dsh-vibe-math 2.4.1 → 2.5.0

package/README.en.md

@@ -79,7 +79,7 @@ flowchart TB
     subgraph FW["⚙️ Framework vibe-v5 —— only a medium (middleware), never assigns tasks"]
         M["Message relay · meetings/debates · task board CAS+DAG<br/>m-vote consensus verification · context and liveness · roster and hiring · scheduler"]
     end
-    PROJ["💾 State/&lt;institute&gt;.v5state.json (hardened JSON, authoritative)<br/>11 kinds of events · pure fold applyV5Event · serial writes · recovery = load before read"]
+    PROJ["💾 State/&lt;institute&gt;.v5state.json (hardened JSON, authoritative)<br/>12 kinds of events · pure fold applyV5Event · serial writes · recovery = load before read"]
     FS["📁 Members/&lt;id&gt;/* · Shared/* · Verified/ · Problems/"]
     RULE{{"Truth gate: Boolean unanimity and Boolean votes ≥ m = min(quorumCap, number of registered voting members)"}}
     OFF <-->|"vibe_v5_* / /v5 commands　↔　status / report"| M
@@ -402,7 +402,7 @@ Then start a new session and select **"Vibe Math V2"** / **"Vibe Math V3"** / **
 > - **Choose `vibe-math-v5` (institute system)** if you want:
 >   - **organized self-organization** — like a real research institute, with a **leader (academician)** responsible for decomposition, assignment, prioritization, chairing meetings and supervising progress, but **judgment still belongs to each individual**;
 >   - a **roster that can grow or shrink** — resident researchers + temp workers who can be **autonomously hired/dismissed** (temp workers have no voting rights, suitable for chores such as checking, trial computation and material organization);
->   - an **adjustable consistency threshold** — `m = min(quorumCap, number of voting members)` boolean-consistent votes settle the matter (easier to converge than "unanimity", while **opposing votes still block**, so a minority will not be drowned out by abstentions);
+>   - an **adjustable consistency threshold** — `m = min(quorumCap, number of voting members)` boolean-consistent votes settle the matter (because an opposing vote blocks, the effective threshold is still "every Boolean vote points the same way"; it differs only when fewer than `m` members cast a Boolean vote);
 >   - **zero-token-cost state persistence** — the institute state is written to a hardened JSON file under the institute directory and does not consume member context budget.
 >
 > **⚠️ `vibe-math-v2` and `vibe-math-v3` are the classic architectures; `vibe-math-v4` and `vibe-math-v5` are experimental architectures,** all four are peers and all are selectable; the old `vibe-math-v1` has been removed (this package contains only v2/v3/v4/v5).
@@ -651,11 +651,11 @@ The main agent runs `vibe_math_status` and reports in plain language: the number
 **Step 3 — Tune parameters by conversation (optional)**
 
 ```
-I want it to use weighted voting, with the concurrency limit set to 6.
+I want it to use forced verdict mode, with the concurrency limit set to 6.
 ```
 
-The main agent `vibe_math_set_params {"verdictMode":"weighted-vote","maxParallelThreshold":6}`,
-and asks you whether to `vibe_math_save_settings` to save them.
+The main agent `vibe_math_set_params {"verdictMode":"forced","maxParallelThreshold":6}`,
+and asks you whether to `vibe_math_save_settings` to save them. (`verdictMode` accepts only `flat|forced`; any other value is **not an error** — it is silently reverted to that architecture's default, `flat` in v2 and `forced` in v3.)
 
 **Step 4 — Intervene midway (optional)**
 
@@ -704,7 +704,7 @@ The main agent `vibe_math_status`: in `qs/qs.json`, `q1` has been written back w
 | `directionsPerSolver` | 1 | Total number of directions visible to each solver prompt (1 = only its own direction, no mutual interference; N>1 = its own plus summaries of up to N-1 other active directions) |
 | `verifierCount` | 3 | Number of independent verifiers per verification target |
 | `debateMaxRounds` | 5 | Maximum number of rounds of verification debate (chat group) |
-| `verdictMode` | `flat` | `flat` = balanced mechanism (inconsistency judged 0.5) / `forced` = forced verdict (historical accuracy + rigor weighting) |
+| `verdictMode` | `flat` | `flat` = the **equal-weight mean** of the probabilities the verifiers report (an inconsistency is no longer judged 0.5) / `forced` = forced verdict (weighted by historical accuracy + rigor; accuracy is tracked per model and scored only when the object later receives a boolean verdict) |
 | `reportMode` | `file` | `file` = write a report file / `push` = push a report to the main agent / `both` |
 | `promoteValueThreshold` | 0.7 | A proposition in Propos with "value/criticality" ≥ this value and undecided (0,1) is automatically added to qs.json |
 | `priorityAdjust` | `none` | `none` / `deadend-deprioritize` (deprioritize all dead ends) / `survival-map` (recompute by survival rate) |
@@ -732,12 +732,12 @@ Added/adjusted on top of all v2 parameters:
 
