dsh-vibe-math 2.8.1 → 2.8.3

package/README.en.md

@@ -9,7 +9,7 @@ English | [中文](README.md)
 > A set of **agent presets** running inside **DeepSeek Harness** (`vibe-math-v2` / `vibe-math-v3` / `vibe-math-v4` / `vibe-math-v5`),
 > which use multi-agent collaboration to automatically solve mathematical problems and perform multi-agent cross-verification of the conclusions. All four presets share the foundational capabilities of "**checkpoint resume**,
 > **mid-run manual intervention**, **progress reporting**, and **natural-language driving**", but adopt four generations of different solving architectures:
-> **💡 All four architectures are peers** — `vibe-math-v2` and `vibe-math-v3` are the **classic** architectures (mature, usable, actively maintained); `vibe-math-v4` is the "resident self-organizing collaborative research" architecture and `vibe-math-v5` is the latest "institute system" (both labelled experimental); choose according to your actual needs (see "How to choose" below).
+> **💡 v4/v5 first** — `vibe-math-v4` (the "resident self-organizing collaborative research" architecture) and `vibe-math-v5` (the latest "institute system") are the **current focus of maintenance**; `vibe-math-v2` and `vibe-math-v3` are the **classic / early** architectures, now **architecturally frozen** and receiving **compatibility-only** maintenance (new capabilities land in v4/v5). All four still install and work; see [`docs/maintenance-policy.md`](docs/maintenance-policy.md). Choose according to your actual needs (see "How to choose" below).
 >
 > - **`vibe-math-v2` (probability-driven · JSON data layer) · classic**: `qs.json` problem list + `Propos/` proposition library + probability-driven scheduling + code heuristic scheduling;
 > - **`vibe-math-v3` (third generation · paper-style md + planner agent + method library) · classic**: all knowledge is stored and extended in **Markdown paper/research-report form** (`Problems/` problem list + dependencies + source motivation, `Progress/` research log, `Propos/` proposition library, `Methods/` general theory invention library, `Verified/` absolutely trustworthy); before scheduling, the **planner agent** autonomously draws up a plan for the next N steps; theories/frameworks/tools/methods/ideas invented during solving are distilled by the **Method Keeper** into a reusable method system (as in inventing group theory or functional analysis).

package/README.md

@@ -9,7 +9,7 @@
 > 运行在 **DeepSeek Harness** 内的一组 **agent preset**（`vibe-math-v2` / `vibe-math-v3` / `vibe-math-v4` / `vibe-math-v5`），
 > 用多代理协作自动求解数学问题并对结论做多代理交叉验证。四个预设共享「**断点续跑**、
 > **中途人工干预**、**进度汇报**、**自然语言驱动**」底座能力，但采用四代不同的求解架构：
-> **💡 四个架构同级**——`vibe-math-v2` 与 `vibe-math-v3` 是**经典**架构（成熟可用、持续维护），`vibe-math-v4` 是「常驻自组织合作研究」架构、`vibe-math-v5` 是最新的「研究所体系」（两者均标注为实验性）；根据你的实际需求自行选择（详见下方「怎么选」）。
+> **💡 v4/v5 优先**——`vibe-math-v4`（「常驻自组织合作研究」架构）与 `vibe-math-v5`（最新的「研究所体系」）是**当前主要维护方向**；`vibe-math-v2` 与 `vibe-math-v3` 是**经典/早期架构**，已**成熟定型**，今后只做**适配兼容**维护（新能力只进 v4/v5）。四套都随包安装、可正常使用；详见 [`docs/maintenance-policy.md`](docs/maintenance-policy.md)。根据你的实际需求自行选择（详见下方「怎么选」）。
 >
 > - **`vibe-math-v2`（概率驱动 · JSON 数据层）· 经典**：`qs.json` 问题清单 + `Propos/` 命题库 + 概率驱动调度 + 代码启发式调度；
 > - **`vibe-math-v3`（第三代 · 论文式 md + 规划代理 + 方法库）· 经典**：全部知识以 **Markdown 论文/研究报告式** 存储与续写（`Problems/` 问题清单+依赖+来源动机、`Progress/` 研究日志、`Propos/` 命题库、`Methods/` 通用理论发明库、`Verified/` 绝对可信）；调度前由**规划代理**自主制定接下来 N 步计划；解决过程中发明的理论/框架/工具/方法/思想由 **Method Keeper** 沉淀为可复用方法体系（如发明群论、泛函分析那样）。

package/docs/AUDIT-CHECKLIST.md

@@ -548,6 +548,8 @@ DSH 的 **agent preset 交付方式**在 0.1.6 → 0.1.7 之间换过一次，
 | **发布前检查（可执行化）** | 版本**三处一致**（`package.json`／lock 顶层／`packages[""]`）、当前版本的中英两份发布说明存在且**章节顺序符合模板**、两份都登记进 `files[]`、`compatNote` 提到当前版本、tarball **声明了但缺失 = 0**、开发专用文件（`.gitattributes`）**不随包**、包内**无 CRLF** | **随包** `node scripts/release-check.mjs`（对 2.8.0 实跑 `ALL CHECKS PASSED`，扫 179 个文本文件）＋ `--self-test`（**15** 条）＋ `node tests/release-check.mutants.mjs`（**4/4** 具名红：分别去掉版本一致性／章节数量／章节顺序／CRLF 扫描 ⇒ 对应自检**按名**变红）。发布后加 `--registry` 比对线上 `dist.shasum` 与本地 tarball |
 | **v4 交付物必须说明编译结果** | 论文正文在**编译之前**组装，因此"局限与未决"里那两条依赖 `ps.compile` 的 bullet 只在**重入**时才有值 ⇒ 交付的 `paper.md`／`paper.tex` 必须**各自写明**"未检测到引擎（只产出 tex+md）"或"PDF 编译失败"，且**恰好一次**（finalize 可重入 ⇒ 幂等；`.tex` 以注释插在 `\end{document}` 之前）| **随包** `node tests/v4-final-paper.test.mjs`（`★ [deliverable/v4] … (once each; md=1 tex=1)`）＋ `node tests/v4-final-paper.mutants.mjs`（**3/3** 具名红：把中文说明置空 ⇒ 该断言按名红，实测 `md=0 tex=0`）|
 
+| **维护方针（文档）** | `vibe-math-v4`／`vibe-math-v5` 是**主要维护方向**（新能力、新守卫、提示词强化、重构只落这两套）；`vibe-math-v2`／`vibe-math-v3` 已**成熟定型**，只接受**适配兼容**改动：宿主适配、共享模块的正确性/安全修复、行为不变的回归修复。共享模块（`vibe-math-v{2,3,4,5}/math-computation.js` 等）是**四份复制型** ⇒ 修复必须**同批传播**且四份**字节一致**，否则 v2/v3 会在宿主升级后腐坏；**明显的正确性/安全缺陷无论落在哪套都修**（"冻结"= 停止功能投资，不是放任缺陷）| **随包文档** `docs/maintenance-policy.md`（中英同页）＋ `README.md`／`README.en.md` 各一行指引；`node scripts/release-check.mjs` 核对随包清单与 `files[]` 一致（声明了但缺失 = 0）|
+
 > **运行方式（本轮核实）**：`tests/run-tests.mjs:93-110` 收集 `tests/` 下**每一个** `.mjs`（仅跳过自身、`NEEDS_ARGS` 清单与 `replacedBare` 变体），因此 `audit-path-discipline.mjs` 是**门禁内的 probe**，不是"只能手动跑"；未随包发布的守卫（见 `package.json#files` 的 tests 子集）在安装树里不可运行。
 
 

package/docs/final-paper.md

@@ -121,6 +121,10 @@
   **所办必须先与全所交流、商讨、优化、审查**：至少 1 条所办消息（`vibe_v5_message`）**且**至少 1 次会议
   （`vibe_v5_meeting`），结论写进定稿说明，然后调用 `vibe_v5_finalize_paper`。
   两个计数记在 `paper.meta.json`；缺一即 `V5_PAPER_CONSULT_REQUIRED`（不是静默跳过）。
+- **`force` ＝ 一张新论文**：`/vN paper … force` **重写**已定稿的论文 ⇒ 新的 `paper` 状态以
+  `consult = {messages: 0, meetings: 0}` 开始，**重启前的消息与会议不再计入**（这是设计，不是丢状态 ✗）。
+  因此 `paperEditor="office"` 在 `force` 之后必须**重新**征询；`V5_PAPER_CONSULT_REQUIRED` 的响应会**具名**说明
+  这一点（`restarted: true` ＋ `forceReason`），不再让人靠猜。
 
