promptfigure 0.2.0

package/src/doc/docx.mjs

@@ -0,0 +1,69 @@
+// docx.mjs — 服务端解析 .docx：段落文本 + 章节路径推断（§3.2 ¶2）
+// 只读 word/document.xml，不依赖样式文件；标题识别 pStyle Heading1-6 / 标题1-6
+import JSZip from "jszip";
+
+function decodeXml(s) {
+  return s
+    .replace(/&lt;/g, "<").replace(/&gt;/g, ">").replace(/&quot;/g, '"')
+    .replace(/&apos;/g, "'").replace(/&amp;/g, "&");
+}
+
+function headingLevel(style) {
+  if (!style) return 0;
+  const m = /^(?:Heading|heading|标题)\s*(\d)$/.exec(style.trim());
+  return m ? Math.min(parseInt(m[1], 10), 6) : 0;
+}
+
+export async function parseDocx(buf) {
+  const zip = await JSZip.loadAsync(buf);
+  const entry = zip.file("word/document.xml");
+  if (!entry) throw new Error("不是有效的 .docx（缺 word/document.xml）");
+  const xml = await entry.async("string");
+
+  const blocks = []; // { i, type:'heading'|'para', level, secPath, para, text }
+  const counters = [0, 0, 0, 0, 0, 0];
+  let paraIdx = 0;
+
+  const pRe = /<w:p\b[^>]*>([\s\S]*?)<\/w:p>|<w:p\b[^>]*\/>/g;
+  let m;
+  while ((m = pRe.exec(xml))) {
+    const inner = m[1] || "";
+    const styleM = /<w:pStyle w:val="([^"]+)"/.exec(inner);
+    const level = headingLevel(styleM && styleM[1]);
+    let text = "";
+    const tRe = /<w:t(?:\s[^>]*)?>([\s\S]*?)<\/w:t>/g;
+    let t;
+    while ((t = tRe.exec(inner))) text += t[1];
+    text = decodeXml(text).replace(/\s+/g, " ").trim();
+    if (!text && level === 0) continue;
+
+    if (level > 0) {
+      counters[level - 1] += 1;
+      for (let i = level; i < 6; i++) counters[i] = 0;
+      paraIdx = 0;
+    } else {
+      paraIdx += 1;
+    }
+
+    // 当前生效章节路径 = 计数器从第一个 0 处截断（全非 0 则取全部 6 级）
+    const firstZero = counters.indexOf(0);
+    const curSec = counters.slice(0, firstZero === -1 ? 6 : firstZero).join(".");
+
+    blocks.push({
+      i: blocks.length,
+      type: level > 0 ? "heading" : "para",
+      level,
+      secPath: level > 0 ? counters.slice(0, level).join(".") : curSec,
+      para: level > 0 ? 0 : paraIdx,
+      text,
+    });
+  }
+
+  // outline：树形章节
+  const outline = [];
+  for (const b of blocks) {
+    if (b.type !== "heading") continue;
+    outline.push({ path: b.secPath, level: b.level, title: b.text, blockIdx: b.i });
+  }
+  return { blocks, outline };
+}

package/src/doc/index.mjs

@@ -0,0 +1,35 @@
+// doc/index.mjs — 按扩展名分发解析
+import fs from "node:fs";
+import path from "node:path";
+import { parseDocx } from "./docx.mjs";
+import { parseTex } from "./tex.mjs";
+
+export function docKind(filePath) {
+  const base = filePath.split(/[\\/]/).pop();
+  const dot = base.lastIndexOf(".");
+  // 🔴 红队实测（2026-09-23 fuzz）：无扩展名路径（如目录）会整串当 ext 报出来（拼接损坏）
+  if (dot <= 0) throw new Error(`无法识别文件类型（没有扩展名）：${filePath}——pf open 需要指向 .docx / .tex / .pdf 文件`);
+  const ext = base.slice(dot + 1).toLowerCase();
+  if (ext === "docx") return "docx";
+  if (ext === "tex") return "tex";
+  if (ext === "pdf") return "pdf";
+  if (ext === "wps" || ext === "wpt") {
+    throw new Error(".wps/.wpt 是私有格式，本插件不支持。请在 WPS 里另存为 .docx 后再打开");
+  }
+  throw new Error(`不支持的格式 .${ext}（一期支持 .docx / .tex / .pdf）`);
+}
+
+export async function parseDocument(filePath) {
+  const kind = docKind(filePath);
+  if (kind === "docx") {
+    return { kind, ...(await parseDocx(fs.readFileSync(filePath))) };
+  }
+  if (kind === "tex") {
+    return { kind, ...parseTex(fs.readFileSync(filePath, "utf8"), path.dirname(filePath)) };
+  }
+  // pdf：服务端不做段落解析（段落级高亮走浏览器 textLayer），outline 留空
+  if (kind === "pdf") {
+    return { kind, blocks: [], outline: [] };
+  }
+  throw new Error("unreachable");
+}