 | Parameter | Default | Description |
 |---|---|---|
-| `verdictMode` | `forced` | v3 first makes a **near-consensus determination** (all results on the same side with a mean ≥0.85/≤0.15 takes the mean), otherwise `forced` = weighted by historical accuracy + rigor / `flat` = balanced (0.5). Fixes the v2 flat problem of misjudging "0.9 vs 1" as 0.5 |
+| `verdictMode` | `forced` | v3 first makes a **near-consensus determination** (all results on the same side with a mean ≥0.85/≤0.15 takes the mean) — this runs BEFORE the mode, so 0.9 vs 1 yields ≈0.95 in either mode; otherwise `forced` takes the **equal-weight mean** of the reported probabilities (no accuracy weighting; strict 1/0 still act as absolute votes), while `flat` rules `0.5` (a disagreement is undecided) |
 | `planningHorizon` | 3 | Maximum number of actions in one plan by the planner agent ("the next n times") |
 | `plannerEnabled` | true | false = fully use the built-in heuristic scheduling (planner agent disabled) |
 | `plannerProvider` / `plannerModel` | empty | Planner agent model route (empty = inherit the root agent) |
 | `plannerPersona` | empty | Persona/requirements injected at the start of the planner agent prompt |
-| `planMinIntervalMs` | 30000 | Minimum interval between two planning calls (milliseconds); ignored when the system is idle and there is work |
+| `planMinIntervalMs` | 30000 | Minimum interval between two planning calls (milliseconds); it applies to **every** planning call (empty plans included, and "idle but still work" — the idle bypass is gone) |
 | `plannerMaxFails` | 3 | The planner agent reaching this number of consecutive failures → automatic degradation to heuristics |
 | `methodKeepIntervalMs` | 0 | Method Keeper periodic consolidation interval (0 = event-driven) |
 | `methodKeepEvery` | 5 | Trigger one consolidation every N accumulated inventions/new propositions awaiting sedimentation |
@@ -825,8 +825,8 @@ Common controls: `vibe_v5_configure` (configure first) → `vibe_v5_start` (star
 - **Four sets of Lean prompt corpora**: [`prompt-corpus-v2/formal-verify-v2.md`](prompt-corpus-v2/formal-verify-v2.md) · [`prompt-corpus-v3/formal-verify-v3.md`](prompt-corpus-v3/formal-verify-v3.md) · [`prompt-corpus-v4/formal-verify-v4.md`](prompt-corpus-v4/formal-verify-v4.md) (each covering off / encourage / **require** / fidelity branches / work rounds / receipt contract; the workspace is normalized to `<WS>` and the VibeMath root to `<VIBEMATH>`)
 - **The persona source text of the four presets**: [`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md) (the prompts the main agent actually receives: which tools, which parameters, which slash subcommands; generated by `audit-persona-surface.test.mjs` and shipped with the package)
 - **Lean formal verification (a contract shared by the four architectures)**: [`docs/formal-verification.md`](docs/formal-verification.md)
-- **Test timing baseline and parallel run recipes**: [`docs/test-timing.md`](docs/test-timing.md) (`node tests/run-tests.mjs` runs all suites in parallel in ≈1.9 min; the probe scripts ≈2.6 min; every runner prints its elapsed time/speedup for choosing a strategy next time)
-- **Static prompt-surface consistency (persona ↔ tool registry ↔ slash command hint/usage)**: [`audit-persona-surface.test.mjs`](tests/audit-persona-surface.test.mjs) (197 assertions, and generates [`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md) for manual review) + [`audit-persona-sensitivity.mjs`](tests/audit-persona-sensitivity.mjs) (11 sensitivity probes) — guarding "every registered tool must appear in the persona / every name in the persona must really be registered / the `prefix` and `text` blocks must match line by line / hint, usage, and the actual branch must agree in all three places"
+- **Test timing baseline and parallel run recipes**: [`docs/test-timing.md`](docs/test-timing.md) (`node tests/run-tests.mjs` runs **every suite AND every probe** in parallel: `TOTAL 56` (38 suites + 18 probes/variants) in ≈3.5 min; which 18 of them ship in the package and which are repository-only is in §1.1 of that document; every runner prints its elapsed time/speedup so the next strategy can be chosen from data)
+- **Static prompt-surface consistency (persona ↔ tool registry ↔ slash command hint/usage)**: [`audit-persona-surface.test.mjs`](tests/audit-persona-surface.test.mjs) (196 assertions, and generates [`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md) for manual review) + [`audit-persona-sensitivity.mjs`](tests/audit-persona-sensitivity.mjs) (11 sensitivity probes) — guarding "every registered tool must appear in the persona / every name in the persona must really be registered / the `prefix` and `text` blocks must match line by line / hint, usage, and the actual branch must agree in all three places"
 - **Mandatory checklist for a full review**: [`AUDIT-CHECKLIST.md`](docs/AUDIT-CHECKLIST.md) (this repository's mandatory audit process; §1.9 specifically checks "whether the tool parameter schema can accommodate them")
 - **Prompt/interaction invariants (all four sets together, re-checkable in one command)**: [`audit-prompt-invariants.mjs`](tests/audit-prompt-invariants.mjs) (157 assertions) — encoding, one by one, "the classes of prompt/tool-surface defects that have really occurred historically" as static invariants (abbreviated tool names, projecting a fidelity defect as 0, `defect` written only in the prompt but not implemented, a receipt contract missing `defect`, passing without a note, wrong field names, the `off` tier still being able to write state in its receipt, uncertain corpora, missing probes, **a tool's closed schema that cannot accommodate the parameters in its own documentation**, **keys that the schema declares at the parameter level but silently drops**). Adding `--self-probe` injects these defect shapes in memory and requires the corresponding invariants to **turn red** while the unmutated control run **stays green** (5/5); the script itself additionally carries six self-checks X5–X8b (the comment scanner must recognize regex literals — including regexes **after a keyword** such as `return /…/` — `//` inside strings must be preserved, stripping comments must not change line structure, and the parse-level criterion that "the four sets of source, with comments stripped, must still parse under `node --check`")
 - **Specification ↔ code traceability (all four sets together)**: [`audit-spec-traceability.mjs`](tests/audit-spec-traceability.mjs) (94 assertions) — tools promised in `实现方案.md`/README must really be registered; the four Lean parameters must be accepted by both the documentation and the code
@@ -840,16 +840,16 @@ Common controls: `vibe_v5_configure` (configure first) → `vibe_v5_start` (star
 - **(DSH ≤ 0.1.6, the directory form)** The installer has **versioned auto-update**: on every DSH start it compares the package version against the record in `<presetRoot>/.vibe-math-installed.json` — **as soon as the version changes (or on the first run of an older installation with no record) it replaces the managed files wholesale, regardless of whether they were modified**; replaced hand-edited originals are first backed up to `<presetRoot>/.vibe-math-backup/<old version>/<preset>/` and listed in the log. Within the same version it **rewrites no files at all** (restarting DSH will not rewrite the presets, nor disturb the generation index it records by mtime), and missing files are restored at any time. To force a full reinstall: delete those four preset directories and restart DSH.
   On DSH ≥ 0.1.7 the installer **writes none of these directories** (the composition rows of the "Installation" section above supply the presets); at startup it only logs that leftover copies can safely be deleted.
 - **On DSH ≤ 0.1.6 these four directories are owned by the installer**: deleting any file in them or the whole directory only leads to it being restored on the next DSH start (which is exactly why the "force a full reinstall" above works). On DSH ≥ 0.1.7 they are neither read nor written any more, so you can simply delete them; to get rid of them entirely, uninstall this package (`dsh plugin --profile <your profile> remove dsh-vibe-math`).
-- A `flat` verdict directly rules `0.5` when the debate is inconsistent (high-confidence disagreement such as 0.9 vs 1 is also misjudged as 0.5 — **v3 has fixed this with the near-consensus rule**); `forced` weights by historical accuracy + confidence.
+- A `flat` verdict takes the **equal-weight mean** of the probabilities the verifiers report when the debate is inconsistent (it is no longer judged `0.5`; the early problem of misjudging a high-confidence disagreement such as 0.9 vs 1 as 0.5 is fixed in v2/v3); `forced` is weighted by historical accuracy + rigor, and accuracy is scored only when the object later receives a boolean verdict.
 - Problems/propositions with `never` priority are **never scheduled**, and do not block strict termination (they count as voluntarily abstaining).
 - The four preset files are mutually independent and can coexist; only one preset can be selected in a given session at a time.
 