 ## 8. LaTeX 检测、编译与降级
 

package/docs/maintenance-policy.md

@@ -0,0 +1,31 @@
+# 维护方针：v4/v5 优先，v2/v3 冻结为「适配兼容」
+
+> 本文件说明四套预设的**维护优先级**，供使用者与维护者参考。策略由维护者决定，可能随版本调整。
+
+## 中文
+
+**投资方向：`vibe-math-v4` 与 `vibe-math-v5`。** 新能力、新守卫（测试/审计）、提示词强化与重构都只落这两套。
+
+**`vibe-math-v2` 与 `vibe-math-v3` 已成熟定型（architecturally frozen）。** 它们仍随包安装、可正常使用，但**只接受「适配兼容」类改动**：
+
+- 宿主（DeepSeek Harness）升级带来的适配修复；
+- **共享模块的正确性/安全修复**：`vibe-math-v{2,3,4,5}/math-computation.js` 等是四份**复制型共享** ⇒ 修复必须**同批传播**且四份保持**字节一致**，否则 v2/v3 会在宿主升级后悄悄腐坏；
+- 让既有行为**保持不变**的回归修复。
+
+**因此**：v2/v3 的参数名、默认值与数据格式**不再演进**（新参数、新字段只出现在 v4/v5）；v2/v3 的测试套件保留为**兼容回归守卫**（保持通过即可），不再为它们新增作业。
+
+**例外（不可扩大解释）**：**明显的正确性/安全缺陷**无论落在哪一套都会被修复 —— 「冻结」指停止**功能投资**，不等于放任已发布代码里的缺陷。
+
+## English
+
+**Investment goes to `vibe-math-v4` and `vibe-math-v5`.** New capabilities, new guards (tests/audits), prompt strengthening and refactors land in these two only.
+
+**`vibe-math-v2` and `vibe-math-v3` are architecturally frozen.** They still install and work, but accept **compatibility-only** changes:
+
+- adaptations required by a DeepSeek Harness upgrade;
+- **correctness/security fixes in shared modules**: `vibe-math-v{2,3,4,5}/math-computation.js` and friends are **copied** into each preset, so a fix must be propagated in the same change and the four copies must stay **byte-identical** — otherwise v2/v3 rot silently on the next host upgrade;
+- regression fixes that keep their current behaviour unchanged.
+
+**Consequently** v2/v3 parameter names, defaults and data formats **no longer evolve** (new parameters and fields appear in v4/v5 only), and their test suites remain **compatibility regression guards** (kept green) without new jobs.
+
+**Exception (not to be stretched)**: an outright correctness/security defect is fixed wherever it lives — "frozen" means no further **feature** investment, not tolerating defects in shipped code.

package/docs/release-notes/README.en.md

@@ -8,6 +8,8 @@ Chinese counterpart of this index is [`README.md`](./README.md) (中文索引).
 
 | Version | Key change | Notes |
 |---|---|---|
+| 2.8.3 | Patch: a speech arriving **after the meeting closed** is no longer dropped silently (it becomes a named "late note" in the minutes, which stay append-only); an office finalisation refused after a `force` restart now **explains why** (`restarted: true` + `forceReason`, stating that the counters restarted with the new paper and that a fresh consultation is required); the docs now state "`force` = a new paper". | [中文](./RELEASE-NOTES-2.8.3.md) · [English](./RELEASE-NOTES-2.8.3.en.md) |
+| 2.8.2 | Patch: an explicit `paperLatexCommand` is now probed **alone**, so `triedPaths` no longer reports impossible paths of the form "known TeX root + absolute path"; v5's delivered `paper.md` / `paper.tex` now also **each state** the compile outcome (no engine detected / compilation failed), exactly once and without duplication on re-entry — matching v4 since 2.8.1. Both have real-host reproductions and named-red guards. | [中文](./RELEASE-NOTES-2.8.2.md) · [English](./RELEASE-NOTES-2.8.2.en.md) |
 | 2.8.1 | Patch: v4's delivered `paper.md` and `paper.tex` now each state the compile outcome (no engine detected → tex+md only, or PDF compilation failed with tex/md kept), **exactly once** and idempotently — resolving known limitation #3 of the 2.8.0 notes; added `scripts/release-check.mjs`, which runs the pre-release checks as one command (version triple, section order of both notes, manifest registration, declared-but-missing = 0, repository-only files not shipped, no CRLF in the package). | [中文](./RELEASE-NOTES-2.8.1.md) · [English](./RELEASE-NOTES-2.8.1.en.md) |
 | 2.8.0 | Feature: an explicit `paperLatexCommand` now means "use only this" (v4/v5 no longer fall back silently), the missing-engine warning is actionable and `triedPaths` is recorded, and all four presets share one byte-identical missing-engine guidance block; the test gate is steadier (default concurrency 2, stale scratch swept at start-up, read-only warning for suspected leftover processes); repository line endings are unified on LF (byte-reproducible npm package). | [中文](./RELEASE-NOTES-2.8.0.md) · [English](./RELEASE-NOTES-2.8.0.en.md) |
 | 2.7.3 | Patch: the v5 mailbox acknowledges **only after a successful send** (unacknowledged mail is no longer lost), round numbering is correct on founding and on resume, a failed required paper-artifact write is named once with an idempotent refill, a refused task CAS leaves no trace, the member path contract and v4's Lean library-index write failure are named; eight stale "ack first" passages were corrected and a prose↔code pairing guard now prevents them from returning. | [中文](./RELEASE-NOTES-2.7.3.md) · [English](./RELEASE-NOTES-2.7.3.en.md) |

package/docs/release-notes/README.md

@@ -7,6 +7,8 @@
 
 | 版本 | 关键变化 | 说明 |
 |---|---|---|
+| 2.8.3 | 补丁：会议**收束后到达的发言**不再静默丢弃（作为具名「会后补记」写入纪要、只增不覆盖）；`force` 重启后 `paperEditor="office"` 的定稿拒绝**说明原因**（新增 `restarted:true` ＋ `forceReason`，并写明"计数随新论文重置、需重新征询"）；文档明确「`force` ＝ 一张新论文」。 | [中文](./RELEASE-NOTES-2.8.3.md) · [English](./RELEASE-NOTES-2.8.3.en.md) |
+| 2.8.2 | 补丁：显式 `paperLatexCommand` 改为**单独探测** ⇒ `triedPaths` 不再出现「已知 TeX 根 ＋ 绝对路径」这类**不可能路径**；v5 交付的 `paper.md`／`paper.tex` 也**各自写明**编译结果（未检测到引擎／编译失败），恰好一次且重入不重复 ⇒ 与 2.8.1 的 v4 对齐。两处均有真机复现读数与具名红守卫。 | [中文](./RELEASE-NOTES-2.8.2.md) · [English](./RELEASE-NOTES-2.8.2.en.md) |
 | 2.8.1 | 补丁：v4 交付的 `paper.md`／`paper.tex` 现在各自写明编译结果（未检测到引擎 ⇒ 只产出 tex+md，或 PDF 编译失败并保留 tex/md），**恰好一次**且幂等 ⇒ 2.8.0 说明里的「已知限制」第 3 条已解决；新增 `scripts/release-check.mjs` 把发布前逐项核对做成一条命令（版本三处一致／两份说明章节顺序／清单登记／声明缺失=0／开发专用文件不随包／包内无 CRLF）。 | [中文](./RELEASE-NOTES-2.8.1.md) · [English](./RELEASE-NOTES-2.8.1.en.md) |
 | 2.8.0 | 功能版：显式 `paperLatexCommand` 改为「指定即只用它」（v4/v5 不再静默回落）、检测不到引擎的警告可行动并新增 `triedPaths`、四套预设新增逐字一致的引擎缺失指引；测试门禁更稳（默认并发 2、自动清扫陈旧临时目录、疑似遗留进程只读告警）；仓库行尾统一为 LF（npm 包逐字节可复现）。 | [中文](./RELEASE-NOTES-2.8.0.md) · [English](./RELEASE-NOTES-2.8.0.en.md) |
 | 2.7.3 | 补丁：v5 邮袋改为「发送成功后才确认」（未确认消息不再丢失）、入会与恢复的轮次编号正确、论文必需产物写失败具名告警且补偿幂等、任务 CAS 被拒不留痕、成员库路径契约与 v4 Lean 索引写失败具名；文档中 8 处旧顺序已改正并新增文↔码配对守卫。 | [中文](./RELEASE-NOTES-2.7.3.md) · [English](./RELEASE-NOTES-2.7.3.en.md) |