package/src/doc/para.mjs

@@ -0,0 +1,54 @@
+// para.mjs — 段落引用 "§3.2 ¶2" 的解析、校验与定位
+// 🔴 用章节路径+段序，不用全局段序（全局段序会因增删段落整体漂移）
+
+export function parseRef(ref) {
+  if (!ref || typeof ref !== "string") throw new Error("锚点引用必须是字符串，如 \"§3.2 ¶2\"");
+  const m = /^\s*§?\s*([\d.]+)\s*(?:¶\s*(\d+))?\s*$/.exec(ref);
+  if (!m) {
+    throw new Error(`无法解析锚点引用 "${ref}"。格式：§<章节路径> ¶<段序>，例如 "§3.2 ¶2"（¶ 缺省时指该章节标题本身）`);
+  }
+  return { secPath: m[1].replace(/\.$/, ""), para: m[2] ? parseInt(m[2], 10) : 0 };
+}
+
+export function resolveRef(blocks, ref) {
+  const { secPath, para } = parseRef(ref);
+  const head = blocks.find((b) => b.type === "heading" && b.secPath === secPath);
+  if (!head) {
+    // 🔴 2026-09-21 死路修复：摘要/前言等无标题章节（tex 抽出的 §0）不在 outline 里，
+    // 但 pf doc search / pf doc data 会给出这类引用——不能因为没 heading 就拒绝自家工具给的引用。
+    // 只要该章节下有真实段落/图注块就放行；¶ 缺省时落到该章节第一个段落。
+    const same0 = blocks.filter(
+      (b) => (b.type === "para" || b.type === "caption") && b.secPath === secPath
+    );
+    if (!same0.length) {
+      const avail = blocks.filter((b) => b.type === "heading").map((b) => b.secPath).slice(0, 40);
+      throw new Error(`章节 §${secPath} 不存在。可用章节（前 40）：${avail.join(", ") || "（无）"}。先 pf doc outline 查看。`);
+    }
+    if (para === 0) return same0[0];
+    const t0 = same0.find((b) => b.para === para);
+    if (!t0) {
+      throw new Error(`§${secPath} 下只有 ${same0.length} 个段落/图注，¶${para} 不存在（若该章只有子章节没有直属段落，引用要到段落所在的子章节，如 §4.1 ¶2）`);
+    }
+    return t0;
+  }
+  if (para === 0) return head;
+  // 🔴 caption 块也算段落（tex.mjs 把 \caption 抽成 caption 块并占段号——doc figures 报的
+  // 图位置正是它）。只搜 type==="para" 会让 doc figures 给出的每个图位置都过不了校验。
+  const same = blocks.filter(
+    (b) => (b.type === "para" || b.type === "caption") && b.secPath === secPath
+  );
+  const target = same.find((b) => b.para === para);
+  if (!target) {
+    throw new Error(`§${secPath} 下只有 ${same.length} 个段落/图注，¶${para} 不存在（若该章只有子章节没有直属段落，引用要到段落所在的子章节，如 §4.1 ¶2）`);
+  }
+  return target;
+}
+
+// 变化检测：quote 是否还在原文里（确定性比对，不猜）
+export function checkQuote(blocks, quote) {
+  if (!quote) return "ok";
+  const norm = (s) => s.replace(/\s+/g, "");
+  const q = norm(quote);
+  if (!q) return "ok";
+  return blocks.some((b) => norm(b.text).includes(q)) ? "ok" : "changed";
+}