 **v3**:
 - **Soft specification rather than zero specification**: the md knowledge base only enforces the 4–7 anchor lines at the head of an object (`- ID/类型/状态/概率/优先级/依赖/...`) and the entry title line (`### 解法/证明/证伪 N｜标题｜概率X｜状态Y`), so that the scheduler can index reliably; the body is entirely free paper-style prose, and the scheduler never parses the body. Manually editing the anchors may cause index drift (the scheduler keeps the last valid index and warns).
-- **The planning agent is an enhancement, not a requirement**: when `plannerEnabled=false` or the planning agent fails repeatedly (`plannerMaxFails`), it automatically falls back to v2-style heuristic scheduling; the `planMinIntervalMs` cooldown takes effect while subagents are in flight (when the system is idle and there is work, it plans immediately).
+- **The planning agent is an enhancement, not a requirement**: when `plannerEnabled=false` or the planning agent fails repeatedly (`plannerMaxFails`), it automatically falls back to v2-style heuristic scheduling; the `planMinIntervalMs` cooldown applies to **every** planning call — empty plans and "idle but still work" included (idleness no longer bypasses the cooldown, which is exactly what amplified the planner's per-tick spinning).
 - **Trust tiers in the method library**: a method card's `可信断言` may only link IDs that have entered `Verified/`; all other content of a method entry (including unverified strategies/intuitions/heuristics) is treated as **empirical reference** and must not be cited as a theorem.
 - **Project lock**: only one session may schedule a given project at a time (starting a second session reports "occupied by session X"); the lock is released automatically on pause/termination/when everything is resolved.
-- **Near-consensus verdicts**: when all verifier results fall on the same side and the mean is ≥0.85/≤0.15, the mean is taken (e.g. 0.9 vs 1 → 0.95); otherwise `forced` weighting / `flat` rules 0.5 — this fixes the problem in v2 where conclusions that were "mathematically correct but formally flawed" were misjudged as uncertain.
+- **Near-consensus verdicts**: when all verifier results fall on the same side and the mean is ≥0.85/≤0.15, the mean is taken (e.g. 0.9 vs 1 → 0.95); otherwise `forced` takes the **equal-weight mean** of the votes and `flat` rules `0.5` — this fixes the problem in v2 where conclusions that were "mathematically correct but formally flawed" were misjudged as uncertain.
 
 **v5**:
 - **The framework never assigns tasks**: this is a hard design boundary, not a feature that has yet to be implemented. The creation and allocation of tasks belong to **institute self-governance**
@@ -857,7 +857,7 @@ Common controls: `vibe_v5_configure` (configure first) → `vibe_v5_start` (star
 - **`≥ m` agreement ≠ mathematically proved**: the quorum only guarantees that "a consistent judgment has been reached within the institute", not that the conclusion is really correct.
   The depth of truth-seeking rests on the members' own derivations and the paper trail in the debate records; objects that do not reach quorum are **kept in the library with their average probability attached** and are not forcibly ruled true or false.
 - **The consistency quorum is not "majority rule"**: any single opposing boolean vote blocks a conclusion, and abstention helps neither truth nor falsehood.
-  To converge more easily, lower `quorumCap` or reduce the number of members with voting rights; for stricter behavior, switch to `quorumMode: "all-unanimous"`.
+  Under this rule (`m = min(quorumCap, number of voting members)`) **as soon as anyone casts a Boolean vote, the threshold is effectively "every Boolean vote points the same way"**: lowering `quorumCap` only lowers how many Boolean votes are required (`m`) and does **not** make convergence easier; for stricter behavior, switch to `quorumMode: "all-unanimous"` (which does not even allow abstentions).
 - **A member whose round never ends will not be forcibly released** (the same boundary as v4): the heartbeat re-arms it every time,
   so the scheduler will not freeze permanently; when manual intervention is needed, use `vibe_v5_fire` (temp worker) or have the institute office add or remove positions.
 - **Meetings and verification are strictly mutually exclusive**: while one is in progress, the other queues/is held. Therefore "concluding an object on the spot in a meeting" first queues,

package/README.md

@@ -79,7 +79,7 @@ flowchart TB
     subgraph FW["⚙️ 框架 vibe-v5 —— 只是媒介（middleware），绝不指派任务"]
         M["消息中继 · 会议/辩论 · 任务板 CAS+DAG<br/>m 票共识验证 · 上下文与活性 · 编制与雇佣 · 调度器"]
     end
-    PROJ["💾 State/&lt;研究所&gt;.v5state.json（加固 JSON 权威源）<br/>11 类事件 · 纯折叠 applyV5Event · 串行写 · 恢复=读前必 load"]
+    PROJ["💾 State/&lt;研究所&gt;.v5state.json（加固 JSON 权威源）<br/>12 类事件 · 纯折叠 applyV5Event · 串行写 · 恢复=读前必 load"]
     FS["📁 Members/&lt;id&gt;/* · Shared/* · Verified/ · Problems/"]
     RULE{{"求真门槛：布尔一致 且 布尔票 ≥ m = min(quorumCap, 在册有表决权人数)"}}
     OFF <-->|"vibe_v5_* / /v5 命令　↔　status / report"| M
@@ -402,7 +402,7 @@ dsh plugin --profile <你的 profile> add github:ChongCyrus/Vibe-Mathematics
 > - **选 `vibe-math-v5`（研究所体系）**，如果你想要：
 >   - **有组织的自组织**——现实研究所那样有**领头人（院士）**负责拆解、分派、定优先级、主持会议、督导进度，但**判断仍归每个人自己**；
 >   - **可增减的编制**——常驻研究员 + 可**自主雇佣/解雇**的临时工（临时工无表决权，适合处理核对、试算、资料整理等杂活）；
->   - **可调的一致性门槛**——`m = min(quorumCap, 有表决权人数)` 票布尔一致即定论（比"全体一致"更容易收敛，同时**反向票仍然阻塞**，少数派不会被弃权淹没）；
+>   - **可调的一致性门槛**——`m = min(quorumCap, 有表决权人数)` 票布尔一致即定论（因为反向票阻塞，实际门槛仍是"全部布尔票同向"；只有投出布尔票的人少于 `m` 时才不同）；
 >   - **零 token 成本的状态持久化**——研究所状态写在研究所目录下的加固 JSON 里，不占成员上下文预算。
 >
 > **⚠️ `vibe-math-v2` 与 `vibe-math-v3` 是经典架构；`vibe-math-v4`、`vibe-math-v5` 是实验性架构，** 四者同级、均可选择；老 `vibe-math-v1` 已被移除（本包仅含 v2/v3/v4/v5）。
@@ -652,11 +652,11 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 **Step 3 — 问答式调参（可选）**
 
 ```
-我想让它用加权投票，并发数设成 6。
+我想让它用强制裁决模式（forced）跑，并发数设成 6。
 ```
 
