amos 8b7616bfd8 Batch 22 (20261027-C-hwe-hard-22): add 63 calibrated HWE hard-design tasks
From the 101 lk/man/pth code-gen sources: 63 tasks (C11 36, C++17 27,
11 canonical families), 38 skipped for scope or duplication. Each task
has at least one independent contract-edge defect, boundary types
rotate, and public docs state trap clauses as plain contract prose.
Claude reviewed 63/63 USABLE against current digests. Difficulty is
unmeasured.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-27 16:14:57 +08:00

117 lines
14 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# 任务规格:concur-batch-commit-tail
## 1. 身份与来源
| 字段 | 填写 |
| --- | --- |
| task-id / 任务版本 | concur-batch-commit-tail / 1 |
| 作者 / 业务评审人 / 独立评审人 | batch22 pth-cond-group-commit 子Agent / root待审 / root待审 |
| 审核状态 | CALIBRATED;难度UNMEASURED |
| 原项目与冻结版本 | HWE本地批次20261027-C-hwe-hard-22只读材料;原始指纹见本题root-source-review.json |
| 来源路线 | C 素材写题;原创修复工程,不声称历史现场故障 |
| 来源ID | hwe/pth-cond-group-commit |
| 来源与授权 | 用户2026-10-27批次BRIEF;metadata visibility=public;只读metadata、prompt与reference,未读checks或执行来源脚本 |
| 隐私处理 | 无真实用户内容或凭据 |
| 替身测试 | software:显式调用表达调度账本,无真实线程、时间、硬件、特权或外部服务 |
| split分组键 | pth-cond-group-commit |
| 来源组 / 根因族 | HWE cond-group-commit / linux-concurrency |
| 与已有任务关系 | 原创的批量发布与尾批模型;来源与ID初筛由root核对,无共用其他新题业务骨架 |
| 验证范围 | C++17公开账本、backend、helper的确定性行为与文本域 |
| 未验证范围 | 真实线程调度、pthread原语实现、硬件、模型能力及真实Agent轨迹 |
设计动机:把原料的“攒满N再提交”改写为可以逐条重放与复算的生产者/提交者软件账本。队列由Engine维护,发布由独立Journal完成,policy决定谓词可选数量,text独立解析配置。原料代码不复制进工程。
缺陷一:policy.cpp的take_count在closing且queued<width时返回0,Engine::pump因此留住最后未满批记录并一直返回WAIT,status.ready也同源错误。指定精确合同边界类型为“最后不满批次的精确收尾”;公开依据R08,smoke不覆盖这一边界。
缺陷二:journal.cpp的新ticket在检查故障注入前已经插入发布记录。Engine收到-11仍保留队列与ticket,而Journal的published计数提前变化。涉及Journal提交与Engine重试两个模块,报告来自console;公开smoke可独立复现。
只修tail保留提前发布,only_fix_tail在smoke与template一致;只修publish保留尾批等待,only_fix_publish在private-tail与template一致。二者独立对照保存在本题factory/selftest。公开文档按2026-10-27 plain-docs要求使用普通条款,不标出陷阱或给边界专门举例。
## 2. 初态与环境
- 初态资产:template的group.hpp、policy.cpp、text.cpp、journal.cpp、engine.cpp、console.cpp、Makefile、smoke.txt、README.md、docs/contract.md及docs/console.md。所有源码与输入随题公开;没有遗漏的依赖文件。
- 公开资产:完整工程、R01–R17合同、命令/结果字段、公开smoke。候选只交付五个声明文件。template不含答案、隐藏输入、spec、历史或构建产物。
- 受保护资产:check.py、gold及fixture、spec与校准证据。verify.py逐字复制rollout/bos_verify.py。
- 环境族:固定trace6 Linux runtime,C++17及Python 3;离线使用g++和标准库,无额外依赖或服务。
- 镜像引用:ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1。实际image ID由正式receipt记录。
- 资源:正式评分512MiB/1CPU;check内编译35秒、观察20秒;平台总评分120秒。执行超时只防挂死。
- 本题C++/头文件独有非空无注释行720,按template单份统计,包含491行可提交实现和229行独立观察控制台。fixture复制不计入。各文件细分与去注释算法见本题report/selfcheck记录。
- 外部依赖:无。初态行为见本题factory/selftest内template-public-smoke.stdout和template-private-tail.stdout;不将本地受信执行当不可信隔离证据。
## 3. 公开题面与接口
题面为中文现场报告式描述,见prompt.md:操作返回可重试而报告已发布记录先变化;make smoke复现;要求满足docs/contract.md全部条款并提交五个声明文件。没有点名根因函数/文件/类别。
公开构建argv为["make"],执行argv为["./group-console"]并从stdin读取命令。成功运行退出0,非法控制台命令退出2。stdout是JSON数组原始观察;语义按整数比较,可容许JSON空白差异,bool不视为整数。stderr仅供诊断,不限制大小或字节编码。完整API参数、返回、副作用、输出引用、顺序和边界见R01–R17;console字段在docs/console.md。
## 4. 判据检查项 ↔ 依据与合法解
| 判据检查项 | 公开依据 | 输入/场景 | 可观察结果与方法 | 公开/受保护 | 超时 |
| --- | --- | --- | --- | --- | --- |
| R01 Independent Engines/Journals and unchanged inputs | template/docs/contract.md:3 | Interleaved objects and every text/batch helper | U alternation and independent raw byte/item digests; exact integer fields | private inputs / public rule | 20s process |