package/src/doc/tex.mjs

@@ -0,0 +1,161 @@
+// tex.mjs — 解析 .tex：\section 层级 + 段落（一期只读预览，不编译）
+// 支持 \input/\include 递归展开（CVPR 等模板 main.tex 只是壳，正文在 sec/*.tex）
+// 🔴 多文件共享同一个章节计数器 —— 否则每个 input 文件里的 § 都从 1 重新数
+// 图/表环境整体跳过，但 \caption{...} 单独抽成 caption 段（锚点定位高频目标）
+import fs from "node:fs";
+import path from "node:path";
+
+const SKIP_ENVS = /(figure|table|algorithm|wraptable|wrapfigure)\*?/;
+const SKIP_CMDS = /^\\(documentclass|usepackage|title|author|affiliations|email|maketitle|bibliographystyle|bibliography|input|includegraphics|graphicspath|label|ref|cite|pagestyle|thispagestyle|setlength|newcommand|renewcommand|def|let|providecommand|DeclareMathOperator|hypersetup|nocite|balance|printacmref|settopmatter|ccsdesc|keywords|terms|acmConference|acmBooktitle|acmPrice|acmISBN|acmDOI|fancyhead|fancyfoot|newenvironment|renewenvironment|vspace|hspace|quad|qquad|bibliographyplot|onecolumn|twocolumn|clearpage|newpage|tableofcontents|appendix|part|centering|raggedright|raggedleft|parbox|color|textcolor|definecolor|usebox|savebox|resizebox|scalebox|rotatebox|adjustbox|toprule|midrule|bottomrule|cmidrule|multicolumn|multirow|hline|cline|rowcolor|columncolor|arraystretch|tabcolsep|specialrule|addlinespace|makeatletter|makeatother|newlength|setcounter|addtocounter|value|numberwithin|allowdisplaybreaks|displaybreak|allowbreak|newline|linebreak|noindent|indent|parskip|par|smallskip|medskip|bigskip|vfill|cleardoublepage|cleardoublepage|frontmatter|mainmatter|backmatter|maketitle|printbibliography|endinput|includeonly|externaldocument|subfile|import|subimport|graphicspath)\b/;
+
+function readTexFile(baseDir, rel, visited) {
+  let p = path.resolve(baseDir, rel);
+  if (!/\.tex$/i.test(p)) p += ".tex";
+  if (visited.has(p) || !fs.existsSync(p)) return null;
+  visited.add(p);
+  return fs.readFileSync(p, "utf8");
+}
+
+export function parseTex(text, baseDir = ".") {
+  const state = { blocks: [], counters: [0, 0, 0], paraIdx: 0 };
+  parseInto(text, path.resolve(baseDir), state, new Set());
+  const outline = state.blocks.filter((b) => b.type === "heading")
+    .map((b) => ({ path: b.secPath, level: b.level, title: b.text, blockIdx: b.i }));
+  return { blocks: state.blocks, outline };
+}
+
+function parseInto(text, baseDir, state, visited) {
+  const blocks = state.blocks;
+  const counters = state.counters;
+  let paraIdx = state.paraIdx;
+  let buf = [];
+  let env = null;
+
+  const curSecPath = () => {
+    const k = counters.findIndex((c) => c === 0);
+    return counters.slice(0, k === -1 ? 3 : Math.max(k, 1)).join(".");
+  };
+
+  const flush = () => {
+    const t = buf.join(" ").replace(/\s+/g, " ").trim();
+    buf = [];
+    if (!t) return;
+    paraIdx += 1;
+    state.paraIdx = paraIdx;
+    blocks.push({ i: blocks.length, type: "para", level: 0, secPath: curSecPath(), para: paraIdx, text: t });
+  };
+
+  const pushHeading = (level, title) => {
+    flush();
+    counters[level - 1] += 1;
+    for (let i = level; i < 3; i++) counters[i] = 0;
+    paraIdx = 0;
+    state.paraIdx = 0;
+    blocks.push({ i: blocks.length, type: "heading", level, secPath: counters.slice(0, level).join("."), para: 0, text: title });
+  };
+
+  const pushCaption = (cap, images) => {
+    flush();
+    const t = cap.replace(/\s+/g, " ").trim();
+    if (!t && !images?.length) return;
+    paraIdx += 1;
+    state.paraIdx = paraIdx;
+    const b = { i: blocks.length, type: "caption", level: 0, secPath: curSecPath(), para: paraIdx, text: t };
