v2.8.4

本页为 dsh-vibe-math 2.8.4 的双语发布说明（中文在前，English below）。

dsh-vibe-math 2.8.4 — 发布说明

上一版本：2.8.3。本版是补丁版：修正 v5 的会议看门狗——它此前会在"已被点名的成员仍在进行中"时就把会议放弃，导致成员的正常发言退化成「会后补记」，并让 paperEditor="office" 的定稿被反复拒绝。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

• 会议不再在"被点名、且仍在进行中"的成员交回发言之前被放弃：这类成员最多把会议续命到停滞预算的 3 倍，超过硬界才放弃（真正卡死的一轮仍不能拖住整场会议）。
• 结果是正常发言回到正常小节：真机实测修复前一次运行出现 1067 次「会后补记」、863 次定稿被拒；修复后会议能真正收齐发言，「会后补记」退回为保险而非常态。
• paperEditor="office" 的定稿闸门因此更容易被真正满足（它要求本张论文至少有 1 条所办消息与 1 次记录在案的会议）。
• 文档写明会议何时会被放弃、放弃如何具名、以及放弃之后到达的发言如何处置。

新增

• 无。（本版为修复版：没有新增参数、字段或工具。）

变更

• 会议停滞判定：软界（2 × activityTimeoutMs，默认 240 s）内，只要仍有被点名的成员在进行中就不放弃；超过 3 倍（默认 720 s）的硬界则照旧放弃。
• docs/final-paper.md §7 增补"会议何时会被放弃"的口径（含软界／硬界与「会后补记」的关系）。

修复

• 症状：成员（含召集人）在会议里发言后，纪要里却只留下「会后补记」条目；一次真机运行累计 1067 次补记，并且 paperEditor="office" 的定稿被 V5_PAPER_CONSULT_REQUIRED 拒绝 863 次。原因：会议停滞只按"最后一次发言"计时，不看被点名的成员是否仍在忙；真机上成员的一整轮常常超过 240 s 的预算，于是会议在看门狗手里被放弃，随后到达的发言只能走补记通道。现在"被点名且仍在忙"的成员会把放弃推迟到硬界（3 倍预算），正常发言因此回到会议的正常小节。守卫：tests/e2e-v5-round2.test.mjs（★ [real1004-stall]，源码级不变量：该分支的真机行为无法在随包套件里确定性触发，故判据是"源码不变量 ＋ 真机复测"）、tests/v5-institute-fixes.mutants.mjs（19/19 具名红）。

兼容性与迁移

• 破坏性变更：无。
• 纯新增：无新增字段（仅行为与文档变化）。
• 迁移：无需迁移。
• DSH 支撑窗口不变（见 package.json 的 engines.dsh）。

已知限制

• 会议看门狗分支的真机行为只能由真机复测确认（随包套件里心跳延时不可控 ⇒ 该分支不能确定性触发；因此守卫是源码不变量，这一点已在测试注释与提交信息里写明）。
• 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。
• 随包套件不覆盖宿主层行为（例如"中断后调用方被丢弃"）。
• 根代理提问会阻塞整所流程（运维注意项，不改产品）：长跑时请把问题与验收条件一次说清。详见 slv-playbook.md §5。

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；发布后加 --registry 比对线上 shasum 与本地 tarball。
• 本版修复：node tests/e2e-v5-round2.test.mjs（★ [real1004-stall]，实测 passed=436 failed=0）与 node tests/v5-institute-fixes.mutants.mjs（19/19 具名红）。
• 真机复测配方：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），统计「会后补记」次数（修复前 1067 次）与定稿被拒次数（修复前 863 次）应显著下降。

依赖

• 无新增运行时依赖；DSH 支撑窗口不变。

---

dsh-vibe-math 2.8.4 — Release Notes

Previous release: 2.8.3. This is a patch release: it fixes v5's meeting watchdog, which used to abandon a meeting while an ASKED member was still working — turning ordinary speeches into "late notes" and making paperEditor="office" finalisations fail again and again. 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 meeting is no longer abandoned before an ASKED member still in flight hands its speech in: such members may hold the meeting open up to 3× the stall budget, after which the hard bound still closes it (a genuinely wedged turn must not wedge the whole meeting).
• The practical effect is that ordinary speeches land in the ordinary sections again: one real run before the fix showed 1067 "late notes" and 863 refused finalisations; with the fix, meetings actually collect their speakers and the "late note" path goes back to being an insurance policy rather than the norm.
• The paperEditor="office" gate therefore becomes satisfiable in practice (it requires at least one office message and one recorded meeting for this paper).
• The documentation now states when a meeting is abandoned, how the abandonment is named, and what happens to speeches that arrive after it.

Added

• None. (Patch release: no new parameters, fields or tools.)

Changed

• Meeting stall decision: within the soft bound (2 × activityTimeoutMs, default 240 s) the meeting is kept open as long as an asked member is still in flight; past the 3× hard bound (default 720 s) it is still abandoned.
• docs/final-paper.md §7 now documents "when a meeting is abandoned" (soft/hard bounds and their relation to the "late note" path).

Fixed

• Symptom: after members (including the convener) spoke in a meeting, the minutes only gained "late note" entries — one real run reached 1067 of them, and paperEditor="office" finalisations were refused by V5_PAPER_CONSULT_REQUIRED 863 times. Cause: the stall check measured only the time since the last INPUT and ignored whether an asked member was still busy; a member's turn on a real host frequently outlives the 240 s budget, so the meeting was abandoned by the watchdog and later speeches could only take the late-note path. Members that were asked and are still busy now hold the meeting open until the hard bound (3× the budget), so ordinary speeches return to the ordinary sections. Guards: tests/e2e-v5-round2.test.mjs (★ [real1004-stall], a source-level invariant — this branch cannot be triggered deterministically inside the shipped suites, so the criterion is "source invariant + real-host re-test"), tests/v5-institute-fixes.mutants.mjs (19/19 named reds).

Compatibility & Migration

• Breaking changes: none.
• Additive only: no new fields (behaviour and documentation only).
• Migration: none needed.
• The DSH support window is unchanged (see engines.dsh in package.json).

Known Limitations

• The real-host behaviour of the meeting watchdog can only be confirmed by a real-host re-test (inside the shipped suites the heartbeat delay is not controllable, so the branch cannot be triggered deterministically; the guard is therefore a source-level invariant, as stated in the test comments and the commit message).
• TeX installed in a non-standard location is still not auto-discovered: the default probe covers PATH and the documented common install roots.
• The shipped suites do not cover host-level behaviour (for example "the caller is dropped after an interrupt").
• A root-agent question blocks the whole institute (operational note, product unchanged): state the problem and acceptance criteria up front for unattended runs. See slv-playbook.md §5.

Verification

• node tests/run-tests.mjs — all green (106 items: 44 suites + 62 probes).
• node scripts/release-check.mjs — exits 0; after publishing, add --registry to compare the registry shasum with the local tarball.
• This release's fix: node tests/e2e-v5-round2.test.mjs (★ [real1004-stall], measured passed=436 failed=0) and node tests/v5-institute-fixes.mutants.mjs (19/19 named reds).
• Real-host re-check recipe: install the exact published version (-Pin <version>) and run a trimmed meeting + paper task (no research), then count "late notes" (1067 before the fix) and refused finalisations (863 before the fix) — both should drop sharply.

Dependencies

• No new runtime dependencies; the DSH support window is unchanged.