@feng0531/dsh-code-review 0.85.4 → 0.85.5

package/CHANGELOG.md

@@ -2,6 +2,14 @@
 
 本文件介绍对使用者有意义的变化。安装与参数见 [README](README.md)，本次发布正文见 [发布说明](RELEASE-NOTES.md)。历史提交和安装包保留在 [GitHub Releases](https://github.com/f-e-n-g-0531/dsh-code-review/releases)；本文不把开发日志当作功能介绍。
 
+## 0.85.5 — 凭据来源保护与审核边界修复
+
+- SVN 独立替换树在读取正文前检查复制来源及祖先来源，阻断从凭据路径复制到普通目标的内容；调用方扫描不再把带源码扩展名的目录当作文件读取。
+- Git 默认工作区范围合并暂存删除后重建的同路径记录，避免重复文件身份；自动上下文允许精确目录索引导入探测，但不扩大后代正文授权。
+- 暂存区审核复用已选变更作为上下文时，按相同 index 版本验证，不再误与未暂存工作区内容比较。
+- 综合证据接受同一快照、来源和版本侧连续或重叠读取的完整区间覆盖；仍拒绝有缺口、错来源或错版本的证据。
+- 超过 20 个窗口批次时保留全部可执行主窗口及此前成果，单独披露综合受限并保持未完整审核状态，不再使整轮异常退出。
+
 ## 0.85.4 — 审查正确性、失败恢复与全项目提速
 
 `v0.85.2` 因发布说明缺少导航链接、`v0.85.3` 因 Windows 临时目录短路径别名未被新增测试 mock 规范化而未通过 CI，均未发布 npm 或 GitHub Release。本版本补齐链接、规范化测试匹配路径并重新执行完整门禁，不覆盖旧冻结标签、不削弱拒绝断言。

package/RELEASE-NOTES.md

@@ -1,28 +1,21 @@
-# v0.85.4 发布说明
+# v0.85.5 发布说明
 
 使用方式见 [README](README.md)，详细变化见 [变更记录](CHANGELOG.md)，发布门禁见 [开发说明](DEVELOPMENT.md)。
 
-`v0.85.2` 因发布说明缺少导航链接、`v0.85.3` 因 Windows runner 临时目录短路径别名未被两个新增测试 mock 规范化而未通过 CI，均未发布 npm 或 GitHub Release。本版本保留旧冻结标签，补齐文档链接并修正测试路径匹配后重新执行完整门禁，不削弱原有拒绝断言。
+## 安全修复
 
-## 审查正确性与失败恢复
+SVN 独立替换树现在在读取正文前检查复制来源及复制祖先来源，阻断把凭据文件复制到普通目标后绕过来源过滤的路径。限定调用方扫描也会区分目录与普通文件，带 `.js` 等源码扩展名的目录不再令预览整体失败。
 
-- 模型流现在支持合法的完整文本块；完整块替换对应增量而不是重复拼接，文本和推理块继续受输出预算限制。真正重复或非法的 JSON 仍会拒绝，不靠截断响应通过。
-- 后台保存队列在一次失败后仍可保存终态；成功落盘后释放已结束任务的内存，最终保存失败也会唤醒等待者并保留诊断。
-- 修复模型拒绝值为 null/undefined 时丢失报告、非法分段参数导致死循环、SVN 历史正文大小低报影响累计预算等问题。
-- 补全进程终止的 SIGKILL 回退及清理、注册表网络异常重试、Git 根路径空格保留、锚点越界重定位、具名导入尾随逗号，以及 JavaScript/C#/C++ 导航边界。
+## 审核正确性与可用性
 