| R02 Return values and WAIT/DONE observations | template/docs/contract.md:5 | Inactive, malformed, range, full, conflict, retry, idle and done | Typed integer array comparison; exact integer fields | private inputs / public rule | 20s process |
| R03 All out fields and state retained on errors | template/docs/contract.md:7 | pump/peek/read/parse/config/commit failures | Nonces 0, intermediate, uint64 max; before/after status; exact integer fields | private inputs / public rule | 20s process |
| R04 Configuration domains, reconfiguration and reset | template/docs/contract.md:9 | width 0/1/16/17/MAX, capacities below/equal/64/65 | V direct helper and C/Z sequences; exact integer fields | private inputs / public rule | 20s process |
| R05 Inactive checks precede remaining validation | template/docs/contract.md:11 | unconfigured enqueue/fail/pump/peek/read/close | Malformed inactive operations plus status; exact integer fields | private inputs / public rule | 20s process |
| R06 Ordered queue, lifetime ids and admission checks | template/docs/contract.md:13 | full queues, invalid fields, duplicates before/after pump and close | Q and receipts plus status; exact integer fields | private inputs / public rule | 20s process |
| R07 Closing generation and replaceable failure setting | template/docs/contract.md:15 | empty/full queues, repeated close, invalid replacement | X/X, F 0..9/MAX, P then S; exact integer fields | private inputs / public rule | 20s process |
| R08 Batch width, final partial group and one-group pump | template/docs/contract.md:17 | width 1/2/3/4/7/16; 0, width-1,width,width+1,2width±1,64 Items | Independent K matrix; engine prefix and completion observations; exact integer fields | private inputs / public rule | 20s process |
| R09 Atomic failure and idempotent ticket publication | template/docs/contract.md:19 | failure 0/2/8, repeated/conflicting/reordered tickets and retries | JC/JS/JR direct backend; Engine P/S/R; exact integer fields | private inputs / public rule | 20s process |
| R10 Batch validation, value domain and duplicate ids | template/docs/contract.md:21 | empty/17 Items, invalid ticket/id/value and mixed-error batches | B independently and JC while failures are pending; exact integer fields | private inputs / public rule | 20s process |
| R11 Publication indexing and queued indexing | template/docs/contract.md:23 | first/last/outside/MAX indexes, reads after close/reset | JR/R/Q with multiple initial output patterns; exact integer fields | private inputs / public rule | 20s process |
| R12 Summary, same_items and modulo total | template/docs/contract.md:25 | empty, 1,16,17,64 Items, full uint64 fields, order/value/id changes | D/E independently of Journal and Engine; exact integer fields | private inputs / public rule | 20s process |
| R13 Item digest byte layout and empty digest | template/docs/contract.md:27 | asymmetric ids/values, maximum words, repeated records | H direct helper and independent Python FNV arithmetic; exact integer fields | private inputs / public rule | 20s process |
| R14 Full status, counters and instance isolation | template/docs/contract.md:29 | initial, admitted, attempted, committed, failed, reused and reset | S/JS after commands; exact integer fields | private inputs / public rule | 20s process |