-主代理 `vibe_math_set_params {"verdictMode":"weighted-vote","maxParallelThreshold":6}`，
-并问你是否 `vibe_math_save_settings` 保存。
+主代理 `vibe_math_set_params {"verdictMode":"forced","maxParallelThreshold":6}`，
+并问你是否 `vibe_math_save_settings` 保存。（`verdictMode` 只接受 `flat|forced`；写其它值**不报错**，会被静默回退成该架构的默认值——v2 是 `flat`、v3 是 `forced`。）
 
 **Step 4 — 中途干预（可选）**
 
@@ -705,7 +705,7 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 | `directionsPerSolver` | 1 | 每个 solver 提示词可见的方向总数（1 = 只看自己方向、互不干扰；N>1 = 自己 + 最多 N-1 个其他活跃方向摘要） |
 | `verifierCount` | 3 | 每个验证对象的独立验证器数量 |
 | `debateMaxRounds` | 5 | 验证辩论（交流群）最大轮数 |
-| `verdictMode` | `flat` | `flat` = 均衡机制（不一致判 0.5）/ `forced` = 强制裁决（历史准确率+严谨性加权） |
+| `verdictMode` | `flat` | `flat` = 取各验证者报告概率的**等权平均**（不一致时不再一律判 0.5）/ `forced` = 强制裁决（按历史准确率+严谨性加权；准确率按模型追踪，且只在对象此后获得布尔裁决时计分） |
 | `reportMode` | `file` | `file` = 写报告文件 / `push` = 推送主代理汇报 / `both` |
 | `promoteValueThreshold` | 0.7 | Propos 中「价值/关键性」≥ 该值且未决(0,1) 的命题自动加入 qs.json |
 | `priorityAdjust` | `none` | `none` / `deadend-deprioritize`（全死路降优先级）/ `survival-map`（按存活率重算） |
@@ -733,12 +733,12 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 
 | 参数 | 默认 | 说明 |
 |---|---|---|
-| `verdictMode` | `forced` | v3 先做**近共识判定**（全部结果同侧且均值 ≥0.85/≤0.15 取均值），否则 `forced`=按历史准确率+严谨性加权 / `flat`=均衡（0.5）。修复了 v2 flat 把"0.9 vs 1"误判成 0.5 的问题 |
+| `verdictMode` | `forced` | v3 先做**近共识判定**（全部结果同侧且均值 ≥0.85/≤0.15 取均值）——这一步先于模式生效，所以 0.9 vs 1 在两种模式下都得到 ≈0.95；否则 `forced` 取各验证者概率的**等权平均**（不再按准确率加权；严格的 1/0 仍以绝对票生效），`flat` 则判 `0.5`（不一致即不确定） |
 | `planningHorizon` | 3 | 规划代理一次计划的最多动作数（"接下来 n 次"） |
 | `plannerEnabled` | true | false = 完全走内置启发式调度（规划代理禁用） |
 | `plannerProvider` / `plannerModel` | 空 | 规划代理模型路由（空 = 继承根代理） |
 | `plannerPersona` | 空 | 注入规划代理提示词开头的人格/要求 |
-| `planMinIntervalMs` | 30000 | 两次规划调用的最小间隔（毫秒）；系统空闲且有工作时忽略 |
+| `planMinIntervalMs` | 30000 | 两次规划调用的最小间隔（毫秒）；对**每一次**规划调用都生效（含空计划、含"空闲但仍有工作"，不再有空闲绕过） |
 | `plannerMaxFails` | 3 | 规划代理连续失败达此值 → 自动降级启发式 |
 | `methodKeepIntervalMs` | 0 | Method Keeper 定时整理间隔（0 = 事件驱动） |
 | `methodKeepEvery` | 5 | 每积累 N 个待沉淀发明/新命题触发一次整理 |
@@ -826,8 +826,8 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 - **四套 Lean 提示词语料**：[`prompt-corpus-v2/formal-verify-v2.md`](prompt-corpus-v2/formal-verify-v2.md) · [`prompt-corpus-v3/formal-verify-v3.md`](prompt-corpus-v3/formal-verify-v3.md) · [`prompt-corpus-v4/formal-verify-v4.md`](prompt-corpus-v4/formal-verify-v4.md)（各自覆盖 off / encourage / **require** / 忠实性分支 / 工作轮 / 回执契约；工作区归一化为 `<WS>`、VibeMath 根为 `<VIBEMATH>`）
 - **四个预设的 persona 原文**：[`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md)（主代理实际收到的提示词：有哪些工具、哪些参数、哪些斜杠子命令；由 `audit-persona-surface.test.mjs` 生成，随包发布）
 - **Lean 形式化验证（四架构共用契约）**：[`docs/formal-verification.md`](docs/formal-verification.md)
-- **测试耗时基线与并行跑法**：[`docs/test-timing.md`](docs/test-timing.md)（`node tests/run-tests.mjs` 并行跑全部套件 ≈1.9 min；探针脚本 ≈2.6 min；每个 runner 都会打印耗时/加速比供下次选策略）
-- **静态提示词面一致性（persona ↔ 工具注册表 ↔ 斜杠命令 hint/usage）**：[`audit-persona-surface.test.mjs`](tests/audit-persona-surface.test.mjs)（197 条断言，并生成 [`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md) 供人工复核）+ [`audit-persona-sensitivity.mjs`](tests/audit-persona-sensitivity.mjs)（11 条灵敏度探针）——守"注册的工具必须在 persona 里出现 / persona 里的名字必须真的注册 / `prefix` 与 `text` 两块逐行一致 / hint、usage、实际分支三处必须一致"
+- **测试耗时基线与并行跑法**：[`docs/test-timing.md`](docs/test-timing.md)（`node tests/run-tests.mjs` 并行跑**全部套件 + 全部探针**：`TOTAL 56`（38 套件 + 18 探针/变体）≈3.5 min；哪 18 项随包发布、哪些仅仓库见该文档 §1.1；每个 runner 都会打印耗时/加速比供下次选策略）
+- **静态提示词面一致性（persona ↔ 工具注册表 ↔ 斜杠命令 hint/usage）**：[`audit-persona-surface.test.mjs`](tests/audit-persona-surface.test.mjs)（196 条断言，并生成 [`prompt-corpus-persona/persona-corpus.md`](prompt-corpus-persona/persona-corpus.md) 供人工复核）+ [`audit-persona-sensitivity.mjs`](tests/audit-persona-sensitivity.mjs)（11 条灵敏度探针）——守"注册的工具必须在 persona 里出现 / persona 里的名字必须真的注册 / `prefix` 与 `text` 两块逐行一致 / hint、usage、实际分支三处必须一致"
 - **全面检查必查清单**：[`AUDIT-CHECKLIST.md`](docs/AUDIT-CHECKLIST.md)（本仓库的强制审计流程；§1.9 专门查"工具参数 schema 收不收得下"）
 - **提示词/交互不变式（四套一起，可一键复核）**：[`audit-prompt-invariants.mjs`](tests/audit-prompt-invariants.mjs)（157 条断言）——把"历史上真实发生过的提示词/工具面缺陷类别"逐条编码成静态不变式（缩写工具名、把忠实性缺陷投成 0、`defect` 只写在提示词里没实现、回执契约缺 `defect`、无 note 放行、字段名错、`off` 档回执仍能写状态、语料不确定、探针缺失、**工具的封闭 schema 收不下它自己文档里的参数**、**schema 声明了参数层却静默丢弃的键**）。加 `--self-probe` 会在内存里注入这些缺陷形状，要求对应不变式**变红**、未变异的对照跑**仍为绿**（5/5）；脚本自身另带 X5–X8b 六条自检（注释扫描器必须认正则字面量——包括 `return /…/ ` 这种**关键字后面**的正则——字符串里的 `//` 必须保留、抹注释不改变行结构，以及"四套源码抹掉注释后仍必须能被 `node --check` 解析"这条解析级判据）
 - **规格 ↔ 代码可追溯（四套一起）**：[`audit-spec-traceability.mjs`](tests/audit-spec-traceability.mjs)（94 条断言）——`实现方案.md`/README 里承诺的工具必须真的注册；四个 Lean 参数必须同时被文档与代码接受
@@ -841,16 +841,16 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 - **（DSH ≤ 0.1.6 的目录形式）** 安装器带**版本化自动更新**：每次 DSH 启动时对比包版本与 `<presetRoot>/.vibe-math-installed.json` 记录——**版本一变（或无记录的老安装首次运行）就整体替换受管文件，不看它是否被改过**；被替换的手改原文先备份到 `<presetRoot>/.vibe-math-backup/<旧版本>/<preset>/` 并在日志里列出。同一个版本内**不重写任何文件**（重启 DSH 不会改写 preset、也不会扰动它按 mtime 记的生成代际），缺失文件随时补回。想强制全量重装：删除那四个 preset 目录后重启 DSH。
   在 DSH ≥ 0.1.7 上安装器**不写这些目录**（预设由上文「安装」里的组合行提供），启动时只会提示旧副本可安全删除。
 - **在 DSH ≤ 0.1.6 上，这四个目录由安装器负责**：删掉其中任一文件或整个目录，只会在下次启动 DSH 时被补回（这正是上面"强制全量重装"可行的原因）。在 DSH ≥ 0.1.7 上它们不再被读取、也不再被写入，直接删掉即可；想彻底不要本包的预设，就卸载本包（`dsh plugin --profile <你的 profile> remove dsh-vibe-math`）。
-- `flat` 裁决在辩论不一致时直接判 `0.5`（高置信分歧如 0.9 vs 1 也会被误判 0.5——**v3 已用近共识规则修复**）；`forced` 按历史准确率+置信度加权。
+- `flat` 裁决在辩论不一致时取各验证者概率的**等权平均**（不再直接判 `0.5`；早期把高置信分歧 0.9 vs 1 误判为 0.5 的问题已在 v2/v3 修复）；`forced` 按历史准确率+严谨性加权，且准确率只在对象此后获得布尔裁决时计分。
 - `never` 优先级的问题/命题**永不调度**，且不阻塞严格终止（视为主动弃权）。
 - 四个 preset 文件互相独立、可共存；同一会话同时只能选一个预设。
 