-## 全项目审核提速
+- Git 默认工作区审核合并暂存删除后重建的同路径记录，保持文件身份唯一。
+- 自动上下文的目录索引导入不再被误判为越界；仍只读取精确选中的候选正文，不授权目录下任意文件。
+- 已选暂存变更同时作为显式上下文时，复用 index 内容并按同侧版本验证，不把尚未暂存的正文混入变更侧。
+- 长编辑分次读取后，综合校验按同一快照、来源及版本侧计算连续覆盖；跨来源、有缺口或错误版本的读取仍不能作为有效证据。
+- 文件分成超过 20 个窗口批次时，主窗口继续审核并保留已有成果；综合受限单独披露，整体仍为部分完成，不再以异常丢失整轮报告。
 
-默认独立批次并发从 3 提高到 8，可用插件配置 projectConcurrency 在 1～8 内调整。保留主审、检索、证据复核及共享调用预算；遇到提供方并发限制或 429 时请调低。
+## 验证与边界
 
-相同 12 个批次、每次模拟模型延迟 40ms 的离线对照中，旧默认 3 路约 182ms，新默认 8 路约 92ms，约快 1.98 倍；调用数、完成数和结果顺序一致。这不是实际模型吞吐承诺，不减少必需调用或承诺 token 成本下降。
+本版为六项已复现问题添加回归测试，并执行完整本地测试及隔离安装验证；正式发布须通过同一冻结标签提交的 Ubuntu/Windows × Node.js 22/24 四组 CI、安装包验证和 npm 实际工件一致性门禁。具体执行结果以该标签 CI 和发布工作流为准。
 
-本次历史审核的 778 次模型调用中，263 次为额外 JSON 重生成。完整块兼容修复可避免对应场景的无谓纠正，但尚未证明它解释历史全部格式失败；真实提供方效果需要更新后复测。
-
-## 验证与安全边界
-
-本地完整测试 685 项：684 通过、1 项 Windows 平台跳过、0 失败、0 取消；另对新增流式、持久化及导航回归进行独立验证。正式发布必须经过冻结标签提交的 Linux/Windows × Node 22/24 四组 CI，以及隔离安装和注册表工件一致性门禁。
-
-文件及祖先身份检查是路径竞态检测加固，不是原子目录相对打开；不能保证抵御恶意进程精确交换并恢复路径。严格隔离场景应在不可并发修改的独立快照中审核。
-
-本版本不自动安装、升级或重启 DSH；需更新并重新加载插件后生效。
+本版本不保证模型始终输出合法 JSON，不扩大源码读取权限，不自动安装、升级或重启 DSH；更新并重新加载插件后生效。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@feng0531/dsh-code-review",
-  "version": "0.85.4",
+  "version": "0.85.5",
   "private": false,
   "license": "SEE LICENSE IN LICENSE",
   "publishConfig": { "access": "public", "registry": "https://registry.npmjs.org/" },

package/src/coverage-followup.mjs

@@ -5,7 +5,7 @@ export function coverageFollowup(report) {
   const items = report.files.map(file => {
     if (!file || typeof file.fileId !== 'string' || !file.fileId || ids.has(file.fileId) || typeof file.path !== 'string' || !file.path || paths.has(file.path) || !['completed', 'excluded', 'blocked', 'failed', 'cancelled', 'pending'].includes(file.status)) throw new Error('Invalid coverage file');
     ids.add(file.fileId); paths.add(file.path);
-    const incompleteSynthesis=(file.stageFailures??[]).length>0||file.synthesisStatus==='failed';
+    const incompleteSynthesis=(file.stageFailures??[]).length>0||!!file.synthesisStatus&&file.synthesisStatus!=='completed';
     const incompleteVerification = report.findings.some(f => f.fileId === file.fileId && ['pending', 'incomplete'].includes(f.verification?.status));
     return { fileId: file.fileId, path: file.path, status: file.status, incompleteVerification,incompleteSynthesis,
       followup: file.status !== 'excluded' && (file.status !== 'completed' || incompleteVerification || incompleteSynthesis) };

