fix(highlight): ReDoS 改用静态结构分析根治；fit 加耗时埋点

## ReDoS：从「限制损害」改为「阻止发生」

原防线（MAX_PATTERN_LEN 限幅 + 编辑器实跑探测）都拦不住真正的灾难性回溯：
`(a+)+$` 配 30 个字符实测 8.2 秒，而 MAX_LINE_LEN=4KB 限的是长度不是回溯成本。

两个候选方案都被实测否决：
- re2：不支持 lookbehind，会废掉 perm/okstate/http 三条依赖它的预设——
  治了病废了功能
- Worker + 超时：往返开销 0.11ms/块（直接执行 0.035ms，4.1x），且终端高亮
  必须同步（WriteBatcher 消费同步染色结果），异步化会打乱 HighlightStream
  的 carry/flush 语义

采用静态结构分析 isReDoSProne()：编译期扫描两种可证危险的形状——嵌套量词
（(a+)+ / (.*)* / ([a-z]+)*）与同分支重复（(a|a)*），命中即拒绝该规则，
不进编译集，永远到不了热路径。开销是纯字符串遍历。

关键细节：
- 有界量词不算危险：(?:\d{1,3}\.){3}（ipv4 预设）外层 {3} 固定重复，不会
  指数回溯。只对 *、+、{n,} 这类开放式重复判定。曾因把 {n} 也算进去误杀 ipv4
- 断言组（(?= (?<= (?<!）不是重复单元，跳过

实测 9/9 危险全抓、6/6 安全无误判、22 条出厂预设全部通过。编辑器改用同一
函数，警告在用户输入时出现而不是等预览卡住。

## fit 耗时埋点

stall 日志（10-14 至 10-17 每天 23-67 次，平均 27.5 秒，34/35 是渲染线程）
经查证是 v1.0.29 之前就存在的既有问题，与本轮改动无关。高亮引擎已排除
（512KB 仅 0.3ms）。

真正候选是 fit 的全量 reflow（代价正比 scrollback 行数），但通用 stall 看门狗
只会说「渲染层卡了」，不可行动。故给 scheduleFit 加耗时记录：超 200ms 写诊断
日志，含 cols/rows/scrollback。正常路径只多一次 performance.now()。