package/docs/release-notes/RELEASE-NOTES-2.8.2.en.md

@@ -0,0 +1,47 @@
+# dsh-vibe-math 2.8.2 — Release Notes
+
+> Previous release: 2.8.1. This is a **patch release** fixing two v5 issues around diagnostics and deliverables: `triedPaths` no longer reports **impossible paths**, and v5's delivered `paper.md` / `paper.tex` now state "no engine detected / compilation failed" (2.8.1 did this for v4 only). Parameter names and defaults, data formats and the directory layout are **unchanged**, the DSH support window is unchanged, and **no migration is needed**.
+
+## Overview
+
+- An explicit `paperLatexCommand` is now probed **alone**: `compile.triedPaths` in `paper.meta.json` no longer contains impossible paths of the form "known TeX root + the absolute path you supplied".
+- v5's delivered `paper.md` and `paper.tex` now **each state** the compile outcome (no engine detected → tex+md only; compilation failed → tex/md kept), **exactly once** and idempotently on re-entry, matching what v4 has done since 2.8.1.
+- Nothing else changed: with TeX in a non-standard location, keep pointing `paperLatexCommand` at an absolute path.
+
+## Added
+
+- None. (Patch release: no new parameters, fields or tools.)
+
+## Changed
+
+- v5's `paper.md` / `paper.tex` carry one extra outcome line (in the language of `paperLanguage`). Apart from that line, the paper's section order, fixed wording and other fields are unchanged.
+
+## Fixed
+
+- Symptom: after setting `paperLatexCommand` to an **absolute** path, `compile.triedPaths` in `paper.meta.json` contained impossible paths such as `…\MiKTeX\miktex\bin\x64\Z:/no/such/xelatex.exe` — and `triedPaths` is precisely the "where did it look" field, so it misled the reader. Cause: the explicit value was treated as a **command name** and joined onto every known install root. It is now probed **alone**, so `triedPaths` records only candidates that were really probed. Measured on a real host (reproduced with a build pinned to 2.8.1): bogus path → `status="not-detected"`, `triedPaths=["Z:/…"]`; real engine → `status="compiled"`, `paper.pdf` 150,044 bytes. Guards: `tests/e2e-v5-round2.test.mjs` (`★ [task-17/v5]`), `tests/v5-institute-fixes.mutants.mjs`.
+- Symptom: v5's delivered `paper.md` / `paper.tex` did **not** say why no PDF was produced (only the `compile` metadata and the log did). Cause: the paper body is composed **before** compilation. Finalisation now writes the outcome into the deliverables **after** compiling (a bullet at the end of `paper.md`; a `%` comment before `\end{document}` in `paper.tex`), **once each** and without duplication on re-entry. Guards: `tests/e2e-v5-round2.test.mjs` (`★ [task-18/v5]`), `tests/v5-institute-fixes.mutants.mjs` (**16/16** named reds).
+
+## Compatibility & Migration
+
+- Breaking changes: **none**.
+- Additive only: one outcome line inside the deliverables (v5 only).
+- Migration: none needed; existing `paper.meta.json` files remain readable.
+- The DSH support window is unchanged (see `engines.dsh` in `package.json`).
+
+## Known Limitations
+
+- **TeX installed in a non-standard location is still not auto-discovered**: the default probe covers PATH and the documented common install roots; use an absolute path in `paperLatexCommand` otherwise.
+- The shipped suites **do not cover host-level behaviour** (for example "the caller is dropped after an interrupt"); that needs a real host.
+- **A root-agent question blocks the whole institute** (operational note, product unchanged): nothing else proceeds until the answer arrives — state the problem and acceptance criteria up front for unattended runs. See [`slv-playbook.md`](./slv-playbook.md) §5.
+- On machines without Lean / LaTeX, the related suites **skip loudly** rather than passing silently.
+
+## Verification
+
+- `node tests/run-tests.mjs` — all green (**106** items: 44 suites + 62 probes).
+- `node scripts/release-check.mjs` — exits 0 (version triple, section order of both notes, `files[]` registration, no CRLF in the package); after publishing, `node scripts/release-check.mjs --registry` also compares the registry shasum with the local tarball.
+- The two fixes in this release: `node tests/e2e-v5-round2.test.mjs` (`★ [task-17/v5]`, `★ [task-18/v5]`) and `node tests/v5-institute-fixes.mutants.mjs` (**16/16** named reds).
+- Real-host re-check recipe: install the exact published version (`-Pin <version>`) and run a **trimmed** paper task (no research) — about **8 minutes** — then inspect the `compile` block and whether `paper.pdf` exists.
+
+## Dependencies
+
+- No new runtime dependencies; the DSH support window is unchanged.

package/docs/release-notes/RELEASE-NOTES-2.8.2.md

@@ -0,0 +1,47 @@
+# dsh-vibe-math 2.8.2 — 发布说明
+
+> 上一版本：2.8.1。本版是**补丁版**，修两处 v5 的「交付物与诊断」问题：`triedPaths` 不再出现**不可能路径**；v5 交付的 `paper.md`／`paper.tex` 也会写明「未检测到引擎／编译失败」（2.8.1 只做了 v4，本版补齐）。参数名与默认值、数据格式与目录结构**未变**，DSH 支撑窗口不变，**无需迁移**。
+
+## 概述
+
+- 显式 `paperLatexCommand` 现在被**单独探测**：`paper.meta.json` 的 `compile.triedPaths` 不再出现「已知 TeX 根 ＋ 你给的绝对路径」这种不可能路径。
+- v5 交付的 `paper.md` 与 `paper.tex` 现在**各自写明**编译结果（未检测到引擎 ⇒ 只产出 tex+md；编译失败 ⇒ 已保留 tex/md），**恰好一次**且定稿重入不重复 ⇒ 与 2.8.1 的 v4 行为**对齐**。
+- 其余行为不变：本机 TeX 装在非标准位置时，仍请用 `paperLatexCommand` 指定绝对路径。
+
+## 新增
+
+- 无。（本版为修复版：没有新增参数、字段或工具。）
+
+## 变更
+
+- v5 的 `paper.md`／`paper.tex` 多一行结果说明（随 `paperLanguage` 中英各一）。除此之外，论文的章节顺序、固定文案与其它字段均未变。
+
+## 修复
+
+- 症状：把 `paperLatexCommand` 设为**绝对**路径后，`paper.meta.json` 的 `compile.triedPaths` 里出现 `…\MiKTeX\miktex\bin\x64\Z:/no/such/xelatex.exe` 这类**不可能路径**——而 `triedPaths` 正是「找过哪里」的字段，会误导用户。原因：显式值被当作**命令名**拿去和已知安装根拼候选。现在显式值被**单独探测**，`triedPaths` 只记录真正探过的候选。真机实测（pin 到 2.8.1 后复现）：坏路径 ⇒ `status="not-detected"`、`triedPaths=["Z:/…"]`；真引擎 ⇒ `status="compiled"`、`paper.pdf` 150,044 B。守卫：`tests/e2e-v5-round2.test.mjs`（`★ [task-17/v5]`）、`tests/v5-institute-fixes.mutants.mjs`。
+- 症状：v5 交付的 `paper.md`／`paper.tex` **不说** PDF 为什么没有产出（此前只有 `compile` 元数据与日志里有）。原因：论文正文在**编译之前**组装。现在定稿在编译之后把结果**补写**进交付物（`paper.md` 末尾一条要点；`paper.tex` 在 `\end{document}` 之前插一行 `%` 注释），**各恰好一次**、重入不重复。守卫：`tests/e2e-v5-round2.test.mjs`（`★ [task-18/v5]`）、`tests/v5-institute-fixes.mutants.mjs`（**16/16** 具名红）。
+
+## 兼容性与迁移
+
+- 破坏性变更：**无**。
+- 纯新增：交付物里的一行结果说明（仅 v5）。
+- 迁移：无需迁移；既有 `paper.meta.json` 仍可读。
+- DSH 支撑窗口不变（见 `package.json` 的 `engines.dsh`）。
+
+## 已知限制
+
+- **非标准位置的 TeX 仍不会被自动发现**：默认探测只覆盖 PATH 与文档化的常见安装根；装在其它目录请用 `paperLatexCommand` 指绝对路径。
+- 随包套件**不覆盖宿主层行为**（例如「中断后调用方被丢弃」），这类行为需要真机才能观察。
+- **根代理提问会阻塞整所流程**（运维注意项，不改产品）：答复到达前其余流程停摆；长跑时请把问题与验收条件一次说清。详见 [`slv-playbook.md`](./slv-playbook.md) §5。
+- 缺少 Lean / LaTeX 等引擎的机器上，相关套件会**响亮跳过**，不会静默通过。
+
+## 验证方式
+
+- `node tests/run-tests.mjs` —— 全部通过（**106** 项：44 套件 + 62 探针）。
+- `node scripts/release-check.mjs` —— 退出 0（版本三处一致／两份说明章节顺序／`files[]` 登记／包内无 CRLF）；发布后 `node scripts/release-check.mjs --registry` 再比对线上 `dist.shasum` 与本地 tarball。
+- 本版两处修复：`node tests/e2e-v5-round2.test.mjs`（`★ [task-17/v5]`、`★ [task-18/v5]`）与 `node tests/v5-institute-fixes.mutants.mjs`（**16/16** 具名红）。
+- 真机复测方式：用 registry 上的精确版本（`-Pin <version>`）跑一个**精简**论文任务（跳过研究），约 **8 分钟**，核对 `compile` 段与 `paper.pdf` 是否存在。
+
+## 依赖
+
+- 无新增运行时依赖；DSH 支撑窗口不变。