package/src/coverage-plan.mjs

@@ -21,7 +21,7 @@ export function buildCoveragePlan(input, options) {
     if (file.eligibility !== 'reviewable') return item;
     const prepared = preparePrimaryInput(input, file, options);
     const riskPlanning = options.enableRiskPlanning === true ? options.skipSmallRiskPlans === true ? riskPlanningDecision([file]) : {required:true,reason:'unconditional'} : {required:false,reason:'disabled'};
-    return { ...item, ...prepared.metadata, riskPlanning, status: prepared.budget.fits ? 'ready' : 'input-blocked', minimumCalls: prepared.budget.fits ? ((prepared.batches?.length ?? 1) * (rounds + (riskPlanning.required ? 1 : 0)) + (prepared.synthesis?.budget.fits ? rounds : 0)) : 0 };
+    return { ...item, ...prepared.metadata, riskPlanning, status: prepared.budget.fits ? 'ready' : 'input-blocked', minimumCalls: prepared.budget.fits ? ((prepared.batches?.length ?? 1) * (rounds + (riskPlanning.required ? 1 : 0)) + (prepared.synthesis?.status === 'ready' ? rounds : 0)) : 0 };
   });
   const local = options.enableBusinessGrouping === true ? localBusinessGroups(input.files) : null;
   const groups = [...(options.grouping?.groups ?? [])];

package/src/git.mjs

@@ -41,7 +41,13 @@ function parseChanges(bytes) {
 
 /** Keep only the entries the requested comparison actually observes. */
 function changesInScope(changes, scope) {
-  if (scope === 'worktree') return changes;
+  // Porcelain reports a staged deletion and its untracked recreation separately.
+  // Worktree compares HEAD to the current file once, retaining the tracked entry
+  // (and its baseline/selection authority), not a second untracked identity.
+  if (scope === 'worktree') {
+    const tracked = new Set(changes.filter(change => change.rawStatus !== '??').map(change => change.path));
+    return changes.filter(change => change.rawStatus !== '??' || !tracked.has(change.path));
+  }
   if (scope === 'staged') return changes.filter(change => change.rawStatus[0] !== ' ' && change.rawStatus[0] !== '?');
   return changes.filter(change => change.rawStatus[1] !== ' ' || change.rawStatus === '??');
 }

package/src/primary-input.mjs

@@ -167,8 +167,9 @@ export function preparePrimaryInput(input, file, { instructions, scope, grouping
 
   if (batches?.length > 1) {
     synthesis = prepareWindowSynthesis(batches.map(batch => batch.payload), measure);
-    metadata.synthesisStatus = synthesis.budget.fits ? 'ready' : 'input-blocked';
-    metadata.synthesisInputBytes = synthesis.budget.bytes;
+    metadata.synthesisStatus = synthesis.status === 'ready' ? 'ready' : 'blocked';
+    if (synthesis.reason) metadata.synthesisBlockedReason = synthesis.reason;
+    if (synthesis.budget) metadata.synthesisInputBytes = synthesis.budget.bytes;
   }
 
   const budget = measure(payload);