+    // 原图清单：figure 环境里捕获的 \includegraphics 路径挂到图注块上（pf doc figures 用）
+    if (images?.length) b.images = images;
+    blocks.push(b);
+  };
+
+  const handleLine = (rawLine) => {
+    const line = rawLine.replace(/(?<!\\)%.*$/, "").trim();
+    if (!line) { if (!env) flush(); return; }
+
+    // ---- \input / \include 递归展开（共享计数器与 blocks）----
+    const inc = /^\\(?:input|include)\{([^}]+)\}/.exec(line);
+    if (inc) {
+      flush();
+      const sub = readTexFile(baseDir, inc[1], visited);
+      if (sub) parseInto(sub, baseDir, state, visited);
+      const rest = line.slice(inc[0].length).trim();
+      if (rest) handleLine(rest);
+      return;
+    }
+
+    // ---- 图/表环境：跳过内容，抓 caption ----
+    const beginEnv = /\\begin\{([^}]+)\}/.exec(line);
+    if (beginEnv && SKIP_ENVS.test(beginEnv[1])) {
+      flush();
+      env = { name: beginEnv[1] };
+    }
+    if (env) {
+      // 捕获 figure 环境内的原图引用（\includegraphics[opts]{path}，subfigure 多图全收）
+      const ig = /\\includegraphics(?:\[[^\]]*\])?\{([^}]+)\}/g;
+      let igm;
+      while ((igm = ig.exec(line))) (env.images ||= []).push(igm[1]);
+      const capM = /\\caption\*?(?:\[([^\]]*)\])?\{/.exec(line);
+      if (capM) {
+        const start = line.indexOf("{", capM.index) + 1;
+        let d = 1, cap = "";
+        for (let i = start; i < line.length && d > 0; i++) {
+          const ch = line[i];
+          if (ch === "{") d++;
+          else if (ch === "}") d--;
+          if (d > 0) cap += ch;
+        }
+        pushCaption(cap, env.images);
+        env.images = null;
+        env.captioned = true;
+      }
+      if (/\\end\{([^}]+)\}/.test(line) && RegExp.$1 === env.name) {
+        // 有原图引用但没写 caption 的环境（少见）：落一个 caption 块兜底，别让原图清单漏图
+        if (env.images?.length && !env.captioned) {
+          paraIdx += 1;
+          state.paraIdx = paraIdx;
+          blocks.push({
+            i: blocks.length, type: "caption", level: 0, secPath: curSecPath(), para: paraIdx,
+            text: "(figure without caption)", images: env.images,
+          });
+        }
+        env = null;
+      }
+      return;
+    }
+
+    // ---- 章节标题（括号配平，支持嵌套花括号）----
+    if (/^\\(?:sub){0,2}section\*?\s*\{/.test(line)) {
+      const level = line.startsWith("\\subsubsection") ? 3 : line.startsWith("\\subsection") ? 2 : 1;
+      const start = line.indexOf("{") + 1;
+      let d = 1, title = "";
+      for (let i = start; i < line.length && d > 0; i++) {
+        const ch = line[i];
+        if (ch === "{") d++;
+        else if (ch === "}") d--;
+        if (d > 0) title += ch;
+      }
+      title = title.replace(/\\[a-zA-Z]+\{([^}]*)\}/g, "$1").replace(/[{}]/g, "").trim();
+      pushHeading(level, title);
+      return;
+    }
+
+    // ---- 顺序敏感：\end{document} 之后的行全忽略（常有注释草稿）----
+    if (/^\\end\{document\}/.test(line)) { flush(); state.stop = true; return; }
+    if (state.stop) return;
+
+    if (SKIP_CMDS.test(line)) { flush(); return; }
+
+    if (/^\\begin\{(center|enumerate|itemize|quote|abstract|spacing|list|description)\}/.test(line)) { flush(); return; }
+
+    if (/^\\item\b/.test(line)) {
+      flush();
+      buf.push(line.replace(/^\\item\s*/, "").replace(/\[[^\]]*\]/, "").trim());
+      return;
+    }
+    buf.push(line);
+  };
+
+  for (const raw of text.split(/\r?\n/)) {
+    if (state.stop) break;
+    handleLine(raw);
+  }
+  flush();
+}