package/docs/release-notes/RELEASE-NOTES-2.8.3.en.md

@@ -0,0 +1,50 @@
+# dsh-vibe-math 2.8.3 — Release Notes
+
+> Previous release: 2.8.2. This is a **patch release** fixing two v5 issues around meetings and finalisation: a speech that arrives **after the meeting closed** is no longer dropped silently (it becomes a clearly-marked late note), and an office finalisation refused after a `force` restart now **explains why** (the counters restarted with the new paper, so a fresh consultation is required). Parameter names and defaults, data formats and the directory layout are **unchanged**, the DSH support window is unchanged, and **no migration is needed**.
+
+## Overview
+
+- A member's meeting speech **no longer disappears silently**: even when it arrives **after** the meeting closed (the watchdog abandons a stalled meeting; the office may finalise early) it is written into the minutes as a clearly-marked "late note", with one named line in the group chat.
+- Meeting minutes are **append-only**: earlier speakers are always preserved.
+- A refusal after a `force` restart is now **explainable**: the response carries `restarted: true` plus the `forceReason`, and the message states that the counters restarted with the new paper, that earlier messages/meetings no longer count, and that a fresh consultation is required.
+- The documentation now states it plainly: **`force` means a new paper** (both `consult` counters start at 0), so a restart requires a **new** consultation.
+
+## Added
+
+- When `vibe_v5_finalize_paper` is refused for insufficient consultation, the response now carries **`restarted`** (boolean: was this paper started by a `force` restart?).
+
+## Changed
+
+- Meeting minutes gained one kind of entry: `### <member>（会后补记：该发言到达时会议已收束）` — used for speeches that arrive **after** the meeting closed; identical text submitted again is recorded only once.
+- The `V5_PAPER_CONSULT_REQUIRED` message appends the reset explanation when `restarted` (**behaviour is unchanged**; it only states the cause).
+- `docs/final-paper.md` §7 gained the "`force` = a new paper" rule.
+
+## Fixed
+
+- Symptom: a member's speech — on a real host, the **convener's own** speech — never appeared in the minutes: the header said `- 发言顺序: r-1 → acad` while the body contained only `### r-1`; the paper's `consult.meetings` stayed 0 and finalisation was refused repeatedly. Cause: a member's end-block `input` was only consumed when a meeting was **live**, and a stalled meeting is closed by the watchdog — speeches arriving later had nowhere to go and were **dropped silently**. Such speeches are now appended as "late notes" (named, de-duplicated, appended rather than overwritten). Guards: `tests/e2e-v5-round2.test.mjs` (`★ [real1004-minutes]`), `tests/v5-institute-fixes.mutants.mjs`.
+- Symptom: after a `force` restart of the paper, the office's usual finalisation was refused by `V5_PAPER_CONSULT_REQUIRED` with **no hint** that the restart had reset the counters (on a real host: **11** consecutive refusals, leaving the members to infer it). Cause: that semantics existed **only in the code** — neither the docs nor the refusal message said it. The refusal now names the reset and asks for a fresh consultation, and the docs state "`force` = a new paper". Guards: `tests/e2e-v5-round2.test.mjs` (`★ [real1004-consult]`), `tests/v5-institute-fixes.mutants.mjs` (**18/18** named reds).
+
+## Compatibility & Migration
+
+- Breaking changes: **none**.
+- Additive only: the `restarted` response field; one new kind of minutes entry.
+- Migration: none needed; existing `paper.meta.json` files, meeting minutes and the directory layout are unchanged.
+- The DSH support window is unchanged (see `engines.dsh` in `package.json`).
+
+## Known Limitations
+
+- **TeX installed in a non-standard location is still not auto-discovered**: the default probe covers PATH and the documented common install roots; use an absolute path in `paperLatexCommand` otherwise.
+- The shipped suites **do not cover host-level behaviour** (for example "the caller is dropped after an interrupt"); that needs a real host.
+- **A root-agent question blocks the whole institute** (operational note, product unchanged): nothing else proceeds until the answer arrives — state the problem and acceptance criteria up front for unattended runs. See [`slv-playbook.md`](./slv-playbook.md) §5.
+- After a meeting is abandoned by the watchdog, the minutes keep the abandonment marker and **all earlier speeches**; speeches that arrive later land in the "late note" section.
+
+## Verification
+
+- `node tests/run-tests.mjs` — all green (**106** items: 44 suites + 62 probes).
+- `node scripts/release-check.mjs` — exits 0 (version triple, section order of both notes, `files[]` registration, no CRLF in the package); after publishing, add `--registry` to compare the registry shasum with the local tarball.
+- The two fixes in this release: `node tests/e2e-v5-round2.test.mjs` (`★ [real1004-minutes]`, `★ [real1004-consult]`) and `node tests/v5-institute-fixes.mutants.mjs` (**18/18** named reds).
+- Real-host re-check recipe: install the exact published version (`-Pin <version>`) and run a **trimmed** meeting + paper task (no research), then check that the minutes contain both speakers and that a refusal after a `force` restart is named.
+
+## Dependencies
+
+- No new runtime dependencies; the DSH support window is unchanged.