package/src/review.mjs

@@ -136,8 +136,8 @@ export async function reviewSnapshot(snapshot, model, options = {}) {
     const { changes, payload } = prepared;
     Object.assign(state, prepared.metadata);
     if (state.contextMode === 'approved-retrieval') report.limitations.push({ fileId: file.id, text: '窗口初始输入仅带上下文身份目录，全文须按需读取；不声明所有上下文已核查。' });
-    if (prepared.synthesis) state.synthesisStatus = prepared.synthesis.budget.fits ? 'pending' : 'blocked';
-    if (prepared.synthesis && !prepared.synthesis.budget.fits) report.limitations.push({ fileId: file.id, text: '跨窗口综合输入超限，未发送；单窗口完成不表示交互影响已检查' });
+    if (prepared.synthesis) state.synthesisStatus = prepared.synthesis.status === 'ready' ? 'pending' : 'blocked';
+    if (prepared.synthesis?.status === 'blocked') report.limitations.push({ fileId: file.id, text: (prepared.synthesis.reason === 'batch-limit' ? '跨窗口综合批次数超过上限（20），未发送；全部主窗口仍独立审查' : '跨窗口综合输入超限，未发送') + '；单窗口完成不表示交互影响已检查' });
     if (state.sourceMode === 'change-windows') state.windowCoverage = { completed: [], pending: [...state.windowIds] };
     if(changes.precision==='envelope')report.limitations.push({fileId:file.id,text:'精确差异不可得，按保守包络分段检查；包络包含可能未变的行，不证明精确编辑或回归因果。'});
     if (changes.status === 'limited') report.limitations.push({ fileId: file.id, text: '精确变更分析受限：' + changes.reason + '；仍按完整两侧内容审查' });
@@ -192,7 +192,7 @@ export async function reviewSnapshot(snapshot, model, options = {}) {
           }
         }
         const batches = [...(prepared.batches ?? [{ payload, windowIds: state.windowIds ?? [] }])];
-        if (prepared.synthesis?.budget.fits) batches.push({ payload: prepared.synthesis.payload, windowIds: [], synthesis: true });
+        if (prepared.synthesis?.status === 'ready') batches.push({ payload: prepared.synthesis.payload, windowIds: [], synthesis: true });
         const matchingInteractionPlans = interactionPlans.filter(p => input.files.filter(f => p.group.fileIds.includes(f.id)).at(-1)?.id === file.id);
         for (const interactionPlan of matchingInteractionPlans.filter(p => p.prepared.status === 'ready')) batches.push({ payload: interactionPlan.prepared.payload, windowIds: [], interaction: interactionPlan.group });
         const scheduled = Array.from({length:rounds},(_,i)=>batches.map(b=>({...b,round:i+1}))).flat();

package/src/snapshot.mjs

@@ -147,7 +147,13 @@ export async function captureSnapshot(cwd, options = {}) {
 
   // Re-capture verifies both selected sides while collecting explicit context.
   if ((await capture(cwd, indexOptions)).id !== snapshot.id) throw new Error('Snapshot changed while collecting context');
-  for (const item of context) if ((await readLocal(snapshot.root, item.path, indexOptions)).hash !== item.hash) throw new Error('Context changed during capture');
+  for (const item of context) {
+    // Overlapping staged context reuses the approved index side, not local bytes.
+    // The same-side re-capture above verifies its content and index identity,
+    // including when the working file differs or no longer exists.
+    if (type === 'git' && snapshot.changeScope === 'staged' && snapshot.files.some(file => file.path === item.path)) continue;
+    if ((await readLocal(snapshot.root, item.path, indexOptions)).hash !== item.hash) throw new Error('Context changed during capture');
+  }
   if (index && (await (type === 'git' ? gitContextIndex : svnContextIndex)(snapshot.root, indexOptions)).fingerprint !== index.fingerprint) throw new Error('Tracked context index changed during capture');
   const verifiedRules = await captureProjectRules(snapshot.root, rulePaths, { signal: options.signal, files: snapshot.files });
   if (JSON.stringify(verifiedRules) !== JSON.stringify(rules)) throw new Error('Rules changed during capture');