package/src/doc/texbuild.mjs

@@ -0,0 +1,158 @@
+// texbuild.mjs — LaTeX 本地编译（零侵入：输出只进 ~/.promptfigure，绝不写用户目录）
+// 策略：找本机 TeX 引擎 → 编译成 PDF 给 GUI 渲染（保真）；没有引擎时 CLI 给明确指引
+// 引擎优先级：tectonic（自动拉宏包、自动跑 bib）> latexmk > pdflatex（两遍，引用可能为 ?）
+import fs from "node:fs";
+import path from "node:path";
+import { spawnSync } from "node:child_process";
+import os from "node:os";
+
+const PF_BIN = path.join(os.homedir(), ".promptfigure", "bin");
+
+// ---- 引擎探测 ----
+function existsExec(p) {
+  try { fs.accessSync(p, fs.constants.X_OK); return fs.statSync(p).isFile(); } catch { return false; }
+}
+
+function findInPath(name) {
+  const exts = process.platform === "win32" ? [".exe", ".cmd", ".bat", ""] : [""];
+  for (const dir of (process.env.PATH || "").split(path.delimiter)) {
+    if (!dir) continue;
+    for (const ext of exts) {
+      const p = path.join(dir, name + ext);
+      if (existsExec(p)) return p;
+    }
+  }
+  return null;
+}
+
+function findTexLivePdflatex() {
+  const roots = ["C:/texlive"];
+  for (const root of roots) {
+    try {
+      const years = fs.readdirSync(root).filter((d) => /^\d{4}/.test(d)).sort().reverse();
+      for (const y of years) {
+        for (const bin of ["bin/windows", "bin/win32"]) {
+          const p = path.join(root, y, bin, "pdflatex.exe");
+          if (fs.existsSync(p)) return p;
+        }
+      }
+    } catch {}
+  }
+  return null;
+}
+
+function findMikTeX() {
+  const candidates = [
+    path.join(process.env.APPDATA || "", "MiKTeX", "miktex", "bin", "x64", "pdflatex.exe"),
+    "C:/Program Files/MiKTeX/miktex/bin/x64/pdflatex.exe",
+    "C:/Users/" + os.userInfo().username + "/AppData/Local/Programs/MiKTeX/miktex/bin/x64/pdflatex.exe",
+  ];
+  return candidates.find((c) => fs.existsSync(c)) || null;
+}
+
+export function detectTexEngine() {
+  // ① 插件自带的（~/.promptfigure/bin，pf setup-tex 下载到这里）
+  const local = path.join(PF_BIN, process.platform === "win32" ? "tectonic.exe" : "tectonic");
+  if (fs.existsSync(local)) return { engine: "tectonic", cmd: local, origin: "plugin" };
+  // ② PATH 与常见安装位置
+  const tectonic = findInPath("tectonic");
+  if (tectonic) return { engine: "tectonic", cmd: tectonic, origin: "system" };
+  const latexmk = findInPath("latexmk");
+  if (latexmk) return { engine: "latexmk", cmd: latexmk, origin: "system" };
+  const pdflatex = findInPath("pdflatex") || findTexLivePdflatex() || findMikTeX();
+  if (pdflatex) return { engine: "pdflatex", cmd: pdflatex, origin: "system" };
+  return null;
+}
+
+// ---- 编译 ----
+export function compileTex(docPath, { timeoutMs = 300000, outDir } = {}) {
+  const eng = detectTexEngine();
+  if (!eng) {
+    return {
+      ok: false,
+      reason: "no-engine",
+      message: "本机没有 TeX 编译环境。两条路：① 执行 pf setup-tex（自动下载便携版 tectonic 到插件目录）② 自装 TeX Live 或 MiKTeX 后重开",
+    };
+  }
+  if (!outDir) throw new Error("compileTex 需要 outDir（编译产物只允许写进插件目录）");
+  fs.mkdirSync(outDir, { recursive: true });
+
+  const srcDir = path.dirname(docPath);
+  let args, opts;
+  if (eng.engine === "tectonic") {
+    args = ["--outdir", outDir, docPath];
+    opts = {};
+  } else if (eng.engine === "latexmk") {
+    args = ["-pdf", "-interaction=nonstopmode", "-halt-on-error", `-outdir=${outDir}`, docPath];
+    opts = {};
+  } else {
