpack、esbuild、Next.js 等）——默认入口，写法不变：`import { generatePptx, validateIr } from "@liustack/pptfast"`。`react`/`react-dom`/`zod`/`jszip`/`dagre`/`pptxgenjs` 依旧是 external，从你自己的 `node_modules` 解析，和你项目里其他依赖一起去重。
+- **裸 `<script type="module">`，零构建**——改用 `@liustack/pptfast/browser`。每个依赖都已打包进去（react + react-dom/server + zod + jszip + dagre + pptxgenjs，约 1.7MB 原始 / 约 455KB gzip），页面上不需要任何 import map 或额外脚本就能加载：
+
+  ```html
+  <script type="module">
+    import { validateIr, generatePptx } from "https://esm.sh/@liustack/pptfast/browser"
+    const { ok, ir, errors } = validateIr(deckJson)
+    if (!ok) throw new Error(errors.map((e) => e.message).join("\n"))
+    const bytes = await generatePptx(ir) // Uint8Array —— new Blob([bytes]) 即可触发下载
+  </script>
+  ```
+
+- **只嵌入校验器**——一个校验用户粘贴/编辑的 IR JSON 并展示错误的页面，完全没理由带上整条渲染/导出链。`@liustack/pptfast/validate` 导出 `validateIr`、`formatIssues`/`formatWarnings`、`irJsonSchema`/`styleJsonSchema`、`listThemes`，以及 IR/style 的 zod schema——`react`、`react-dom/server`、`pptxgenjs`、`jszip`、`dagre` 从依赖闭包层面就不存在（不是「运行时用不到」，而是构建时就没打进去），体积只有 `/browser` 的一个零头（约 730KB 原始 / 约 155KB gzip）。
+
+浏览器里不需要调用 `installPlatform()`——DOM 解析，以及（等价于 `--pixels` 的）SVG 光栅化，都会自动 fallback 到真实的浏览器全局对象（`DOMParser`、`OffscreenCanvas`），`installNodePlatform()`（见上文快速开始）只是 Node 场景下的硬性要求。两条如实说明的限制，而非缺陷：图片资产要么已经是 `data:` URI，要么是浏览器自身 `fetch` 能跨域读取的 `http(s)` URL（宿主需要开放的 CORS 头——把原始字节读出来内联，比 `<img>` 标签单纯显示图片的要求更高）——`generatePptx` 会在导出时替你 fetch 这类 URL，但 `auditDeck` 的像素审查路径（`{ pixels: true }`）会直接拒绝还带着远程 `http(s)` 图片引用的页面，和 Node/Sharp 路径共享同一条「绝不擅自发起意外网络请求」的规则。等价于 `--pixels` 的审查需要 `OffscreenCanvas`，现代浏览器都有，Node 一侧则需要可选依赖 `sharp`（见下文审查一节）。
+
+`pptfast preview --html`（见下文「面向 AI agent」）是这个项目目前最成熟的「产物直接跑在浏览器里」的例子：Node CLI 跑一次生成一个自包含的审阅页，打开后零网络请求、零进一步依赖——和「把 SDK 实时 import 进你自己的页面」是两回事，但如果你要的正好是「零依赖浏览器产物」，这个已经是现成答案。
+
 ## CLI
 
 | 命令 | 作用 |
@@ -78,9 +100,9 @@ const bytes = await generatePptx(ir) // Uint8Array，可直接写成 .pptx 文
 
 运行 `node dist/cli.js schema` 获取完整 JSON Schema——让模型写 IR 之前先读它。一份 deck（`PptxIR`）包含 `version`（现为 `"4"`，且省略时默认就是它）、`filename`、一个可选的 `narrative`（预设 id 字符串，或部分轴对象——详见下文「叙事」一节）、`theme`（`id` 加可选的 `style`/`brand` 覆盖）、`meta`、`assets`——均可省略、有默认值——另外还有一个独立的可选 `brand`（logo 位置）字段，以及必填的有序 `slides` 列表。每张 slide 有一个 `type`（`cover`、`chapter`、`content`、`ending`）、一个可选的 `layout`（显式指定页面版式 id，一经设置恒生效、优先于自动选型——省略则由 pptfast 自动选型，详见下文「版式选型」）、一个可选的 `arrangement`（content slide 正文的排布方式，如 `two_column`、`kpi_focus`），以及一组带类型的 `components`（`bullets`、`kpi_cards`、`image`、`chart` 等）。`assets` 的形状是 `{ images: { [id]: { src, alt? } } }`，component 通过 `asset_id` 引用图片，同一张图可以在多页复用而不必重复内嵌。
 