 **v3**：
 - **软规范而非零规范**：md 知识库只强制对象头部的 4~7 行锚点（`- ID/类型/状态/概率/优先级/依赖/...`）与条目标题行（`### 解法/证明/证伪 N｜标题｜概率X｜状态Y`），供调度器可靠索引；正文完全自由论文式叙述，调度器从不解析正文。手工编辑锚点可能导致索引漂移（调度器会保留上次有效索引并告警）。
-- **规划代理是增强而非必需**：`plannerEnabled=false` 或规划代理连续失败（`plannerMaxFails`）时自动回退 v2 式启发式调度；`planMinIntervalMs` 冷却在有在途子代理时生效（系统空闲时有工作则立即规划）。
+- **规划代理是增强而非必需**：`plannerEnabled=false` 或规划代理连续失败（`plannerMaxFails`）时自动回退 v2 式启发式调度；`planMinIntervalMs` 冷却对**每一次**规划调用生效——包括空计划和"空闲但仍有工作"的情形（空闲不再绕过冷却，那正是规划代理每 tick 空转的放大器）。
 - **方法库可信分层**：方法卡的 `可信断言` 只允许链接已进 `Verified/` 的 ID；方法条目的其余内容（含未验证的策略/直觉/启发式）一律视为**经验参考**，不得当定理引用。
 - **项目锁**：同一项目同一时刻只允许一个会话调度（第二个会话启动会提示"被会话 X 占用"）；锁在暂停/终止/全部解决时自动释放。
-- **近共识裁决**：全部验证器结果同侧且均值 ≥0.85/≤0.15 时取均值（如 0.9 vs 1 → 0.95），否则 `forced` 加权 / `flat` 判 0.5——修复了 v2 中"数学上正确但形式有瑕疵"的结论被误判为不确定的问题。
+- **近共识裁决**：全部验证器结果同侧且均值 ≥0.85/≤0.15 时取均值（如 0.9 vs 1 → 0.95），否则 `forced` 取各票的**等权平均**、`flat` 判 `0.5`——修复了 v2 中"数学上正确但形式有瑕疵"的结论被误判为不确定的问题。
 