package/src/svn-caller-index.mjs

@@ -7,4 +7,11 @@ import {lstat} from 'node:fs/promises';
 import path from 'node:path';
 const canonical=v=>Array.isArray(v)?v.map(canonical):v&&typeof v==='object'?Object.fromEntries(Object.keys(v).sort().map(k=>[k,canonical(v[k])])):v;
 // Offline scoped status only, never svn list or repository-wide enumeration.
-export async function svnCallerIndex(root,scopes,options={}){const approved=validateCallerScopes(scopes),entries=[],chunks=[];let bytes=0;const seen=new Set();for(const scope of approved){const ancestors=await svnCallerAncestors(root,scope,options);chunks.push(ancestors.fingerprint);let cursor=root;for(const part of scope.split('/')){cursor=path.join(cursor,part);const s=await lstat(cursor);if(s.isSymbolicLink()||!s.isDirectory())throw new Error('SVN caller scope requires real directories');}const raw=await checked('svn',['--non-interactive','status','--xml','--verbose','--ignore-externals','--',scope+'@'],{...options,cwd:root,maxBytes:64*1024});bytes+=raw.length;if(bytes>64*1024)throw new Error('SVN caller index byte limit exceeded');const found=(xml(raw).status?.target??[]).flatMap(t=>t.entry??[]);for(const e of found){const name=relativePath(e['@_path'].replaceAll('\\','/'));if(name!==scope&&!name.startsWith(scope+'/'))throw new Error('SVN caller index escaped scope');if(seen.has(name))throw new Error('Duplicate SVN caller index path');seen.add(name);if(seen.size>256)throw new Error('SVN caller index path limit exceeded');entries.push({...e,'@_path':name});}if(!found.some(e=>e['@_path'].replaceAll('\\','/')===scope))throw new Error('SVN caller scope missing from status');chunks.push(scope);}entries.sort((a,b)=>a['@_path']<b['@_path']?-1:1);const unsafe=s=>!s||s['@_item']!=='normal'||!['none','normal'].includes(s['@_props'])||['@_switched','@_copied','@_file-external','@_tree-conflicted'].some(k=>s[k]==='true');for(const scope of approved){const e=entries.find(e=>e['@_path']===scope);if(unsafe(e?.['wc-status']))throw new Error('Unsafe SVN caller directory state');}return {scopes:approved,paths:entries.filter(e=>!approved.includes(e['@_path'])&&!unsafe(e['wc-status'])).map(e=>e['@_path']),entries,records:entries.length,bytes,fingerprint:hash(JSON.stringify(canonical({scopes:chunks,entries})))};}
+// Status XML has no node kind: bind local kinds and omit directories from body candidates.
+// Links and other unsafe non-directory nodes still reach the strict source validator.
+export async function svnCallerIndex(root,scopes,options={}){const approved=validateCallerScopes(scopes),entries=[],chunks=[];let bytes=0;const seen=new Set();for(const scope of approved){const ancestors=await svnCallerAncestors(root,scope,options);chunks.push(ancestors.fingerprint);let cursor=root;for(const part of scope.split('/')){cursor=path.join(cursor,part);const s=await lstat(cursor);if(s.isSymbolicLink()||!s.isDirectory())throw new Error('SVN caller scope requires real directories');}
+ const raw=await checked('svn',['--non-interactive','status','--xml','--verbose','--ignore-externals','--',scope+'@'],{...options,cwd:root,maxBytes:64*1024});bytes+=raw.length;if(bytes>64*1024)throw new Error('SVN caller index byte limit exceeded');const found=(xml(raw).status?.target??[]).flatMap(t=>t.entry??[]);
+ for(const e of found){const name=relativePath(e['@_path'].replaceAll('\\','/'));if(name!==scope&&!name.startsWith(scope+'/'))throw new Error('SVN caller index escaped scope');if(seen.has(name))throw new Error('Duplicate SVN caller index path');seen.add(name);if(seen.size>256)throw new Error('SVN caller index path limit exceeded');entries.push({...e,'@_path':name});}if(!found.some(e=>e['@_path'].replaceAll('\\','/')===scope))throw new Error('SVN caller scope missing from status');chunks.push(scope);}
+ entries.sort((a,b)=>a['@_path']<b['@_path']?-1:1);const unsafe=s=>!s||s['@_item']!=='normal'||!['none','normal'].includes(s['@_props'])||['@_switched','@_copied','@_file-external','@_tree-conflicted'].some(k=>s[k]==='true');for(const scope of approved){const e=entries.find(e=>e['@_path']===scope);if(unsafe(e?.['wc-status']))throw new Error('Unsafe SVN caller directory state');}
+ const nodes=[];for(const e of entries.filter(e=>!approved.includes(e['@_path'])&&!unsafe(e['wc-status']))){options.signal?.throwIfAborted();const stat=await lstat(path.join(root,e['@_path']));nodes.push({path:e['@_path'],kind:stat.isDirectory()?'dir':stat.isSymbolicLink()?'link':stat.isFile()?'file':'other'});}
+ return {scopes:approved,paths:nodes.filter(n=>n.kind!=='dir').map(n=>n.path),entries,records:entries.length,bytes,fingerprint:hash(JSON.stringify(canonical({scopes:chunks,entries,nodes})))};}