| R15 Decimal grammar, lexical-before-range and output retention | template/docs/contract.md:31 | empty/sign/SP/TAB/CR/LF/VT/FF/NUL/non-ASCII, overflow plus trailing invalid byte | N direct helper with original-input digests; exact integer fields | private inputs / public rule | 20s process |
| R16 Complete config grammar and numeric validation | template/docs/contract.md:33 | ordering, separators, whitespace, mixed errors, overflow, complete-line trailing bytes | T direct helper with original-input digests; exact integer fields | private inputs / public rule | 20s process |
| R17 Unbounded leading-zero strings and raw diagnostics | template/docs/contract.md:35 | Whole strings transported as hex, strict stdout UTF-8, stderr replacement decode | 20,000 leading zeros and overflowing numeric tokens; exact integer fields | private inputs / public rule | 20s process |
| 控制台观察格式与输入内容保留 | template/docs/console.md:3,5,7,9,11,13,15,17,19,21 | 所有命令、同进程U双实例、独立helper与Journal操作 | 固定原始观察器只输出实值;Python独立建模并严格整数语义比较 | public protocol / private script | 20s |
| 编译与交付边界 | prompt.md:5;template/README.md:7 | 五个声明文件重新构建 | 私有固定console与候选源码编译,其他文件不收集 | public | 35s |
- Oracle以Python重新实现公开数学与状态合同,不执行gold获取答案。固定seed 0xBA7C22用于700步交错回归。每次生成同样的完整输入,所有输出字段都比较。
- 所有候选文件先一次性读取,mkdtemp(dir=Path.cwd())下写好全部源码,再启动首个child;之后check不遍历、读、写或清理共享scratch。平台外层root最终统一回收。无候选同进程导入、ctypes或shell执行;harness无断言或预期结果。
- stdout以bytes捕获后按UTF-8/JSON解码;stderr只诊断时以replace解码。JSON递归同类型比较,拒绝bool冒充整数。原始observer独立计算输入指纹,不借用被测helper计算其输入是否改变。
- 合法替代解valid_cow使用deque队列、unordered_set接受索引、Journal整份新状态写时复制后swap,选择数量采用整除计算;与gold的vector/线性搜索/原地append不同。
- 拒绝硬编码公开批宽、忽略失败设置、输出PASS替代行为等捷径;不要求候选使用特定调用路径。每个公开helper均通过直接原始观察验证,Journal也独立于Engine验证。cheat_read_grader仅核对判据隔离。
- grader的输出由统一verify包装器提供一行JSON;PASS/FAIL检查数1,退出0/1;缺环境或执行故障另按平台合同,不计能力失败。
- 隐藏输入、答案及历史候选不提供给Agent;没有发布到Salty的历史答案或任何外挂检索依赖。
## 5. 四类样例与校准
| 样例 | 预期 | 检验目的 |
| --- | --- | --- |
| gold | PASS | 满足全部公开合同 |
| template | FAIL | 两处独立缺陷初态 |
| only_fix_tail | FAIL | 仅修尾批,保留失败提前发布 |
| only_fix_publish | FAIL | 仅修失败发布,保留尾批等待 |
| noop | FAIL | configure修改状态后返回未配置,破坏核心功能 |
| cheat | FAIL | 按公开smoke的固定批宽4处理所有参数 |
| mask_symptom_retry | FAIL | 忽略失败注入掩盖公开症状 |
| valid_cow | PASS | 不同队列、索引与事务提交结构 |
| cheat_read_grader | FAIL(Docker) | 判据隔离;本地可读而PASS为合同允许例外 |
现成两门均每类5次,共45次;不降低每类次数,不改平台。最终回执及任务摘要由本题report.json记录,不手写或修改回执。若本地reader PASS导致总门FAIL,保留真实原件并注明允许例外,以Docker整体PASS为正式校准结果。最终校准前删除template构建产物并冻结任务摘要。
## 6. 真实试采与证据
- 本批不用Qwen或其他额外模型;未做真实Agent采样,没有原生轨迹、broker对账、模型成功率或难度分池结论。
- 任务状态只能为CALIBRATED/UNMEASURED;不能宣称HARD、ADMITTED或真实调度/硬件通过。
- 本次唯一运行证据在runs/20261027-C-hwe-hard-22/pth-cond-group-commit。每次attempt唯一,不覆盖历史;环境异常立即停止本题精确PID并报root。
- 容器由平台逐个管理并核对cleanup;不清理其他作业。没有本题专属网络、外部服务或凭据。尚未封存或上传Salty,收尾由root统一安排。
## 7. 评审与准入
- [x] 来源、公开机制、软件替身、完整初态与交付边界已核对。
- [x] 全部17条公开合同以及固定observer输出均建立依据映射;未为边界条款加提示。
- [x] 正例、结构不同正例、定向修复负例和症状遮盖样例齐全;本题预检与缺陷独立对照保留真实输出。
- [x] 最终双门回执摘要一致且Docker全部45次符合预期、cleanup confirmed;本地reader PASS例外使原始整体FAIL,原件保留。
- [ ] root只读独审完成。
- [ ] 真实采样和难度测量(本批明确不做)。
作者自检:2026-10-27;未发现需要用户补充的材料或环境依赖。最终状态与根审结果以report/review为准,不在此宣称已准入。任务包发生变化后须重新两门校准。
最终校准记录(2026-10-27):任务摘要 `7e997b5270fd31dd896fa191e711732d5a1ae5d37d3dc767b6def2a5795ee0dd`。本地原始FAIL(仅许可reader例外),正式Docker PASS,各类5次。
- local: `runs/20261027-C-hwe-hard-22/pth-cond-group-commit/local-final-2c39d6b931274b74b2724c3db5ab028f/calibration-636db7df601f49e586492faab646363e`;cleanup confirmed。
- docker: `runs/20261027-C-hwe-hard-22/pth-cond-group-commit/docker-final-db4f5b3fd35d4d8495933d4902f31824/calibration-02955503f71e4b67a7fe62cf97946196`;cleanup confirmed。