 **v5**：
 - **框架绝不指派任务**：这是设计上的硬边界，不是尚未实现的功能。任务的产生与分配属于**所内自治**
@@ -858,7 +858,8 @@ v5 的完整架构（含成员生命周期、一轮时序、共识状态机、
 - **`≥ m` 一致 ≠ 数学上已证明**：门槛只保证"所内达成了一致判断"，不保证结论真的正确。
   求真的纵深靠成员自己的推导与辩论录留痕；未达门槛的对象会**留库附平均概率**，不会被强行判真判假。
 - **一致性门槛不是"少数服从多数"**：任何一张反向布尔票都阻塞定论，弃权既不帮真也不帮假。
-  要更容易收敛就调低 `quorumCap` 或减少有表决权人数；要更严格可切 `quorumMode: "all-unanimous"`。
+  因此在本口径下（`m = min(quorumCap, 有表决权人数)`）**只要有人投布尔票，门槛就等于"全部布尔票同向"**：
+  调低 `quorumCap` 只是把"至少几张布尔票"降到 `m`，**不会让收敛更容易**；要更严格可切 `quorumMode: "all-unanimous"`（它连弃权都不允许）。
 - **一个回合永不结束的成员不会被强行释放**（与 v4 同一边界）：心跳每次都会重新武装，
   所以调度器不会永久冻结；需要人工介入时用 `vibe_v5_fire`（临时工）或所办增删编制。
 - **会议与验证严格互斥**：一方进行中，另一方排队/暂存。因此"在会议上当场定论一个对象"会先排队，

package/cordis.patch.yml

@@ -86,6 +86,9 @@
                 Key rules to remember when reporting: a problem is "solved" when one of its solutions reaches
                 正确概率 = 1; a proposition reaches 布尔估计 = 1/0 when a proof/refutation in its lists reaches
                 正确概率 = 1; Propos propositions with 价值/关键性 ≥ promoteValueThreshold auto-promote to qs.json.
+                Closure rule (identical across the presets): the 1/0 values above are written ONLY by the scheduler,
+                after ≥2 independent verifier agents settle the object (one lone reviewer can never conclude), and
+                Verified/ is written by that same scheduler path — an agent's own self-reported 1 is never a conclusion.
                 When the user asks about progress, call vibe_math_report and summarize in plain language.
 
                 Configuration highlights (all tunable via vibe_math_set_params / the settings file):
@@ -329,7 +332,8 @@
                 - vibe_math_add_problem {id, description, priority, dependencies?} — add a problem (creates Problems/<id>.md).
                 - vibe_math_add_proposition {id, 概述, 概率, 优先级, 价值/关键性, 分类} — add a proposition (Propos/<分类>/<id>.md).
                 - vibe_math_list_propositions — list the proposition knowledge base.
-                - vibe_math_start / vibe_math_resume — start / resume the scheduler (resume = continue after checkpoint/restart).
+                - vibe_math_start / vibe_math_resume — start / resume the scheduler (resume = continue after checkpoint/restart);
+                  both accept override=true to take the project lock over from another session (only after you verified that session is gone).
                 - vibe_math_status / vibe_math_report — read status / full progress report (report writes Progress_Logs/report.json + Logs/报告.md).
                 - vibe_math_pause / vibe_math_abort — pause / abort.
                 - vibe_math_set_mode {mode: manual|auto} — switch manual / auto (manual gates: 计划审批 / 裁决 / 方法晋升).
@@ -342,7 +346,8 @@
                 - vibe_math_plan {force?} — show queued plan / last plan, or force a planning round.
                 - vibe_math_index — rebuild State/index.json from the Markdown knowledge base.
                 - vibe_math_method_add / vibe_math_method_list — manually add / list method cards (Methods/ + global).
-                - vibe_math_lock_status — project lock occupancy.
+                - vibe_math_lock_status — project lock occupancy (the lock is a renewed lease: while a session is running it never expires; taking it over needs override=true, or happens automatically once the owner's lease has clearly expired after a crash).
+                  Only ONE session may run a project: a second session is refused with PROJECT_LOCKED instead of silently writing the same knowledge base in parallel.
                 - vibe_math_new_project / vibe_math_set_project / vibe_math_list_projects — per-project folders.
                 - vibe_math_list_decisions / vibe_math_decide {id, action: approve|reject|override, verdict?} — resolve manual decisions.
                 - vibe_math_list_agents / vibe_math_message_agent / vibe_math_interrupt_agent — inspect / steer / interrupt subagents.
@@ -350,14 +355,20 @@
                   verification (execute / archive / list the reuse library). The scheduler's child agents
                   use them too; they work in every mode.
 
-                A /vibe slash command mirrors the main controls (/vibe start|resume|pause|abort|status|report|mode
+                A /vibe slash command mirrors the main controls (/vibe start [override]|resume [override]|pause|abort|status|report|mode
                 <auto|manual>|setup|save|add <id> <desc>|add-proposition <id> <概述>|list-propositions|methods|index|plan|lock|...).
                 Data survives restarts via vibe_math_resume.
 
                 Key rules when reporting: a problem is "solved" when one of its 解法候选 entries reaches 概率 = 1
                 (the scheduler writes Verified/问题/<id>.md); a proposition reaches 已验证·真/假 when a 证明/证伪
-                entry reaches 概率 = 1 (Verified/命题/<id>.md); Propos propositions with 价值/关键性 ≥
-                promoteValueThreshold auto-promote into Problems/ as "判断下述命题是否成立：<命题>" (verification
+                entry reaches 概率 = 1 (Verified/命题/<id>.md).
+                Closure rule (identical across the presets): those 概率 = 1 values are written ONLY by the scheduler,
+                after ≥2 independent verifier agents settle the object (a lone reviewer can never conclude; see
+                minVotes), and Verified/ is written by that same scheduler path — a self-reported 1 is never a
+                conclusion. Propos propositions whose 价值/关键性 ≥
+                promoteValueThreshold (write that anchor on the proposition card, or report it as lemmas[].价值/关键性 in
+                vibe_math_sync_meta — default 0.5 never promotes) and that are still undecided
+                auto-promote into Problems/ as "判断下述命题是否成立：<命题>" (verification
                 results sync back to the source proposition); a solver-reported sub-question q_sub registers THREE
                 objects (q_sub problem + judge problem + p-tmp temporary-assumption proposition) with full 来源与动机.
 
@@ -626,14 +637,24 @@
                     (execute / archive / list the reuse library). Residents use them too; they work in every mode.
                   - vibe_v4_formal_report — human-readable Lean formal-verification mirror (mode, Lean-passed
                     objects, recorded blockers, formalization TODO, library paths).
-                  - vibe_v4_prompts {which: brainstorm|normal|heartbeat|verify|coreRules, member?, target?, stage?} — read the exact prompt text a resident would receive (prompt auditing; prompt text is the product).
-                A /v4 slash command mirrors the main controls (configure|start|resume|pause|abort|status|report|message <to|all> <content>|meeting|members|add|remove|set).
+                  - vibe_v4_prompts {which: brainstorm|normal|heartbeat|meeting|verify|coreRules, member?, target?, stage?} — read the exact prompt text a resident would receive (prompt auditing; prompt text is the product).
+                A /v4 slash command mirrors the main controls (configure|start [problem]|resume|pause|abort|status|report|message <to|all> <content>|meeting|members|add|remove|set <k=v>...).
 