package/src/svn-replacement-tree.mjs

@@ -6,11 +6,19 @@ import {readReplacementSource} from './svn-replacement-source.mjs';
 import {svnWorkingIdentity} from './svn-working-identity.mjs';
 import {hash,decode,relativePath,readLocal} from './content.mjs';
 import {isSecretPath} from './secret-path.mjs';
+import {canonicalSvnUrl} from './svn-url.mjs';
 import {svnKeywordMapping} from './svn-keyword-mapping.mjs';
 import {svnEolMapping} from './svn-eol-mapping.mjs';
 import {svnSpecialData} from './svn-special.mjs';
 import {readWorkingSvnSpecial} from './svn-working-special.mjs';
 
+// Copy provenance is not permission to capture credential bytes at a safe target name.
+function assertCopySource(source,suffix=''){
+ if(source===null||source===undefined)return;
+ const url=canonicalSvnUrl(source),name=decodeURIComponent(new URL(url).pathname);
+ if(isSecretPath(name)||isSecretPath(name+(suffix?'/'+suffix:'')))throw new Error('Credential replacement copy source excluded');
+}
+
 // Fixed repository tree versus explicitly selected working replacement tree.
 // Every descendant is a separate source/coverage unit, never an implicit approval.
 export async function captureReplacementTree(root,rootInfo,name,options={}){
@@ -26,13 +34,14 @@ export async function captureReplacementTree(root,rootInfo,name,options={}){
  const ancestors=async p=>{
   const parts=p.split('/').slice(0,-1),names=['.',...parts.map((_,i)=>parts.slice(0,i+1).join('/'))],records=[];
   for(const name of names){
-   if(!verifying&&ancestorReceipts.has(name)){records.push(ancestorReceipts.get(name));continue;}
+   if(!verifying&&ancestorReceipts.has(name)){const record=ancestorReceipts.get(name);assertCopySource(record.copyFromUrl,name==='.'?p:p.slice(name.length+1));records.push(record);continue;}
    const stat=await lstat(name==='.'?root:path.join(root,name));if(stat.isSymbolicLink()||!stat.isDirectory())throw new Error('Replacement ancestor is not a real directory');
    const node=xml(await run(['info','--xml','--',name+'@'])).info?.entry?.[0],identity=svnWorkingIdentity(node,rootInfo,name);
    const entries=xml(await run(['status','--xml','--verbose','--ignore-externals','--depth','empty','--',name+'@'])).status?.target?.flatMap(t=>t.entry??[])??[],s=entries[0]?.['wc-status'];
    if(entries.length!==1||!s||!['normal','modified','added','replaced'].includes(s['@_item'])||s['@_props']==='conflicted'||['@_switched','@_file-external','@_tree-conflicted'].some(k=>s[k]==='true'))throw new Error('Unsafe replacement ancestor state');
+   assertCopySource(identity.source.copyFromUrl,name==='.'?p:p.slice(name.length+1));
    const properties=await props(name);if(Object.hasOwn(properties,'svn:externals')||Object.hasOwn(properties,'svn:special'))throw new Error('Unsafe replacement ancestor properties');
-   const record={path:name,fingerprint:hash(JSON.stringify([identity.fingerprint,properties]))};
+   const record={path:name,copyFromUrl:identity.source.copyFromUrl,fingerprint:hash(JSON.stringify([identity.fingerprint,properties]))};
    if(verifying){if(ancestorReceipts.get(name)?.fingerprint!==record.fingerprint)throw new Error('Replacement ancestor changed during capture');}else ancestorReceipts.set(name,record);records.push(record);
   }return records;
  };
@@ -40,6 +49,7 @@ export async function captureReplacementTree(root,rootInfo,name,options={}){
   if(['@_switched','@_file-external','@_tree-conflicted'].some(k=>s[k]==='true')||!['normal','modified','added','replaced'].includes(s['@_item'])||s['@_props']==='conflicted')throw new Error('Unsafe replacement descendant state');
   const ancestorEvidence=await ancestors(p);
   const node=xml(await run(['info','--xml','--',p+'@'])).info?.entry?.[0],identity=svnWorkingIdentity(node,rootInfo,p);if(!['file','dir'].includes(node['@_kind']))throw new Error('Unsupported replacement descendant kind');
+  assertCopySource(identity.source.copyFromUrl);
   if(node['@_kind']==='dir'&&identity.source.depth!=='infinity')throw new Error('Sparse replacement target cannot prove descendant inventory');
   const properties=await props(p);if(Object.hasOwn(properties,'svn:externals'))throw new Error('Replacement externals require independent scope');
   const content=node['@_kind']==='dir'?decode(Buffer.alloc(0)):Object.hasOwn(properties,'svn:special')?(await readWorkingSvnSpecial(root,p,properties,options)).content:await readLocal(root,p,options);