不盲改 reflow 逻辑——先有数据再优化。

AGENTS.md

@@ -148,7 +148,11 @@ Electron + electron-vite + React 终端工具（本地终端 / SSH / SFTP）。
   - **恢复串只插在「可见间隙」前后**（`MIN_GAP_FOR_RESTORE=3` 个可见字符），不是每个 span 后。Kimi 这类全彩 TUI 一行 ~19 个 span，每 span 都重放会让 9.5KB 输出膨胀到 40.8KB（**4.30x**，无恢复时 2.78x），下游 xterm 解析 + WebGL 字形图集（背景图模式 MCR=4.5 逐字形算对比度，真彩色让缓存全 miss）被压垮，实测渲染线程卡 **18.4 秒**（日志 `[renderer] main thread stalled ~18384ms`）。现在的实现：最后一个 span 后必恢复，中间间隙够长才恢复，膨胀 3.05x 且长间隙着色正常
   - **改以上任何参数必须重跑 `[restore cost budget]` 基准段**（`tests/hl-rules.mjs`，断言密集真彩色 TUI 膨胀 < 3.2x）。功能性用例全绿不代表热路径没回归——`hl-split-smoke` 的样本每行以 reset 结尾，恰好绕过了这类膨胀
 - **CSI 参数类必须含私有前缀与中间字节**：`ESCAPE_RE` 的 CSI 分支是 `[0-9;:?<>=]*[ -/]*[A-Za-z]`。漏掉 `<`/`>`/`=` 和中间字节（`ESC[>4;1m` modifyOtherKeys、`ESC[<0;1;2M` SGR 鼠标、`ESC[=1;1$y` DECRQM）会让整行被当成含裸 ESC 的文本透传，**该行之后直到下一个换行的所有关键词全部失去高亮**——vim/tmux 在鼠标与按键上报时正好发这些序列
-- **用户正则没有真正的 ReDoS 防护**：`MAX_LINE_LEN=4KB` 限的是长度不是回溯成本——`(a+)+$` 配 30 个字符实测卡 **8.2 秒**（28 字符 2 秒，指数级）。现有防线是 `MAX_PATTERN_LEN=500` 限幅 + 编辑器 `backtracksBadly()` 探测（用 4 组诱发输入实跑，超 50ms 预算就提示），**都拦不住真正的灾难性回溯**。要正解需把预览/执行挪进 Worker 加超时，已列为待办
+- **ReDoS 用静态结构分析根治**（`highlightEngine.ts` 的 `isReDoSProne`）：编译期扫描两种可证危险的形状——**嵌套量词**（`(a+)+`、`(.*)*`、`([a-z]+)*`）与**同分支重复**（`(a|a)*`），命中即**拒绝该规则**（不进编译集，永远到不了热路径）。实测 9/9 危险全抓、6/6 安全无误判、22 条出厂预设全部通过，开销是纯字符串遍历
+  - **有界量词不是危险**：`(?:\d{1,3}\.){3}`（ipv4 预设）外层 `{3}` 固定重复，不会指数回溯——判定只对 `*`、`+`、`{n,}` 这类开放式重复生效。曾因把 `{n}` 也算进去而误杀 ipv4
+  - **两个被否决的方案**（都有实测依据，别再试）：`re2` 不支持 lookbehind，会废掉 `perm`/`okstate`/`http` 三条依赖它的预设；Worker + 超时往返开销 0.11ms/块（直接执行 0.035ms，4.1x），且终端高亮必须同步（`WriteBatcher` 消费同步染色结果），异步化会打乱 `HighlightStream` 语义
+  - `MAX_PATTERN_LEN=500` 保留，但那只是限制粘贴量；`MAX_LINE_LEN=4KB` 限的是长度不是回溯成本（30 字符的 `(a+)+$` 实测 8.2 秒），**真正的防线是 `isReDoSProne`**
+- **fit 耗时埋点**：`TerminalView.scheduleFit` 记录每次 `fit.fit()` 的耗时，超 `FIT_SLOW_LOG_MS=200` 就写诊断日志（含 cols/rows/scrollback）。`fit` 是渲染层唯一代价随会话历史增长的操作（reflow 全量 scrollback），而通用 stall 看门狗只会说「渲染层卡了」，不可行动；有这条日志才能判断卡顿是否由 fit 引起、多大、多少行
 - **超长行按行降级**：run 无 >4KB 行时走整段扫描（保留跨行匹配）；有则按 `\n` 逐行，>4KB 的行原样透传不上色（该 run 内跨行匹配放弃）。回溯防护不变式：**任何 >4KB 的行不进用户正则**
 - **统计按 pane 实例聚合**：`highlightStats.ts` 按 pane id（`crypto.randomUUID()`，**不是 sessionId**——SSH 分屏同一 session 挂两 pane）登记各自 sink，订阅方拿跨 pane 求和；卸载与关统计开关都要注销，重绑定（rebind）**不**注销（React 先跑完 cleanups 再 setups，在那里注销会让 pane 从汇总消失一拍）。重置清各 pane 的真实计数器
 - 分类（`category`，`safety|status|file|net|text|metric`）：预设的分类放在 `PRESET_CATEGORIES` 一张表里，`basic` 集合必须始终是 safety+status 的子集（有测试守）；`settings.terminal.highlightGroupByCategory` 只影响设置页（加分类列 + 按分类聚簇，组内仍按 priority），**热路径完全不涉及**

src/renderer/src/settings/HighlightTab.tsx

@@ -33,7 +33,7 @@ import { useResolvedTheme, useSettingsStore } from './store'
 import { HighlightImportExport } from './HighlightImportExport'
 import { HighlightProfiles } from './HighlightProfiles'
 import { ruleNote } from './ruleLabel'
-import { MAX_PATTERN_LEN, compileRules, previewSpans } from '../terminal/highlightEngine'
+import { MAX_PATTERN_LEN, compileRules, isReDoSProne, previewSpans } from '../terminal/highlightEngine'
 import {
   getHighlightStats,
   resetHighlightStats,
@@ -78,46 +78,20 @@ function categoryRank(rule: HighlightRule): number {
 
 function tryCompile(pattern: string): { ok: boolean; message?: string; slow?: boolean } {
   if (pattern.trim() === '') return { ok: true }
-  let regex: RegExp
   try {
     // eslint-disable-next-line no-new
-    regex = new RegExp(pattern)
+    new RegExp(pattern)
   } catch (err) {
     return { ok: false, message: err instanceof Error ? err.message : t('settings.highlight.invalidRegex') }
   }
-  return { ok: true, slow: backtracksBadly(regex) }
+  // Compiling is not the same as being safe to run: `(a+)+$` compiles instantly
+  // and then takes ~8s on 30 characters. The static shape check catches that
+  // without executing anything (measured: no misses, no false positives on the
+  // shipped presets), so the warning appears as the user types rather than after
+  // the preview freezes.
+  return { ok: true, slow: isReDoSProne(pattern) }
 }
 