                 TRUST RULE: only Verified/ (and Propos/ entries marked 已验证·真/假) are absolutely
                 trustworthy; everything else — unverified resident claims, Progress/, Method claims —
                 is experiential reference. A proposition / method / theory only reaches Verified/ when
                 ALL residents unanimously agree true (or all agree false); otherwise it stays in its
-                library with a probability estimate.
+                library with a probability estimate. Closure rule (identical across the presets): the framework
+                writes the Verified/ card — a resident's own claim never does — and what counts as a vote is
+                stated once in the VOTE CONTRACT below.
+
+                VOTE CONTRACT: a verification vote is a probability of truth in [0,1], but only an EXACT
+                1 (absolutely true) and an EXACT 0 (absolutely false) count as a vote. Anything strictly
+                between (0.9, 0.95, 0.5 …) is an ABSTENTION carrying that probability estimate: the run
+                does NOT conclude, and the group's MEAN probability is written back to the object's card
+                so it stays in its library with a probability. The same absoluteness applies to the
+                meeting stop-vote: every resident on the roster must answer `voteSolved:true` — one
+                abstention keeps the run alive.
 
                 LEAN FORMAL VERIFICATION (formalVerify, a tunable parameter):
                   - 'off' (default, no extra requirement) | 'encourage' (the residents decide by implementation
@@ -912,8 +933,10 @@
                   Shared/TaskBoard.md (human view), Shared/Chat/<day>.md,
                   Shared/Meetings/<id>.md, Shared/Debates/<target>.md,
                   Problems/<id>.md, Verified/<kind>/<id>.md (read-only; m-vote only),
-                  State/ (a human-readable MIRROR only — the authoritative state is the
-                  session log projection; never hand-edit State/).
+                  State/ (the AUTHORITATIVE state is State/<institute>.v5state.json in this
+                  directory; it is a hardened JSON file, never a host session log. Everything
+                  else here and elsewhere in the tree is a human-readable MIRROR — never
+                  hand-edit any of it).
 
                 Main controls (recommended flow: configure FIRST, then start):
                   - vibe_v5_configure {project?, institute?, problem?, params?} — create/configure the institute WITHOUT starting it.

package/docs/AUDIT-CHECKLIST.md

@@ -166,8 +166,14 @@ v2/v3/v4/v5 是**同构实现**（同一份契约、四份独立代码，刻意
 ### 1.9 工具参数 schema：文档写了 ≠ 工具收得下
 
 §1.6 查的是"persona 里有没有写"，这一节查**工具的参数 schema 收不收得下**。本仓库所有工具 schema 都由
-`objParams` 收口，并以 `additionalProperties:false` **关闭**：schema 没列出的键会被任何遵守 schema 的
-provider **直接拒绝**。于是"提示词/规格/状态表都写着这个参数"完全可以在**功能根本打不开**的同时全绿。
+`objParams` 收口，并以 `additionalProperties:false` **关闭**。**但"关闭"的强度取决于注册路径**（v3 审计 M5
+的更正，Round B 复核扩大）：只有走宿主 `defineTool` 的工具，宿主才会用该 schema 校验实参、把未列出的键
+**直接拒绝**；`tools.register`（裸 `ToolDefinition`）只把 `parameters` 生成给模型的 schema，宿主**从不校验实参**，
+`additionalProperties:false` 在那里只是给模型看的**文档**。**四套预设全部走裸注册**
+（`grep defineTool vibe-math-v{2,3,4,5}.js` → 0 命中；`tools.register` 各 1–3 处），所以"schema 与参数层不一致
+就静默成功（fail open）"对**每一套**都成立，不只是 v3。最典型的反例是 `vibe_math_setup` 公布 `mode`、而
+`vibe_math_set_params` 的 schema 里一度没有它：模型看不到这个键，写给它却仍会被 `sanitizeParams` 收下
+（该键已补进 schema，两条注册路径的键集现在一致）。
 
 - [ ] 每个"可调参数"是否同时出现在：**设置工具的参数 schema**、发现面（`vibe_math_setup` 返回的
       `PARAM_SCHEMA` / v4·v5 的 status 参数表）、`<vibe root>/*setting.json` 模板、以及 persona 参数表？
@@ -190,18 +196,30 @@ provider **直接拒绝**。于是"提示词/规格/状态表都写着这个参
 
 测试和脚本"在我机器上是绿的"不等于它们能跑。**真实事故（2.3.13 归档时发现）**：5 个测试/脚本把
 本机绝对路径写死在源码里（`D:/wd/vibemath开发/...`、`C:/Users/admin/...`），其中
-`audit-fuzz-helpers.mjs` **是随包发布的**——对任何用户都跑不了，而仓库里从来没人发现，
-因为作者机器上它恰好是对的。
+`audit-fuzz-helpers.mjs` 当时**是随包发布的**——对任何用户都跑不了，而仓库里从来没人发现，
+因为作者机器上它恰好是对的。（那些路径后来都改成由 `import.meta.url` 推导；Round B 复核：`tests/`
+下已经没有任何真实的写死盘符路径，而且 `audit-fuzz-helpers.mjs` **现在也不再随包发布**——
+它不在 `package.json` 的 `files` 里，也不在 `npm pack --dry-run --json` 的清单里。）
 
 - [ ] 每个文件路径是否都由 `import.meta.url` / `npm root -g` / 环境变量**推导**出来，
       而不是写死盘符或用户名？（推导不出时应当**报出所有试过的候选路径**，而不是静默跳过。）
 - [ ] 文件被移动/改名后，**引用它的每一处**是否都跟着改？（包括：`package.json` 的 `files`、
       `run-tests.mjs` 这类 runner、守卫里按名字读套件的表、文档里的命令行、README 链接。）
+- [ ] 新增/改名 `tests/` 下的脚本后，确认 **`run-tests.mjs` 真的收集了它**：runner 收集 `tests/` 下
+      **所有** `.mjs`（`*.test.mjs` 是套件，其余是探针），只有登记进 `NEEDS_ARGS` 的脚本才会被跳过，
+      而那个登记会在**开发检出**（存在 `.git`）里被逐项校验——条目失效就直接报错退出，所以改名不能
+      悄悄把一条守卫踢出发布门禁。参数变体（同一脚本跑多次，如 `audit-registration.mjs` 四个预设）走
+      `VARIANTS`。
 - [ ] 命令行/相对链接是否需要**从仓库根执行才成立**？写清楚起点（如 `node tests/run-tests.mjs`），
       别留一个照抄就报"找不到文件"的用法行。
 - [ ] **搬动之后必须有不依赖"跑绿了"的证据**：① 每个套件的断言条数与搬动前**逐项相同**；
       ② 建一个 `git worktree` 拿改动前的检出，同一套件在两种布局各跑一遍，归一化路径/临时目录/耗时后
       **逐行比对**（`_oneoff/layout-invariance.mjs`）；③ 相对链接扫描 0 失效（`_oneoff/scan-links.mjs`）。
+- [ ] 讲"全套件 / 门禁 / 多少次全绿"时，是否区分了**随包发布面**与**仓库**？`package.json` 的 `files`
+      只发 `tests/` 的 **18 项**（其中 `.test.mjs` **10** 个），完整门禁（`node tests/run-tests.mjs`，
+      Round B 起 **56 项 = 38 套件 + 18 探针/变体**）只在开发检出里成立。两边的清单见
+      `docs/test-timing.md` §1.1；发布物里的 runner 会把缺失/跳过项**打印出来**（不会静默少跑），
+      所以"安装用户照文档跑得到全套件"这类说法必须避免。
 - [ ] **runner 本身也要跑一遍**：直接跑套件通过 ≠ 并行 runner 通过（2.3.13 就出现过
       "26 个套件直接跑全绿、`run-tests.mjs` 因为按裸文件名 spawn 而全红"）。发布门禁必须包含 runner。
 