-一份 deck 还可以携带一个可选的 `seed`（整数，让自动选型的版式在多次修订之间保持稳定——省略时如何生成，详见下文「版式选型」）。任意 slide 都可以设置一个稳定的 `id`（spec 的页面和校验报错都靠它引用）、`placeholder: true`（还没有内容的占位 slide——由 `assemble` 为 spec 里没人填写的页面注入，内容质量检查会跳过它，`render` 也会因它拒绝导出，除非加 `--draft`），以及一个可选的 `notes`（同义词 `note`/`speaker_notes`/`speakerNotes`），导出为原生 PowerPoint 演讲者备注——只是给主讲人自己看的内容，不会画到幻灯片画布上，也从不计入任何版式容量。模型输出里容易和 schema 对不上的字段名（跨 component 类型共 40 组同义词，例如 kpi 的 `title`→`label`、quote 的 `content`→`text`、swot 的 `strength`→`strengths`、bmc 的 `partners`→`key_partners`）会在校验时静默改写成规范名——`validate`/`render`/`preview` 会打印一条改了什么的提示，从不因此报错。这套救援机制只覆盖弱模型的同义词漂移，不覆盖 v4 之前的旧词汇。标着 v4 却仍写 `scenario`（而不是 `narrative`）、`mode`/`delivery`（而不是 `strategy`/`pacing`）、或轴值还停留在旧的 `narrative`/`text`/`presentation` 的文档，会像任何其他未识别字段或非法值一样直接硬报错，并列出当前正确的名称和取值。显式写 `version: "3"`（或 `"2"`）同样硬拒绝并给出迁移指引——见下文 `pptfast migrate`，这是旧词汇输入唯一支持的路径。
+一份 deck 还可以携带一个可选的 `seed`（整数，让自动选型的版式在多次修订之间保持稳定——省略时如何生成，详见下文「版式选型」）。任意 slide 都可以设置一个稳定的 `id`（spec 的页面和校验报错都靠它引用）、`placeholder: true`（还没有内容的占位 slide——由 `assemble` 为 spec 里没人填写的页面注入，内容质量检查会跳过它，`render` 也会因它拒绝导出，除非加 `--draft`），以及一个可选的 `notes`（同义词 `note`/`speaker_notes`/`speakerNotes`），导出为原生 PowerPoint 演讲者备注——只是给主讲人自己看的内容，不会画到幻灯片画布上，也从不计入任何版式容量。模型输出里容易和 schema 对不上的字段名（跨 component 类型共 53 组同义词，例如 kpi 的 `title`→`label`、quote 的 `content`→`text`、swot 的 `strength`→`strengths`、bmc 的 `partners`→`key_partners`）会在校验时静默改写成规范名——`validate`/`render`/`preview` 会打印一条改了什么的提示，从不因此报错。这套救援机制只覆盖弱模型的同义词漂移，不覆盖 v4 之前的旧词汇。标着 v4 却仍写 `scenario`（而不是 `narrative`）、`mode`/`delivery`（而不是 `strategy`/`pacing`）、或轴值还停留在旧的 `narrative`/`text`/`presentation` 的文档，会像任何其他未识别字段或非法值一样直接硬报错，并列出当前正确的名称和取值。显式写 `version: "3"`（或 `"2"`）同样硬拒绝并给出迁移指引——见下文 `pptfast migrate`，这是旧词汇输入唯一支持的路径。
 
-四种 component 类型是「满幅」的：`swot`（strengths/weaknesses/opportunities/threats 四象限）、`bmc`（九宫格商业模式画布）、`waterfall`（运行合计瀑布图）、`gantt`（共享数轴上的甘特条形图）。各自独占整张 slide 的内容区域，必须是该 slide 唯一的 component——混入其他 component 会在校验时报错，而不是静默丢弃。
+八种 component 类型是「满幅」的：`swot`（strengths/weaknesses/opportunities/threats 四象限）、`bmc`（九宫格商业模式画布）、`waterfall`（运行合计瀑布图）、`gantt`（共享数轴上的甘特条形图）、`pest`（政治/经济/社会/技术宏观环境扫描）、`five_forces`（波特五力竞争结构轮辐图）、`heatmap`（值驱动色阶网格）、`sankey`（分层且量值成比例的流向图——导出为原生可编辑矢量，而非该图表类型在别处常见的栅格图片）。各自独占整张 slide 的内容区域，必须是该 slide 唯一的 component——混入其他 component 会在校验时报错，而不是静默丢弃。
 