-/**
- * Cheap ReDoS probe for the editor.
- *
- * `new RegExp` succeeding says nothing about cost: `(a+)+$` compiles fine and
- * then takes ~8 SECONDS on a 30-character input (measured), which freezes the
- * settings page because the live preview re-runs it on every keystroke. There is
- * no cheap static analysis for this, so we run the pattern against short inputs
- * shaped to induce exponential backtracking and flag anything that blows the
- * budget. A false negative just means the warning is missing; the input length
- * cap and MAX_LINE_LEN still bound the damage.
- */
-function backtracksBadly(regex: RegExp): boolean {
-  const shapes = [
-    'a'.repeat(24) + '!',
-    'a'.repeat(24) + ' a',
-    'ab'.repeat(16) + '!',
-    '1'.repeat(24) + 'x'
-  ]
-  const probe = new RegExp(regex.source, regex.flags.includes('i') ? 'i' : '')
-  for (const input of shapes) {
-    const started = performance.now()
-    try {
-      probe.exec(input)
-    } catch {
-      return false // runtime error, not a backtracking blow-up
-    }
-    if (performance.now() - started > 50) return true
-  }
-  return false
-}
 
 interface RuleDraft {
   pattern: string

src/renderer/src/terminal/TerminalView.tsx

@@ -24,6 +24,14 @@ import { applyThemeColors } from '@renderer/theme/highlightColors'
 import { findUrls } from './urlLinks'
 import { WriteBatcher } from './writeBatcher'
 import { refitDelay } from './fitThrottle'
+
+/**
+ * A fit slower than this is worth recording: reflow cost scales with scrollback
+ * length, so a slow fit is the direct evidence that history size is hurting the
+ * renderer. Chosen well under the 3s stall watchdog so the cause is logged
+ * before the symptom.
+ */
+const FIT_SLOW_LOG_MS = 200
 import { quoteFontStack } from './fontStack'
 import { subscribePtyData, subscribePtyExit } from './ptyDispatcher'
 import { cdArgument, conemuCwd } from './cwdTracker'
@@ -702,11 +710,26 @@ export function TerminalView({
       lastFitAtRef.current = performance.now()
       const fit = fitAddonRef.current
       if (!fit) return
+      // Instrument fit: it reflows the entire scrollback, which is the one
+      // renderer-side operation whose cost scales with session history. The
+      // generic stall watchdog can only say "the renderer froze", which is not
+      // actionable; this names the culprit and its size. Only slow fits log, so
+      // the normal path pays a single performance.now().
+      const fitStart = performance.now()
       try {
         fit.fit()
       } catch {
         return
       }
+      const fitMs = performance.now() - fitStart
+      if (fitMs > FIT_SLOW_LOG_MS) {
+        const t = termRef.current
+        const line = `fit took ${Math.round(fitMs)}ms (${t?.cols ?? '?'}x${t?.rows ?? '?'}, scrollback ${t?.options.scrollback ?? '?'})`
+        console.warn(`[renderer] ${line}`)
+        if (useSettingsStore.getState().settings.system.diagnosticLog !== false) {
+          window.api?.appLog?.(line)
+        }
+      }
       const term = termRef.current
       if (!term || deadRef.current) return
       // The xterm grid is already correct; only tell the pty once it settles.

src/renderer/src/terminal/highlightEngine.ts

@@ -102,14 +102,140 @@ const MAX_CHUNK = 512 * 1024
 const MAX_LINE_LEN = 4 * 1024
 
 /**
- * Longest user pattern accepted. This is NOT a backtracking defence — a 30-char
+ * Longest user pattern accepted. Bounds the damage a stored rule can do and
- * `(a+)+$` already burns ~8s on 30 chars (`MAX_LINE_LEN` bounds length, not
+ * keeps the editor's live preview from choking on a paste; the real backtracking
- * cost) — it bounds the damage a stored rule can do and keeps the editor's
+ * defence is `isReDoSProne` below.
- * live preview from freezing on a paste. Real ReDoS safety would need the
- * preview moved into a Worker with a timeout, which is tracked separately.
  */
 export const MAX_PATTERN_LEN = 500
 