package/docs/release-notes/RELEASE-NOTES-2.8.3.md

@@ -0,0 +1,50 @@
+# dsh-vibe-math 2.8.3 — 发布说明
+
+> 上一版本：2.8.2。本版是**补丁版**，修两处 v5 的会议与定稿问题：**会议收束后到达的发言不再被静默丢弃**（改成具名的"会后补记"）；`force` 重启后 `paperEditor="office"` 的定稿拒绝会**说明原因**（计数已随新论文重置、需重新征询）。参数名与默认值、数据格式与目录结构**未变**，DSH 支撑窗口不变，**无需迁移**。
+
+## 概述
+
+- 成员的会议发言**不会再无声消失**：即使发言在会议收束（看门狗放弃停滞会议，或所办提前定稿）**之后**才到达，也会作为「**会后补记**」写进会议纪要，并在群聊里具名说明一次。
+- 会议纪要**只增不覆盖**：先前发言者的内容一律保留。
+- `force` 重启后的定稿拒绝**可解释**：响应新增 `restarted: true` 并写明 `forceReason`，消息直接告诉你"计数随新论文重置、重启前的消息/会议不再计入、请重新征询"。
+- 文档明确：**`force` ＝ 一张新论文**（`consult` 的两个计数随之从 0 开始），因此重启后必须**重新**征询。
+
+## 新增
+
+- `vibe_v5_finalize_paper` 在因征询不足被拒时，响应新增 **`restarted`** 字段（布尔；该论文是否由 `force` 重启而来）。
+
+## 变更
+
+- 会议纪要新增一种条目：「会后补记」（`### <成员>（会后补记：该发言到达时会议已收束）`）——针对**会议已收束后**到达的发言；同一段文本重复提交只记一次。
+- `V5_PAPER_CONSULT_REQUIRED` 的文案在 `restarted` 时追加重置说明（**行为逻辑不变**，只是把原因说清）。
+- `docs/final-paper.md` §7 增补一条「`force` ＝ 一张新论文」的口径。
+
+## 修复
+
+- 症状：某位成员（真机实测中正是**会议的召集人**）的发言没有出现在会议纪要里——纪要在头部写着 `- 发言顺序: r-1 → acad`，正文却**只有 `### r-1`**；同一张论文的 `consult.meetings` 一直为 0，定稿被连续拒绝。原因：成员结束块里的 `input` 只认「**有活会议**」这一条路径，而会议一旦停滞会被看门狗收束，之后到达的发言**无处可去、静默丢弃**。现在这类发言会作为「会后补记」写入纪要（具名、去重、追加而非覆盖）。守卫：`tests/e2e-v5-round2.test.mjs`（`★ [real1004-minutes]`）、`tests/v5-institute-fixes.mutants.mjs`。
+- 症状：`force` 重启论文后，所办按老流程定稿会被 `V5_PAPER_CONSULT_REQUIRED` 反复拒绝，而**提示里没有任何线索**说明"刚刚重启导致计数清零"（真机实测连续 **11** 次、只能靠成员自行推断）。原因：这一语义**只存在于代码里**，文档与拒绝文案都没写。现在拒绝会具名说明重置与"请重新征询"，文档也把「`force` ＝ 新论文」写清楚。守卫：`tests/e2e-v5-round2.test.mjs`（`★ [real1004-consult]`）、`tests/v5-institute-fixes.mutants.mjs`（**18/18** 具名红）。
+
+## 兼容性与迁移
+
+- 破坏性变更：**无**。
+- 纯新增：`restarted` 响应字段；纪要里一种新的条目类型。
+- 迁移：无需迁移；既有 `paper.meta.json`、会议纪要与目录结构不变。
+- DSH 支撑窗口不变（见 `package.json` 的 `engines.dsh`）。
+
+## 已知限制
+
+- **非标准位置的 TeX 仍不会被自动发现**：默认探测只覆盖 PATH 与文档化的常见安装根；装在其它目录请用 `paperLatexCommand` 指绝对路径。
+- 随包套件**不覆盖宿主层行为**（例如「中断后调用方被丢弃」），这类行为需要真机才能观察。
+- **根代理提问会阻塞整所流程**（运维注意项，不改产品）：答复到达前其余流程停摆；长跑时请把问题与验收条件一次说清。详见 [`slv-playbook.md`](./slv-playbook.md) §5。
+- 会议「因长时间无新发言被看门狗放弃」之后，纪要里会留下放弃标记；**放弃前的发言仍保留**，之后到达的发言进「会后补记」。
+
+## 验证方式
+
+- `node tests/run-tests.mjs` —— 全部通过（**106** 项：44 套件 + 62 探针）。
+- `node scripts/release-check.mjs` —— 退出 0（版本三处一致／两份说明章节顺序／`files[]` 登记／包内无 CRLF）；发布后加 `--registry` 比对线上 shasum 与本地 tarball。
+- 本版两处修复：`node tests/e2e-v5-round2.test.mjs`（`★ [real1004-minutes]`、`★ [real1004-consult]`）与 `node tests/v5-institute-fixes.mutants.mjs`（**18/18** 具名红）。
+- 真机复测方式：用 registry 上的精确版本（`-Pin <version>`）跑一个**精简**会议＋论文任务（跳过研究），核对纪要是否含两位发言人、以及 force 重启后拒绝是否具名。
+
+## 依赖
+
+- 无新增运行时依赖；DSH 支撑窗口不变。

package/docs/release-playbook.md

@@ -37,6 +37,7 @@ node -e "const p=require('./package.json'),l=require('./package-lock.json');cons
 ## 3. 提交前：门禁与派生计数
 
 ```bash
+node scripts/clean-temp.mjs --ps                 # 先看有没有遗留进程（只读；--kill 需显式）
 node tests/run-tests.mjs                       # 必须全绿：TOTAL <n> PASS <n> FAIL 0
 node scripts/update-doc-counts.mjs             # 让文档里的派生计数收敛
 node scripts/update-doc-counts.mjs --check     # 退出 0 = 已收敛
@@ -44,6 +45,11 @@ node scripts/release-check.mjs                 # 版本三处一致／两份说
                                                # tarball 声明缺失=0／开发专用文件不随包／包内无 CRLF
 ```
 
+- **先查遗留进程，再谈"时序抖动"**（实测教训）：一只**两天前**启动、父进程已消失的宿主累计烧掉 **197,410 秒 CPU**，
+  在 4 核机上等于长期占满一个核 ⇒ 同一批作业被放慢 **1.3–4.5×**、门禁耗时总和从 **2822 s 涨到 7875 s**，于是
+  "单跑全绿"的套件在门禁里偶发报红。`scripts/clean-temp.mjs --ps` 只读列出来（判据：命令行属于本工程/团队**且**
+  父进程已消失），`--kill` 才动手；门禁启动时也会打印一行只读告警。
+
 - **`scripts/release-check.mjs` 是本手册的可执行化部分**：把 §1／§2／§6 的逐项核对做成一条命令（含一次真实
   `npm pack` 与解包扫描）。发布前它必须退出 0；`--skip-pack` 只做静态检查，`--registry` 在发布后比对线上
   `dist.shasum` 与本地 tarball 的字节一致性，`--self-test` 自证判据不是空转（配 `tests/release-check.mutants.mjs`

package/docs/slv-playbook.md

@@ -75,7 +75,11 @@ dsh --profile <clone> --port <自有端口> --no-open
 
 - **没有任何引擎**（Octave/Julia/MATLAB/Maple/Wolfram 等）⇒ 这些面只能核对**诚实文本**：`configured`（配置启用）与 `absent[]`（每项带 `why`），这是**部分替代**，不是执行验证；
 - **单客户端** ⇒ 并发要靠两个克隆 + 两个端口模拟，否则如实声明超出范围；
-- **有界运行只能建立"落位与可观测性"**，**不能**建立"收敛/收口"结论——收敛结论必须来自随包套件。
+- **有界运行只能建立"落位与可观测性"**，**不能**建立"收敛/收口"结论——收敛结论必须来自随包套件；
+- **根代理提问会阻塞整所流程（实测观察到，属既有行为，不改产品）**：当**根/主代理**（而非某个成员）调用 `ask_user_question`
+  这类"等待人工答复"的工具时，在答复到达之前**其余流程会停摆**（转录里可看到停在这一步、后续事件不再推进）。
+  影响面：把研究所当成**无人值守**长跑时会卡住；有界影响、**不主张缺陷**。使用建议：长跑时避免让根代理走到提问分支
+  （把问题、范围与验收条件在任务里一次说清），或准备好及时答复；这也是把"**有界运行**"当结论必须慎重的另一条理由。
 