+    // pdflatex 跑两遍解决交叉引用；output-directory 模式下允许在源目录找输入
+    const pass = () => spawnSync(eng.cmd, [
+      "-interaction=nonstopmode", "-halt-on-error", `-output-directory=${outDir}`, docPath,
+    ], { cwd: srcDir, timeout: timeoutMs / 2, encoding: "buffer" });
+    const r1 = pass();
+    if (r1.status !== 0) return fail(eng, r1, docPath);
+    const r2 = pass();
+    return collect(eng, r2, docPath, outDir, timeoutMs);
+  }
+
+  const run = spawnSync(eng.cmd, args, { cwd: srcDir, timeout: timeoutMs, encoding: "buffer", ...opts });
+  return collect(eng, run, docPath, outDir, timeoutMs);
+}
+
+function fail(eng, run, docPath) {
+  const log = run.stdout?.toString() + (run.stderr?.toString() || "");
+  return {
+    ok: false,
+    engine: eng.engine,
+    reason: "compile-error",
+    logTail: log.slice(-1500),
+    message: `TeX 编译失败（${eng.engine}）。常见原因：宏包缺失（tectonic 需要联网拉包，走代理时先设 HTTPS_PROXY）/ 文档本身有错。可用用户目录下的 .log 复核。`,
+  };
+}
+
+function collect(eng, run, docPath, outDir, timeoutMs) {
+  if (run.error?.killed || run.error?.code === "ETIMEDOUT") {
+    return { ok: false, engine: eng.engine, reason: "timeout", message: `编译超时（>${Math.round(timeoutMs / 1000)}s）。首次编译 tectonic 要在线拉宏包，可能较慢，重跑一次通常快很多。` };
+  }
+  if (run.status !== 0) return fail(eng, run, docPath);
+  const pdfPath = path.join(outDir, path.basename(docPath).replace(/\.tex$/i, "") + ".pdf");
+  if (!fs.existsSync(pdfPath)) {
+    return { ok: false, engine: eng.engine, reason: "no-pdf", logTail: (run.stdout?.toString() || "").slice(-1500), message: "编译命令成功但没产出 PDF（检查文档是否有 \\documentclass）" };
+  }
+  return { ok: true, engine: eng.engine, pdfPath };
+}
+
+// ---- pf setup-tex：下载便携 tectonic 到插件目录（国内镜像优先，逐个回退）----
+const MIRRORS = [
+  "https://ghfast.top/https://github.com",
+  "https://gh-proxy.com/https://github.com",
+  "https://ghproxy.net/https://github.com",
+  "https://github.com",
+];
+const TECTONIC_VER = "0.15.0";
+const TECTONIC_URL = `/tectonic-typesetting/tectonic/releases/download/tectonic%40${TECTONIC_VER}/tectonic-${TECTONIC_VER}-x86_64-pc-windows-msvc.zip`;
+
+export async function setupTex({ onLog = console.log } = {}) {
+  const target = path.join(PF_BIN, "tectonic.exe");
+  if (fs.existsSync(target)) { onLog(`✅ 已有 ${target}`); return { ok: true, already: true }; }
+  fs.mkdirSync(PF_BIN, { recursive: true });
+  const { execFileSync } = await import("node:child_process");
+  const zip = path.join(PF_BIN, "tectonic.zip");
+  for (const m of MIRRORS) {
+    try {
+      onLog(`↓ 尝试 ${m.split("/")[2]} ...`);
+      execFileSync("curl", ["-sL", "--max-time", "180", "-o", zip, m + TECTONIC_URL], { stdio: "ignore" });
+      if (fs.existsSync(zip) && fs.statSync(zip).size > 1000000) break;
+    } catch {}
+  }
+  if (!fs.existsSync(zip) || fs.statSync(zip).size < 1000000) {
+    return { ok: false, message: "下载失败（所有镜像都不通）。手动方案：开代理后重试，或自装 TeX Live / MiKTeX" };
+  }
+  // 解压（用 powershell Expand-Archive，零依赖）
+  execFileSync("powershell", ["-NoProfile", "-Command", `Expand-Archive -Force '${zip}' '${PF_BIN}'`], { stdio: "ignore" });
+  fs.rmSync(zip, { force: true });
+  if (!fs.existsSync(target)) return { ok: false, message: "解压后没找到 tectonic.exe" };
+  onLog(`✅ tectonic ${TECTONIC_VER} → ${target}`);
+  return { ok: true, path: target };
+}