package/src/synthesis-evidence.mjs

@@ -1,15 +1,31 @@
 // Audit only successful reads from this synthesis generation, never prior batches/searches.
 export function validateSynthesisReads(findings, changes, reads, catalog, fileId, snapshotId) {
  const sources = new Map(catalog.map(s => [s.id, s]));
+ const intervals = new Map();
+ for (const read of reads) {
+  const source = sources.get(read.sourceId);
+  if (read.kind !== 'read' || read.snapshotId !== snapshotId || source?.fileId !== fileId || !['old', 'new'].includes(source.side)) continue;
+  if (!Number.isSafeInteger(read.start) || read.start < 1 || !Number.isSafeInteger(read.count) || read.count < 1 || !Number.isSafeInteger(read.start + read.count)) continue;
+  const ranges = intervals.get(read.sourceId) ?? [];
+  ranges.push({ start: read.start, end: read.start + read.count });
+  intervals.set(read.sourceId, ranges);
+ }
+ // Never join different source identities, even when their file and side match.
+ for (const [id, ranges] of intervals) {
+  const merged = [];
+  for (const range of ranges.sort((a, b) => a.start - b.start)) {
+   const last = merged.at(-1);
+   if (last && range.start <= last.end) last.end = Math.max(last.end, range.end);
+   else merged.push({ ...range });
+  }
+  intervals.set(id, merged);
+ }
  for (const finding of findings) {
   for (const id of finding.attribution.editIds) {
    const edit = changes.edits.find(e => e.id === id);
    const covered = ['old', 'new'].some(side => {
     const range = edit?.[side];
-    return range?.count > 0 && reads.some(r => {
-     const source = sources.get(r.sourceId);
-     return r.kind === 'read' && r.snapshotId === snapshotId && source?.fileId === fileId && source.side === side && r.start <= range.start && r.start + r.count >= range.start + range.count;
-    });
+    return range?.count > 0 && [...intervals].some(([sourceId, ranges]) => sources.get(sourceId).side === side && ranges.some(r => r.start <= range.start && r.end >= range.start + range.count));
    });
    if (!covered) throw new Error('Synthesis referenced edit was not read in this generation');
   }