 ## 6. 中止触发与处置
 

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-vibe-math",
   "description": "Multi-agent mathematical problem-solving & verification frameworks for DeepSeek Harness — FOUR agent presets in one install: vibe-math-v2 (probability-driven, classic: qs.json + Propos knowledge base + explorer→solver→review/debate verdict), vibe-math-v3 (THIRD-generation, classic: paper-style Markdown knowledge base with Problems/Progress/Propos/Methods/Verified + planner-agent scheduling that decides the next N actions + universal theory/method invention library + agents write their own Markdown directly via a per-file write lock), and vibe-math-v4 (FOURTH-generation: persistent self-organizing resident subagents that message & meet to decide all tasks, verify only by unanimous consensus, /compact at a context threshold, and stop only when all agree the problem is solved), and vibe-math-v5 (FIFTH-generation research institute: an academician as the organizational centre who decomposes and ASSIGNS work and chairs meetings; permanent researchers who hold the vote and may hire/fire their own temp workers; temp workers with no vote; a group chat and meetings; a durable per-recipient mailbox; a compare-and-set task DAG; and a boolean m-vote consensus rule where an object enters Verified/ only when at least m voting members agree AND every one of them returns exactly 1 or exactly 0). Installing this bundle auto-installs all four presets (v1 was removed at v2.0.0). Each preset also writes a FINAL PAPER by default when its run concludes — and does so BEFORE the run is marked complete: on v4/v5 the team co-authors its parts, merges them, cross-reviews, and the editor named by paperEditor finalises; the output is Paper/<run or institute id>/{paper.md,paper.tex,paper.pdf when a LaTeX engine is detected,paper.meta.json,paper.log.md} (params finalPaper/paperFormat/paperLanguage/paperCompilePdf/paperLatexCommand, plus paperEditor on v4/v5; see docs/final-paper.md).",
-  "version": "2.8.1",
+  "version": "2.8.3",
   "type": "module",
   "engines": {
     "node": "^22.19.0 || >=24.0.0",
@@ -24,6 +24,7 @@
     "README.en.md",
     "docs/COMPAT-AUDIT-ROUND2.md",
     "docs/AUDIT-CHECKLIST.md",
+    "docs/maintenance-policy.md",
     "tests/audit-formal-sensitivity.mjs",
     "tests/audit-persona-sensitivity.mjs",
     "tests/audit-persona-surface.test.mjs",
@@ -206,7 +207,11 @@
     "docs/release-notes/RELEASE-NOTES-2.8.0.md",
     "docs/release-notes/RELEASE-NOTES-2.8.0.en.md",
     "docs/release-notes/RELEASE-NOTES-2.8.1.md",
-    "docs/release-notes/RELEASE-NOTES-2.8.1.en.md"
+    "docs/release-notes/RELEASE-NOTES-2.8.1.en.md",
+    "docs/release-notes/RELEASE-NOTES-2.8.2.md",
+    "docs/release-notes/RELEASE-NOTES-2.8.2.en.md",
+    "docs/release-notes/RELEASE-NOTES-2.8.3.md",
+    "docs/release-notes/RELEASE-NOTES-2.8.3.en.md"
   ],
   "license": "MIT",
   "keywords": [
@@ -236,7 +241,7 @@
     },
     "minVersion": "0.1.2-alpha.4",
     "testedVersion": "0.2.0-rc.2",
-    "compatNote": "v2.8.1 补丁版：修掉上一版说明里如实写出的「已知限制」——v4 的最终文稿此前不说明引擎缺失/编译失败。① v4 交付的 paper.md 与 paper.tex 现在各自写明编译结果（未检测到引擎 ⇒ 只产出 tex+md；PDF 编译失败 ⇒ 已保留 tex/md），恰好一次且定稿重入不重复；论文的章节结构与其它字段未变。② 新增随包脚本 scripts/release-check.mjs：把发布前逐项核对做成一条命令（版本三处一致、当前版本两份发布说明的章节顺序与 files[] 登记、compatNote 是否提到当前版本、npm pack 声明缺失=0、开发专用文件不随包、包内无 CRLF），并配 4 条单点变异自证非空转。参数名与默认值、数据格式与目录结构均未变；DSH 支撑窗口不变；无需迁移。",
+    "compatNote": "v2.8.3 补丁版：修 v5 的会议与定稿两处。① 会议收束后到达的成员发言不再静默丢弃 —— 看门狗放弃停滞会议（或所办提前定稿）后到达的 input 现在作为具名「会后补记」追加进纪要（按文本去重、先前发言只增不覆盖）；真机实测原症状是纪要在头部写着「发言顺序: r-1 → acad」而正文只有 ### r-1。② force 重启后 paperEditor=\"office\" 的定稿拒绝现在说明原因：响应新增 restarted 字段并在文案里写明 forceReason 与「计数随新论文重置、重启前的消息/会议不再计入、请重新征询」；docs/final-paper.md §7 明确「force ＝ 一张新论文」。参数名与默认值、数据格式与目录结构均未变；DSH 支撑窗口不变；无需迁移。",
     "compatibility": {
       "dshReleases": {
         "0.1.2-alpha.4": "compatible",

package/tests/e2e-v5-round2.test.mjs

@@ -1299,6 +1299,13 @@ console.log('\n[24] manual paper, office editor: parts → cross-review → cons
   console.log('    [task-9] v5 detection evidence: compile=' + String(meta && meta.compile && meta.compile.status) + ' triedPaths=' + JSON.stringify((meta && meta.compile && meta.compile.triedPaths) || null))
   assert(Array.isArray(meta && meta.compile && meta.compile.triedPaths) && meta.compile.triedPaths.length > 0,
     '★ [task-9] the durable meta records the probed TeX paths (triedPaths)')
+  // task-18 (v4 got this in 2.8.1): the DELIVERED artifacts must state why no PDF was produced — the body is
+  // composed before the compile runs, so the note is written afterwards, exactly once each (idempotent).
+  // The note follows paperLanguage, so the check is language-agnostic and still requires EXACTLY one.
+  const noteHits = (t) => t.split('本机未检测到 LaTeX 引擎：只产出 tex+md').length - 1
+    + t.split('No LaTeX engine detected: tex+md only').length - 1
+  assert(noteHits(md) === 1 && noteHits(tex) === 1,
+    '★ [task-18/v5] the DELIVERED paper.md and paper.tex state that no engine was detected (once each; md=' + noteHits(md) + ' tex=' + noteHits(tex) + ')')
   const warnText = (meta && Array.isArray(meta.warnings) ? meta.warnings : []).join(' | ')
   assert(warnText.indexOf('只交付 paper.tex 与 paper.md。') !== -1 && warnText.indexOf('已探测 PATH 与文档化的常见 TeX 根；可用 paperLatexCommand 指定绝对路径') !== -1,
     '★ [task-9] the not-detected warning keeps its tail AND says where it looked and how to pin an engine')
@@ -1343,6 +1350,20 @@ console.log('\n[25] paper idempotency: repeated triggers only fill missing artif
   const stF = await h.callTool('vibe_v5_status', {})
   assert(stF.paper.status === 'writing' && stF.paper.round === 1, 'the forced run restarted at round 1')
 }
+// real1004-consult (found on a real host with 2.8.2): `force` rewrites the paper, so the NEW paper starts with
+// consult={0,0} — intended — but the office was rejected ELEVEN times with no hint that its earlier meeting had
+// stopped counting (only the members could infer it). The refusal must NAME the reset and say what to do.
+{
+  const hq = makeHost({ pluginModule })
+  await hq.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await hq.settleSpawns()
+  const fq = await hq.callTool('vibe_v5_paper', { force: true, editor: 'office', reason: 'real1004 复测' })
+  assert(fq.ok === true && fq.started === true, 'precondition: a forced office-editor paper starts (' + JSON.stringify(fq).slice(0, 130) + ')')
+  await drivePaper(hq)   // parts + cross-reviews land, so the office's finalisation reaches the CONSULT gate
+  const rq = await hq.callTool('vibe_v5_finalize_paper', { decision: 'deliverable', note: 'real1004：未重新征询，预期被具名拒绝' })
+  assert(rq.ok === false && rq.code === 'V5_PAPER_CONSULT_REQUIRED' && rq.restarted === true && /force/.test(String(rq.message)) && /重置/.test(String(rq.message)) && /重新/.test(String(rq.message)),
+    '★ [real1004-consult] an office finalisation after a force restart NAMES the reset (restarted=true, message explains the required re-consultation): ' + JSON.stringify(rq).slice(0, 300))
+}
 