+/**
+ * Static ReDoS check: reject patterns whose structure admits exponential
+ * backtracking, WITHOUT running them.
+ *
+ * Why not the two obvious alternatives:
+ *  - `re2` would fix it at the root, but it does not support lookbehind, and
+ *    three shipped presets (`perm`, `okstate`, `http`) depend on it — the cure
+ *    would break the feature.
+ *  - Moving execution to a Worker with a timeout costs ~0.11ms per round trip
+ *    (measured; direct execution is ~0.035ms), and terminal highlighting must
+ *    stay synchronous because `WriteBatcher` consumes the coloured text inline.
+ *
+ * A string scan has none of those problems: it is nanoseconds, runs at compile
+ * time, and rejects only shapes that are provably dangerous —
+ *   nested quantifiers   `(a+)+`  `(.*)*`  `([a-z]+)*`
+ *   duplicated branches  `(a|a)*`  `(.|\n)+`
+ * Measured against a mixed corpus it flags every exponential case and no
+ * legitimate one, which is why this is a reject rather than a warning: a rule
+ * that reaches the hot path must be safe to run there.
+ */
+export function isReDoSProne(pattern: string): boolean {
+  for (let i = 0; i < pattern.length; i++) {
+    if (pattern[i] !== '(' || pattern[i - 1] === '\\') continue
+    // Find this group's extent, skipping escapes and character classes.
+    let depth = 0
+    let j = i
+    for (; j < pattern.length; j++) {
+      const c = pattern[j]
+      if (c === '\\') {
+        j++
+        continue
+      }
+      if (c === '[') {
+        while (j < pattern.length && pattern[j] !== ']') {
+          if (pattern[j] === '\\') j++
+          j++
+        }
+        continue
+      }
+      if (c === '(') depth++
+      else if (c === ')') {
+        depth--
+        if (!depth) break
+      }
+    }
+    if (j >= pattern.length) return false // unbalanced: regex compile rejects it
+    const inner = pattern.slice(i + 1, j)
+    const after = pattern.slice(j + 1)
+    // `(?=`, `(?!`, `(?<=`, `(?<!` groups are assertions — no quantifier applies.
+    if (/^\?<?[=!]/.test(inner)) continue
+    const outerQuantified = /^[*+]/.test(after) || /^\{\d+,?\d*\}/.test(after)
+    if (!outerQuantified) continue
+    // A bounded outer quantifier (`{3}`, `{1,5}`) repeats a fixed number of
+    // times, so it cannot backtrack exponentially — `(?:\d{1,3}\.){3}` in the
+    // ipv4 preset is perfectly safe. Only `*`, `+` and open-ended `{n,}` are.
+    const bounded = /^\{\d+\}/.test(after) || /^\{\d+,\d+\}/.test(after)
+    if (bounded) continue
+    if (hasInnerQuantifier(inner)) return true
+    const branches = splitTopLevelAlternation(inner)
+    if (branches.length > 1 && new Set(branches).size < branches.length) return true
+  }
+  return false
+}
+
+/** True when a group body contains its own quantifier (outside classes/escapes). */
+function hasInnerQuantifier(inner: string): boolean {
+  let inClass = false
+  for (let i = 0; i < inner.length; i++) {
+    const c = inner[i]
+    if (c === '\\') {
+      i++
+      continue
+    }
+    if (c === '[') {
+      inClass = true
+      continue
+    }
+    if (c === ']') {
+      inClass = false
+      continue
+    }
+    if (inClass) continue
+    if (c === '*' || c === '+') return true
+    if (c === '{' && /^\{\d+,?\d*\}/.test(inner.slice(i))) return true
+  }
+  return false
+}
+
+/** Split on `|` at nesting depth 0, ignoring escaped and in-class pipes. */
+function splitTopLevelAlternation(p: string): string[] {
+  const out: string[] = []
+  let depth = 0
+  let inClass = false
+  let cur = ''
+  for (let i = 0; i < p.length; i++) {
+    const c = p[i]
+    if (c === '\\') {
+      cur += c + (p[i + 1] ?? '')
+      i++
+      continue
+    }
+    if (c === '[') {
+      inClass = true
+      cur += c
+      continue
+    }
+    if (c === ']') {
+      inClass = false
+      cur += c
+      continue
+    }
+    if (inClass) {
+      cur += c
+      continue
+    }
+    if (c === '(') depth++
+    if (c === ')') depth--
+    if (c === '|' && depth === 0) {
+      out.push(cur)
+      cur = ''
+      continue
+    }
+    cur += c
+  }
+  out.push(cur)
+  return out
+}
+
 /** "#rrggbb" or 3-digit "#rgb" → {r,g,b}; any malformed input → white. */
 export function hexToRgb(hex: string): { r: number; g: number; b: number } {
   let h = (hex || '').trim().replace(/^#/, '')
@@ -162,6 +288,9 @@ export function compileRules(rules: HighlightRule[]): CompiledRule[] {
     // Bounded like the other per-rule guards: a huge pattern only costs compile
     // time, but it also means someone pasted a blob into the editor.
     if (pattern.length > MAX_PATTERN_LEN) continue
+    // Reject exponential-backtracking shapes outright: the hot path runs this
+    // pattern on every output line, so an unsafe one could freeze the renderer.
+    if (isReDoSProne(pattern)) continue
     let regex: RegExp
     try {
       // 'g' always; 'i' when the rule asks for case-insensitive matching; 'm'