package/src/docsearch.mjs

@@ -0,0 +1,96 @@
+// docsearch.mjs — 🔴 原文检索 + 结果数据抽取（2026-09-21 用户反馈"别人用不了：不读文章、找不到结果数据"）：
+// 弱 AI 只有 pf doc outline（章节树），定位内容全靠猜——结果/实验章节里的数字句
+// （准确率、提升幅度、对比基线）更是无处可查。这里给两个确定性工具，全部本地计算、不依赖 daemon：
+//   - searchBlocks：关键词全文检索（para/caption 块），返回 §引用 + 命中片段 → "去哪读"
+//   - dataSentences：抽定量句（%/小数/×/± 模式）→ "结果数据在哪"
+// 纯函数：CLI 与测试共用。分句走 extract.splitSentences（2026-09-22 中英通用，中文论文不再糊成一坨）。
+import { splitSentences } from "./extract.mjs";
+
+// 关键词全文检索。kws 全部小写化匹配；命中块按命中次数排序（多关键词同时命中的排前）。
+// 返回 [{ ref, secPath, para, type, snippet, hits }]，snippet 是首个命中点前后 ~90 字符窗口。
+export function searchBlocks(blocks = [], kws = [], { limit = 12 } = {}) {
+  const terms = (Array.isArray(kws) ? kws : String(kws).split(/[\s,，]+/))
+    .map((k) => String(k).trim().toLowerCase()).filter(Boolean);
+  if (!terms.length) return [];
+  const out = [];
+  for (const b of blocks) {
+    if (b.type !== "para" && b.type !== "caption") continue;
+    const text = String(b.text || "");
+    if (!text) continue;
+    const low = text.toLowerCase();
+    let hits = 0;
+    const hitTerms = [];
+    for (const t of terms) {
+      const n = low.split(t).length - 1;
+      if (n > 0) { hits += n; hitTerms.push(t); }
+    }
+    if (!hits) continue;
+    // 片段窗口：第一个出现的关键词位置
+    let pos = -1;
+    for (const t of terms) { const p = low.indexOf(t); if (p >= 0 && (pos < 0 || p < pos)) pos = p; }
+    const start = Math.max(0, pos - 60);
+    const snippet = (start > 0 ? "…" : "") + text.slice(start, pos + 90).replace(/\s+/g, " ") + (start + 90 < text.length ? "…" : "");
+    out.push({
+      ref: `§${b.secPath}${b.para ? " ¶" + b.para : ""}`,
+      secPath: b.secPath, para: b.para || null, type: b.type,
+      snippet, hits, hitTerms,
+    });
+  }
+  out.sort((a, b) => b.hits - a.hits);
+  return out.slice(0, limit);
+}
+
+// 定量句抽取（"结果数据"定位）：命中任一模式的句子算定量句。
+// 模式全部确定性：百分比 / 带小数的数值 / ×倍数 / ±波动 / dB/mAP 等常见指标后缀。
+// 返回 [{ ref, sentence }]，按原文顺序。
+const METRIC_TAIL = /\b(?:dB|PSNR|SSIM|LPIPS|mAP|FID|Acc|accuracy|F1)\b/i;
+const QUANT_RE = new RegExp(
+  [
+    "\\d+(?:\\.\\d+)?\\s*%",          // 34.2%
+    "\\d+\\.\\d+",                     // 小数（0.913、3.14）
+    "\\d+\\s*[×x]\\s*\\d+",           // 2× 提升 / 1024×1024
+    "\\d+(?:\\.\\d+)?\\s*[±]\\s*\\d+",// 1.2±0.1
+    "[×x]\\s*\\d+(?:\\.\\d+)?",       // ×1.5
+    METRIC_TAIL.source,                // 指标名
+  ].join("|"),
+  "g"
+);
+
+export function dataSentences(blocks = [], { limit = 20 } = {}) {
+  const out = [];
+  for (const b of blocks) {
+    if (b.type !== "para") continue;
+    const text = String(b.text || "");
+    if (!text) continue;
+    const sents = splitSentences(text);
+    for (const s of sents) {
+      // 🔴 最短句过滤 CJK 感知（2026-09-22）：原 s.length<15 按英文字符定标，
+      // "方案二准确率达到91.3%。"（14 字符）被误杀——中文信息密度高，汉字按 2 计
+      const eff = (s.match(/[A-Za-z0-9]+/g) || []).join("").length
+        + (s.match(/[\u4e00-\u9fff]/g) || []).length * 2
+        + (s.match(/[^\u4e00-\u9fffA-Za-z0-9\s]/g) || []).length;
+      if (eff < 15) continue;
+      QUANT_RE.lastIndex = 0;
+      if (!QUANT_RE.test(s)) continue;
+      out.push({
+        ref: `§${b.secPath}${b.para ? " ¶" + b.para : ""}`,
+        sentence: s.slice(0, 200) + (s.length > 200 ? "…" : ""),
+      });
+      if (out.length >= limit) return out;
+    }
+  }
+  return out;
+}
+
+// 数据句按章节聚合视图（pf doc data 输出用）：[{ secPath, items: [{ref, sentence}] }]
+export function dataBySection(blocks = [], { limit = 40 } = {}) {
+  const flat = dataSentences(blocks, { limit });
+  const bySec = new Map();
+  for (const it of flat) {
+    // 归一化章节键：去掉引用里的 § 前缀与段落号（CLI 输出时统一加 §）
+    const key = it.ref.replace(/^§/, "").replace(/ ¶\d+$/, "");
+    if (!bySec.has(key)) bySec.set(key, []);
+    bySec.get(key).push(it);
+  }
+  return [...bySec.entries()].map(([secPath, items]) => ({ secPath, items }));
+}