 v4 IR schema 自 0.4.0 起冻结——后续演进只走加法（新增可选字段、新增枚举值），任何破坏性变更都会启用新的顶层 `version` 值，并沿用 v3 那套硬拒绝 + 迁移提示的处理方式。`pptfast migrate <v3-file.json> -o <out.json>` 能确定性地把一份 v3 文件转成 v4（只做字段改名——theme、版式选型、内容预算与视觉输出都不变）——`deck.plan.json` → `deck.spec.json` 的姊妹转换见下文「Deck 项目」。
 
@@ -172,7 +194,7 @@ pptfast audit examples/basic.json
 
 ## 面向 AI agent
 
-推荐给 agent 的生成回路：先读 `pptfast schema` 学词汇表，写出 IR JSON，跑 `pptfast validate` 并根据报错自纠（错误信息带页码和可直接照抄的修正方式，目的就是让这个回路不必依赖人工介入），再跑 `pptfast audit`——同样是可直接照抄的修正反馈，只是针对一份*合法* deck 在渲染层仍可能出现的问题（溢出、低对比度、重叠——exit code 本身就说明干不干净），最后执行 `pptfast render`。`pptfast preview` 能让 agent 在正式渲染前先看一遍 SVG，自查版式是否合理。加上 `--html` 还会额外写出一个自包含的 `preview.html`，供人工审查（键盘翻页、占位页角标——远程 URL 的图片资产仍是远程链接，这是自包含性上唯一的缺口）。当所有页面都已填写时，这份 `preview.html` 还会叠加同一份 `audit` 检查结果（每页一个数量角标，加一个可点击跳转的 findings 面板），让人工审查者不必打开终端就能看到问题——如果 deck 里还有占位页，则显示一行「audit 已跳过」的提示代替。审查者可以直接在 `preview.html` 里给每页写自由文本批注，并导出为 `revision-request.json`（浏览器 Blob 下载，不联网也不写文件——preview 始终只读），交给 agent 通过 `pages/*.json` 回填。上文的 Claude Code 插件已把这套回路封装成 skill（[`skills/pptfast/SKILL.md`](./skills/pptfast/SKILL.md)）。这套回路本身由一个模型无关的内部基准测试（`tests/bench/`，不发布到 npm）机械化验证——固定题库，评估模型跟随该 skill 的表现，细节见 `tests/bench/README.md`。
+推荐给 agent 的生成回路：先读 `pptfast schema` 学词汇表，写出 IR JSON，跑 `pptfast validate` 并根据报错自纠（错误信息带页码和可直接照抄的修正方式，目的就是让这个回路不必依赖人工介入），再跑 `pptfast audit`——同样是可直接照抄的修正反馈，只是针对一份*合法* deck 在渲染层仍可能出现的问题（溢出、低对比度、重叠——exit code 本身就说明干不干净），最后执行 `pptfast render`。`pptfast preview` 能让 agent 在正式渲染前先看一遍 SVG，自查版式是否合理。加上 `--html` 还会额外写出一个自包含的 `preview.html`，供人工审查（键盘翻页、占位页角标——远程 URL 的图片资产仍是远程链接，这是自包含性上唯一的缺口）——打开后零网络请求、零进一步依赖，是这个项目目前最成熟的零依赖浏览器产物（见上文「浏览器」）。当所有页面都已填写时，这份 `preview.html` 还会叠加同一份 `audit` 检查结果（每页一个数量角标，加一个可点击跳转的 findings 面板），让人工审查者不必打开终端就能看到问题——如果 deck 里还有占位页，则显示一行「audit 已跳过」的提示代替。审查者可以直接在 `preview.html` 里给每页写自由文本批注，并导出为 `revision-request.json`（浏览器 Blob 下载，不联网也不写文件——preview 始终只读），交给 agent 通过 `pages/*.json` 回填。上文的 Claude Code 插件已把这套回路封装成 skill（[`skills/pptfast/SKILL.md`](./skills/pptfast/SKILL.md)）。这套回路本身由一个模型无关的内部基准测试（`tests/bench/`，不发布到 npm）机械化验证——固定题库，评估模型跟随该 skill 的表现，细节见 `tests/bench/README.md`。
 
 ## 路线图