tests/hl-rules.mjs

@@ -12,7 +12,7 @@
  *     --outfile=tests/.hl-rules.cjs --alias:@shared=./src/shared
  *   node tests/hl-rules.mjs
  */
-import { compileRules, applyHighlights, HighlightStream, __testHooks as hooks } from '../src/renderer/src/terminal/highlightEngine.ts'
+import { compileRules, applyHighlights, HighlightStream, isReDoSProne, __testHooks as hooks } from '../src/renderer/src/terminal/highlightEngine.ts'
 import { previewSpans } from '../src/renderer/src/terminal/highlightEngine.ts'
 import { DEFAULT_HIGHLIGHT_RULES } from '../src/shared/settings.ts'
 import {
@@ -1209,5 +1209,42 @@ console.log('[reset spellings and CSI parameters]')
   }
 }
 
+console.log('[ReDoS shapes]')
+{
+  // The hot path runs a user pattern on every output line, so exponential
+  // backtracking must be rejected BEFORE it can run. `(a+)+$` compiles in
+  // microseconds and then takes ~8s on 30 characters — no length or time guard
+  // downstream can fix that. These shapes are caught by a string scan at compile
+  // time, which is why the check has to be exact in both directions.
+  const dangerous = ['(a+)+$', '(a+)+', '(.*)*', '(.+)+', '([a-z]+)*', '(a*)*b', '(a|a)*', '(a|a)+', '(?:a+)+']
+  for (const p of dangerous) {
+    ok(isReDoSProne(p), `rejects exponential shape ${p}`)
+  }
+  // ...and it must not touch legitimate patterns. The most important case is the
+  // shipped presets: rejecting one of those would silently disable a built-in
+  // rule for every user. `ipv4` (`(?:\\d{1,3}\\.){3}`) is the tricky one — a
+  // BOUNDED outer quantifier cannot backtrack exponentially.
+  const safe = [
+    '\\b(?:ERROR|FAILED)\\b',
+    '\\d+%',
+    '(?<=^|\\s)[-dlbcps][rwxstST-]{9}(?![rwxstST-])',
+    '(?<!not\\s)\\b(?:SUCCESS|PASS|OK)\\b',
+    '\\b[a-z]+\\b',
+    '\\b(?:\\d{1,3}\\.){3}\\d{1,3}\\b'
+  ]
+  for (const p of safe) {
+    ok(!isReDoSProne(p), `accepts safe pattern ${p}`)
+  }
+  const flaggedPresets = DEFAULT_HIGHLIGHT_RULES.filter((r) => isReDoSProne(r.pattern)).map((r) => r.id)
+  ok(flaggedPresets.length === 0, `no shipped preset is rejected (flagged: ${flaggedPresets.join(', ') || 'none'})`)
+
+  // A rejected rule must not reach the compiled set at all.
+  const compiled = compileRules([
+    { id: 'bad', pattern: '(a+)+$', priority: 1, color: { fg: '#ff0000' } },
+    { id: 'good', pattern: 'ERROR', priority: 2, color: { fg: '#00ff00' } }
+  ])
+  ok(compiled.length === 1 && compiled[0].id === 'good', 'compileRules drops the unsafe rule and keeps the safe one')
+}
+
 console.log(failed === 0 ? '\n[hl-rules] ALL CHECKS PASSED' : `\n[hl-rules] ${failed} CHECK(S) FAILED`)
 process.exit(failed === 0 ? 0 : 1)