 // ---------- 26. fake LaTeX: the success path produces paper.pdf ----------
 console.log('\n[26] fake LaTeX compiler: success path produces paper.pdf')
@@ -1437,6 +1458,42 @@ console.log('\n[27] fake LaTeX compiler: repair path and persistent-failure degr
   const metap = JSON.parse(readFileSync(join(dp, 'paper.meta.json'), 'utf8'))
   assert(metap.compile.status === 'not-detected',
     '★ [task-10/v5] an unresolvable explicit paperLatexCommand degrades (never falls through to another engine) — got ' + JSON.stringify(metap.compile.status))
+  // task-17 (found on a real host with 2.8.1): the explicit value must be probed ALONE. Probing it as a bare
+  // command name let the known-root stage build impossible paths like `<MiKTeX root>/Z:/…/xelatex.exe`, which
+  // made `triedPaths` — the very field that says WHERE it looked — report paths that cannot exist.
+  const tp = metap.compile.triedPaths || []
+  const joined = tp.filter((p) => /[\\/][A-Za-z]:[\\/]/.test(String(p).replace(/^[A-Za-z]:[\\/]/, '')))
+  assert(tp.length >= 1 && joined.length === 0,
+    '★ [task-17/v5] an ABSOLUTE explicit paperLatexCommand is probed ALONE — triedPaths never joins it onto a known TeX root (triedPaths=' + JSON.stringify(tp) + ')')
+}
+// real1004-minutes (found on a real host with 2.8.2): a speech that arrives AFTER the meeting closed — the
+// watchdog abandons a stalled meeting and the office may finalise early — used to be dropped SILENTLY, so the
+// minutes declared a speaking order ("r-1 → acad") that never appeared in the file (only "### r-1"). It must
+// be appended as a clearly-marked late note, and the earlier speeches must survive (append-only).
+{
+  const hm = makeHost({ pluginModule })
+  await hm.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await hm.settleSpawns()
+  const mt = await hm.callTool('vibe_v5_meeting', { agenda: 'real1004 晚到发言：会议收束后仍有人发言', kind: 'sync' })
+  const mid = mt && mt.meeting
+  assert(!!mid, 'precondition: the meeting is convened (' + JSON.stringify(mt).slice(0, 120) + ')')
+  await hm.drain(40)                                  // everyone speaks -> the meeting finalises, meeting = null
+  const mpath = join(hm.WS, 'VibeMath', 'Projects', 'default', 'Institutes', 'institute', 'Shared', 'Meetings', mid + '.md')
+  const before = readFileSync(mpath, 'utf8')
+  assert(/### /.test(before), 'precondition: the closed minutes already contain at least one speech: ' + before.slice(0, 140))
+  const stc = await hm.callTool('vibe_v5_status', {})
+  assert(!stc.meeting, 'precondition: the meeting is CLOSED (meeting=' + JSON.stringify(stc.meeting) + ') so a later speech has no live meeting')
+  // A LATE speech: the member submits `input` on a LATER wake while no meeting is running (exactly what a
+  // real host showed — the member kept submitting after the meeting was already closed).
+  await hm.callTool('vibe_v5_message', { to: 'acad', content: 'real1004：请再确认一次你的意见。' })
+  const wlate = await hm.peekWakeOf('acad', 8000)
+  assert(!!wlate, 'precondition: a member is woken again after the meeting closed (so the end block is processed)')
+  hm.fireEnd(wlate.childId, { input: 'real1004 晚到发言：会议收束之后我仍提交了意见。', contextPct: 20 })
+  await sleep(200)
+  const after = readFileSync(mpath, 'utf8')
+  assert(after.includes('real1004 晚到发言') && after.includes('会后补记'),
+    '★ [real1004-minutes] a speech arriving after the meeting closed is appended as a late note, never dropped silently')
+  assert(after.includes('### '), 'the earlier speeches are still in the minutes (append-only, never clobbered)')
 }
 
 // ---------- 28. office-only surface + id normalisation ----------

package/tests/v5-institute-fixes.mutants.mjs

@@ -194,13 +194,46 @@ const FAMILIES = [
     to: '只交付 paper.tex 与 paper.md。',
     expect: /\[task-9\] the not-detected warning keeps its tail AND says where it looked/,
   },
+  // RETIRED (task-17, 2.8.1 real-host fix): the former entry here mutated `return [forced]` into the merged
+  // `[forced].concat(order…)`, which used to let an unresolvable explicit command fall through to a real
+  // engine. After task-17 the explicit value is ALSO handed to resolveKnownTool as `explicit` (the explicit
+  // branch probes only it and returns early), so the property is now enforced at TWO independent sites and
+  // NO single-site mutation can defeat it — the merged list merely probes dead candidates. The assertion
+  // itself stays in tests/e2e-v5-round2.test.mjs (`★ [task-10/v5] …degrades`); only the falsifier is retired
+  // (and the task-17 entry below covers the hand-off single-site).
   {
-    // task-10: restoring the merged candidate list lets an unresolvable explicit command fall through
-    // to another engine (the documented rule is "explicit ⇒ only it").
-    name: 'task-10: the explicit LaTeX command falls through to another engine again',
-    from: 'return [forced]',
-    to: 'return [forced].concat(order.filter((n) => n !== forced))',
-    expect: /\[task-10\/v5\] an unresolvable explicit paperLatexCommand degrades/,
+    // task-17 (real host, 2.8.1): probing the explicit value as a BARE COMMAND NAME lets the known-root stage
+    // join it onto every TeX root, so `triedPaths` reports impossible paths. Reverting the hand-off must
+    // redden the guarding assertion by name.
+    name: 'task-17: the explicit paperLatexCommand is probed as a bare command name again (triedPaths gets fake joins)',
+    from: "explicit: forcedTex, kind: 'tex'",
+    to: "explicit: '', kind: 'tex'",
+    expect: /\[task-17\/v5\] an ABSOLUTE explicit paperLatexCommand is probed ALONE/,
+  },
+  {
+    // task-18: the DELIVERED artifacts must say why no PDF was produced (the body is composed before the
+    // compile runs, so the note is written after it). Blunting the zh note must redden that guard BY NAME.
+    // The scenario that guards this runs with lang='en', so the ENGLISH branch is the one to blunt.
+    name: 'task-18: the delivered paper stops stating that no engine was detected',
+    from: ": 'No LaTeX engine detected: tex+md only'",
+    to: ": ''",
+    expect: /\[task-18\/v5\] the DELIVERED paper.md and paper.tex/,
+  },
+  {
+    // real1004-minutes: a speech arriving after the meeting closed must be appended as a late note; dropping
+    // it silently is exactly what a real host showed (minutes declared "r-1 → acad" with only "### r-1").
+    name: 'real1004-minutes: a speech arriving after the meeting closed is dropped silently again',
+    from: "} else if (lastMeetingId && typeof p.input === 'string' && p.input.trim()) {",
+    to: '} else if (false) {',
+    expect: /\[real1004-minutes\] a speech arriving after the meeting closed/,
+  },
+  {
+    // real1004-consult: the office refusal must NAME the force reset (observed: 11 unnamed rejections on a real
+    // host). Dropping the clause must redden the guarding assertion by name.
+    name: 'real1004-consult: the office refusal stops naming the force restart',
+    from: 'const restarted = p.forced === true',
+    to: 'const restarted = false',
+    expect: /\[real1004-consult\] an office finalisation after a force restart NAMES the reset/,
   },
 ]
 let red = 0