package/src/tracked-context.mjs

@@ -16,7 +16,10 @@ export async function gitContextIndex(root, options = {}) {
  let bytes;
  if(options.probePaths !== undefined){
   if(!Array.isArray(options.probePaths)||options.probePaths.length>512||new Set(options.probePaths).size!==options.probePaths.length)throw new Error('Invalid Git context probes');
-  const chunks=[];let totalBytes=0;for(const name of [...options.probePaths].sort()){relativePath(name);const chunk=await checked('git',['--no-optional-locks','--literal-pathspecs','-c','core.fsmonitor=false','-c','core.untrackedCache=false','ls-files','--stage','-z','--',name],{...options,cwd:root,maxBytes:64*1024});const records=new TextDecoder('utf-8',{fatal:true}).decode(chunk).split('\0').filter(Boolean);for(const record of records)if(record.slice(record.indexOf('\t')+1)!==name)throw new Error('Git context probe expanded beyond literal path');totalBytes+=chunk.length;if(totalBytes>1024*1024)throw new Error('Tracked context index limit exceeded');chunks.push(chunk);}
+  // A literal pathspec alone still recursively matches a directory. Exclude
+  // its literal trailing-slash subtree before output, so directory probes are
+  // exact misses without enumerating or authorizing unrequested descendants.
+  const chunks=[];let totalBytes=0;for(const name of [...options.probePaths].sort()){relativePath(name);const chunk=await checked('git',['--no-optional-locks','-c','core.fsmonitor=false','-c','core.untrackedCache=false','ls-files','--stage','-z','--',':(top,literal)'+name,':(top,literal,exclude)'+name+'/'],{...options,cwd:root,maxBytes:64*1024});const records=new TextDecoder('utf-8',{fatal:true}).decode(chunk).split('\0').filter(Boolean);for(const record of records)if(record.slice(record.indexOf('\t')+1)!==name)throw new Error('Git context probe expanded beyond literal path');totalBytes+=chunk.length;if(totalBytes>1024*1024)throw new Error('Tracked context index limit exceeded');chunks.push(chunk);}
   bytes=Buffer.concat(chunks);if(bytes.length>1024*1024)throw new Error('Tracked context index limit exceeded');
  }else bytes = await checked('git', ['--no-optional-locks','-c','core.fsmonitor=false','-c','core.untrackedCache=false','ls-files','--stage','-z'], {...options,cwd:root,maxBytes:1024*1024});
  const records = new TextDecoder('utf-8',{fatal:true}).decode(bytes).split('\0').filter(Boolean);

package/src/window-synthesis.mjs

@@ -1,6 +1,8 @@
 // Pure approved-data preparation, no reads or sends. No source text is silently clipped.
 export function prepareWindowSynthesis(batchPayloads, measure) {
-  if (!Array.isArray(batchPayloads) || batchPayloads.length < 2 || batchPayloads.length > 20) throw new Error('Invalid synthesis batches');
+  if (!Array.isArray(batchPayloads) || batchPayloads.length < 2) throw new Error('Invalid synthesis batches');
+  // A synthesis scheduling limit must not discard any independently reviewable batch.
+  if (batchPayloads.length > 20) return { status: 'blocked', reason: 'batch-limit', batchCount: batchPayloads.length };
   const batches = batchPayloads.map(payload => JSON.parse(payload));
   const base = batches[0];
   const windows = batches.flatMap(batch => batch.windows);
@@ -8,5 +10,5 @@ export function prepareWindowSynthesis(batchPayloads, measure) {
   const windowIndex = windows.map(w => ({ id: w.id, editIds: w.editIds, old: { start: w.old.start, count: w.old.count }, new: { start: w.new.start, count: w.new.count } }));
   const payload = JSON.stringify({ ...base, windows: undefined, windowIndex, sourceMode: 'window-synthesis', windowNotice: '本请求综合检查所有窗口的共同影响，不重报单窗口问题；windowIndex仅为原始定位目录，不含源码，必须在批准catalog中读取相关旧新片段，不能仅凭目录认定缺陷。输出仍仅主file的发现，归因至少关联两个不同窗口的编辑；证据不足明确限制。' });
   const budget = measure(payload);
-  return { payload, windows, windowIds: windows.map(w => w.id), budget };
+  return { status: budget.fits ? 'ready' : 'blocked', ...(budget.fits ? {} : { reason: 'input-budget' }), payload, windows, windowIds: windows.map(w => w.id), budget };
 }