package/src/entity-pair.mjs

@@ -0,0 +1,17 @@
+// entity-pair.mjs — 实体双语对照解析（2026-09-22 用户拍板补差距②）。
+// 独立叶子模块：quality.mjs 与 craft.mjs 都要用，放任何一边都会成环。
+//
+// 实测痛点：中文用户把实体译成英文喂 craft（standard 档中文乱码率高），但译完溯源就断——
+// checkEntitySource 对"脚本不一致"的实体只能 skipped（fail open）。对照格式
+// 「英文标签|原文词」（如 "Weighted Box Fusion|加权框融合"）两头都占：
+//   - 溯源对账查原文侧（确定性）
+//   - 卡面/提示词用英文侧（不乱码）
+export function splitEntityPair(e) {
+  const raw = String(e || "").trim();
+  // 🔴 渗透实测（2026-09-22 子智能体 B）：一侧为空的残缺对照（"Label|" / "|原文"）不能当
+  // 普通实体放行——英文侧会连 "|" 一起印进卡面，原文侧静默丢掉。标记 invalid 让质量门拦。
+  if (!raw.includes("|")) return { label: raw, src: null };
+  const parts = raw.split("|").map((s) => s.trim());
+  if (parts.length === 2 && parts[0] && parts[1]) return { label: parts[0], src: parts[1] };
+  return { label: raw, src: null, invalid: true };
+}

package/src/events.mjs

@@ -0,0 +1,52 @@
+// events.mjs — events.jsonl 只 append + SSE 广播
+// 🔴 永不改写/删除已有行 —— AI 回放审批历史全靠这个文件
+// 🔴 不用 fs.watch —— 进程间通知走本地 HTTP/SSE（跨平台踩过的坑）
+import fs from "node:fs";
+import path from "node:path";
+import { projectByDocId } from "./store.mjs";
+
+const clients = new Set(); // { docId, res }
+
+function eventsFile(docId) {
+  const dir = projectByDocId(docId);
+  fs.mkdirSync(dir, { recursive: true });
+  return path.join(dir, "events.jsonl");
+}
+
+export function appendEvent(docId, type, data = {}) {
+  const line = JSON.stringify({ t: new Date().toISOString(), type, ...data });
+  fs.appendFileSync(eventsFile(docId), line + "\n");
+  broadcast(docId, { type, ...data });
+  return line;
+}
+
+export function readEvents(docId, limit = 200) {
+  try {
+    const raw = fs.readFileSync(eventsFile(docId), "utf8").trim();
+    if (!raw) return [];
+    const lines = raw.split("\n");
+    return lines.slice(-limit).map((l) => { try { return JSON.parse(l); } catch { return { type: "corrupt", raw: l }; } });
+  } catch {
+    return [];
+  }
+}
+
+export function broadcast(docId, payload) {
+  const msg = `data: ${JSON.stringify(payload)}\n\n`;
+  for (const c of clients) {
+    if (c.docId !== docId) continue;
+    try { c.res.write(msg); } catch { clients.delete(c); }
+  }
+}
+
+export function sseHandler(docId, req, res) {
+  res.writeHead(200, {
+    "Content-Type": "text/event-stream",
+    "Cache-Control": "no-store",
+    Connection: "keep-alive",
+  });
+  res.write(`data: ${JSON.stringify({ type: "hello", docId })}\n\n`);
+  const entry = { docId, res };
+  clients.add(entry);
+  req.on("close", () => clients.delete(entry));
+}