package/vibe-math-v5/vibe-math-v5.js

@@ -990,6 +990,11 @@ export function apply(ctx) {
     const inflight = new Map()        // childId -> turn token (dedupes duplicate subagent/end)
     let heartbeatDisposer = null
     let meeting = null                // in-flight meeting round state
+    // real1004-minutes: the id of the MOST RECENT meeting. A speech can arrive AFTER the meeting was
+    // closed (the watchdog abandons a stalled meeting, and the office may finalise early); such a late
+    // submission used to be dropped SILENTLY, so the minutes claimed a speaking order that never appeared
+    // in the file. Keep the pointer so a late speech can still be appended as a clearly-marked note.
+    let lastMeetingId = ''
     let pendingMeeting = null          // parked meeting (never preempts verification)
     let digestTimer = null
     let lastProgressAt = now()
@@ -4971,6 +4976,7 @@ export function apply(ctx) {
         return idx
       } })
       const id = idx.id
+      lastMeetingId = id   // real1004-minutes: late speeches are appended to the most recent minutes
       meeting = {
         id, agenda: opts.agenda, kind: opts.kind || 'sync', target: opts.target || '',
         by: opts.by || 'office', order, inputs: {}, extras: {}, lastInputAt: now(), startedAt: now(),
@@ -5623,11 +5629,20 @@ export function apply(ctx) {
       }
       const consult = Object.assign({ messages: 0, meetings: 0 }, p.consult || {})
       if (!(Number(consult.messages) > 0) || !(Number(consult.meetings) > 0)) {
+        // real1004-consult: `force` REWRITES the paper (docs/final-paper.md: "force 重写已定稿的论文"), so the
+        // new paper starts with consult={0,0}. That is intended — but the refusal used to say nothing about it:
+        // on a real host the office was rejected ELEVEN times and only the members could infer why their
+        // earlier meeting had stopped counting. Name the reset, its reason, and what to do about it.
+        const restarted = p.forced === true
         return {
           ok: false, code: 'V5_PAPER_CONSULT_REQUIRED',
           message: 'paperEditor="office" 要求所办先与全所交流、商讨、优化、审查：至少 1 条所办消息（vibe_v5_message）+ 至少 1 次会议（vibe_v5_meeting），然后才能定稿。当前：messages=' +
-            Number(consult.messages || 0) + ', meetings=' + Number(consult.meetings || 0),
+            Number(consult.messages || 0) + ', meetings=' + Number(consult.meetings || 0) +
+            (restarted
+              ? '。注意：本论文由 force 重启（forceReason=' + String(p.forceReason || '') + '）⇒ consult 计数随新论文重置为 0，重启前的消息/会议不再计入，请**重新**征询（两个计数记录在 paper.meta.json）'
+              : ''),
           consult: { messages: Number(consult.messages || 0), meetings: Number(consult.meetings || 0) },
+          restarted,
         }
       }
       const fin = {
@@ -5893,8 +5908,13 @@ export function apply(ctx) {
       }
       const engines = []
       const triedPaths = []   // task-9: union of the known-root candidates probed across THIS detection
+      // task-17: when paperLatexCommand is set, paperEngineCandidates() returns THAT VALUE ALONE. Probing it as
+      // a bare command name made resolveKnownTool join it onto every known TeX root, so `triedPaths` reported
+      // impossible paths like `<MiKTeX root>/Z:/no/such/xelatex.exe` (observed on a real host with 2.8.1).
+      // Handing it over as the EXPLICIT value probes only it (docs/final-paper.md §B); '' keeps auto-detection.
+      const forcedTex = String(params.paperLatexCommand || '').trim()
       for (const name of paperEngineCandidates(lang)) {
-        try { const _vr = await resolveKnownTool(sub, { name, explicit: '', kind: 'tex' }); triedPaths.push(..._vr.tried); const exe = _vr.exe; if (exe) engines.push({ name: String(name), exe: String(exe), via: _vr.via }) } catch (e) { /* not installed */ }
+        try { const _vr = await resolveKnownTool(sub, { name, explicit: forcedTex, kind: 'tex' }); triedPaths.push(..._vr.tried); const exe = _vr.exe; if (exe) engines.push({ name: String(name), exe: String(exe), via: _vr.via }) } catch (e) { /* not installed */ }
       }
       return { engines, triedPaths, reason: engines.length ? '' : ('未检测到 LaTeX 引擎（' + paperEngineCandidates(lang).join('/') + '）') }
     }
@@ -6065,6 +6085,27 @@ export function apply(ctx) {
       // broken engine. Whatever the compile outcome, the DELIVERED tex must be the real,
       // fully-generated one (the attempts are recorded in the meta either way).
       if (p.format !== 'md') await writeTextRel(PAPER_FILE(id, 'paper.tex'), tex)
+      // task-18: the body is composed BEFORE compilation runs, so the outcome cannot live in the sections.
+      // v4 states it in the delivered artifacts since 2.8.1; v5 must too — ONCE and idempotently (finalisation
+      // can be re-entered). The marker sentence is the same family as the warning below, and the presence
+      // check keeps the two paths from ever duplicating it.
+      if (compile.status === 'failed' || compile.status === 'not-detected') {
+        const zhNote = p.lang !== 'en'
+        const note = compile.status === 'failed'
+          ? (zhNote ? 'PDF 编译失败：已保留 paper.tex 与 paper.md 并上报' : 'PDF compilation failed: paper.tex and paper.md were kept and reported')
+          : (zhNote ? '本机未检测到 LaTeX 引擎：只产出 tex+md' : 'No LaTeX engine detected: tex+md only')
+        const annotate = async (rel, file, isTex) => {
+          const cur = await readTextAbs(paperAbs(id, file))
+          if (typeof cur !== 'string' || cur.indexOf(note) !== -1) return
+          const line = (isTex ? '% ' : '- ') + note
+          const next = isTex && cur.indexOf('\\end{document}') !== -1
+            ? cur.replace('\\end{document}', line + '\n\\end{document}')
+            : cur.replace(/\s*$/, '') + '\n' + line + '\n'
+          await writeTextRel(rel, next)
+        }
+        if (p.format !== 'tex') await annotate(PAPER_FILE(id, 'paper.md'), 'paper.md', false)
+        if (p.format !== 'md') await annotate(PAPER_FILE(id, 'paper.tex'), 'paper.tex', true)
+      }
       if (await fileExistsAbs(paperAbs(id, 'paper.pdf'))) files.push('paper.pdf')
       const warnings = (p.warnings || []).slice()
       if (compile.status === 'failed') warnings.push('LaTeX 编译失败（已保留 paper.tex 与 paper.md，不阻塞定稿）：' + (compile.attempts || []).map((a) => a.engine + '/' + a.label + ' exit=' + a.exitCode + (a.message ? '(' + a.message + ')' : '')).join('；'))
@@ -6782,6 +6823,20 @@ export function apply(ctx) {
         const rel = 'Shared/Meetings/' + meeting.id + '.md'
         const prev = (await readTextRel(rel)) || ('# 会议纪要｜' + meeting.id + '\n\n')
         await writeTextRel(rel, prev + '### ' + member.id + '\n' + (text || '（无发言）') + '\n\n')
+      } else if (lastMeetingId && typeof p.input === 'string' && p.input.trim()) {
+        // real1004-minutes: no live meeting, yet the member submitted a speech (a real host showed exactly
+        // this: `input` kept arriving after the meeting had been closed/abandoned). It used to be dropped
+        // SILENTLY — the minutes then declared a speaking order that never appeared in the file (`- 发言顺序:
+        // r-1 → acad` with only `### r-1` present). Append it as a clearly marked late note instead, keyed on
+        // ANY wake (not only meeting wakes: the observed submissions arrived on later rounds), de-duplicated by
+        // text so repeated submissions do not pile up.
+        const late = String(p.input).trim()
+        const rel = 'Shared/Meetings/' + lastMeetingId + '.md'
+        const prev = (await readTextRel(rel)) || ('# 会议纪要｜' + lastMeetingId + '\n\n')
+        if (prev.indexOf(late) === -1) {
+          await writeTextRel(rel, prev + '### ' + member.id + '（会后补记：该发言到达时会议已收束）\n' + late + '\n\n')
+          await saveChatLine('【会议 ' + lastMeetingId + '】' + member.id + ' 的发言在会议收束后到达，已作为补记写入纪要（不再静默丢弃）。')
+        }
       }
       if (p.solved !== undefined) {
         await writeTextRel('Shared/State-of-institute.md', [