@@ -353,8 +371,9 @@ provider **直接拒绝**。于是"提示词/规格/状态表都写着这个参
 ## Upgrade
 ```
 
-- [ ] 两个 H1 都在；**英文那份的 H1 必须含 `English`**（发布门禁阶段 1c 会检查）；
-- [ ] 中文部分 ≥ 150 个汉字（门禁同阶段检查），英文部分不得残留未翻译段落；
+- [ ] 两个 H1 都在；**英文那份的 H1 必须含 `English`**（人工检查：本仓库**没有任何 CI 工作流**，
+      `scripts/` 只有 `build-preset-rows.mjs`，不存在所谓的"发布门禁阶段 1c"）；
+- [ ] 中文部分 ≥ 150 个汉字（同样人工检查），英文部分不得残留未翻译段落；
 - [ ] 每条变更都回答"**对使用者有什么影响**"——不写内部编号、探针名、断言条数；
 - [ ] 「兼容性」明确回答：四套预设的字节/行为是否变化、参数与工具面是否变化；
 - [ ] 「升级」给出可直接复制的命令（`npm i dsh-vibe-math@latest`；若 profile 里把版本钉死了，

package/docs/COMPAT-AUDIT-ROUND3.md

@@ -73,7 +73,22 @@ DSH 0.1.7 起 agent preset **只**由组合行声明；`$DSH_HOME/.agent-presets
 
 宿主的 `CommandResult` 联合是 `{kind:'success',text?} | {kind:'error',text}`；四套预设的 `/vibe`、`/v4`、`/v5`
 在**所有失败路径**上都返回 `{kind:'success', text:'{"ok":false,…}'}` → UI 永远显示不出"这条命令被拒绝了"。
-修复：按 dispatch 结果自身的 `ok:false` 返回 `kind:'error'`（该联合在 0.1.5 上同样存在，故旧线安全）。
+
+**更正（据实收窄）：这条修复不是一次"全仓"修复，而是分两批落地的，本节初稿把"打算做的"写成了"已经做完的"。**
+
+- **2.4.0（`cc0c220`）只改了 v2 与 v3**（`vibe-math-v2.js` / `vibe-math-v3.js` 的命令 handler）。当时 v4 与 v5 的
+  handler 仍然无条件 `return {kind:'success', text:JSON.stringify(r,null,2)}`（各自的"无 session"分支也返回
+  `kind:'success'`）——所以"四套都修了"在 2.4.0 这个提交上是不成立的。
+- **v4 与 v5 在 round-A（`8ec5a9c`）才各自补上同一形状**：v4 的命令 handler 改成
+  `const failed = !r || r.ok === false || (r.ok === undefined && !!r.error)` → `{kind: failed?'error':'success'}`；
+  v5 改成 `const failed = r !== null && typeof r === 'object' && r.ok === false` → 同样三元返回，并把"无 session"
+  分支一并改成 `kind:'error'`。判据：`git log -G "kind:\s*'error'" -- vibe-math-v4/vibe-math-v4.js`（及 v5 同）
+  **只命中 `8ec5a9c`**；round-C（`4eb3175`）与 round-D（`0dc6ee3`）都没有再碰这两个 handler（前者的 v4 hunk 在
+  1354/1854 行，后者的 v5 hunk 止于 4564 行，均在命令 handler 之外）。
+- **仓库内的回归证据是 round-C（`4eb3175`）补进来的**：`tests/host-failure-paths.test.mjs` 在该提交首次入库，
+  它断言 `/v4 set` 未知键、`/v4` 未知子命令与 `/v5` 未知子命令都返回 `kind:'error'`。round-D（`0dc6ee3`）只是
+  把"本节仍把 `/v4` 的修复说成全仓，需据实收窄"记进了待办，没有改代码。
+- 该联合在 0.1.5 上同样存在，故旧线安全。也就是说：**"四套最终一致"成立，"四套一次修完"不成立**。
 
 ### G-9 文档、镜像与工具描述与实现不符
 

package/docs/release-notes/RELEASE-NOTES-2.5.0.md

@@ -0,0 +1,173 @@
+# dsh-vibe-math 2.5.0 — 发布说明（中文）
+
+> 上一版：2.4.1。本版是**深度审计后的修复版本**：四套预设的提示词、工具面、组合声明与数据格式未做破坏性变更，
+> 但修复了多处"看起来能跑、实际目标功能不成立"的缺陷（含三处会丢数据或让运行停摆的问题），并新增多项守卫。
+
+---
+
+## 概览
+
+- 本轮对 v2/v3/v4/v5 四套预设 + 安装器 + 全部守卫做了多轮独立审计与交叉证伪（审计结论：**所有宿主接口/工具/命令调用在 DSH 0.2.0-rc.2 上均存在且调用形状正确**；问题集中在目标功能与内部逻辑）。
+- 修复了 **3 处会造成实际损害的问题**：v5 的状态文件会被空研究所覆盖、v2 的验证日志文件名可越出项目目录、v4 的邮箱投递失败会让调度器永久停摆。
+- 还修掉 **v4 的"真实 /compact"从未生效**、**v3 空闲时每 tick 空转派生 planner**、**v3 单票即可写入 `Verified/`**、**v5 同 tick 并发提议会丢提案**等。
+- 在本机 DSH 0.2.0-rc.2 上做了**真实宿主端到端验证**：四个预设正常注册（order 20–23）、零激活警告、组合行线上不再写旧的 `.agent-presets` 目录。
+- 测试与守卫大幅加强：全仓 **38 套件 + 18 探针**全部纳入自动门禁（此前有若干探针未被收集），并修正了一处测试抖动与一处语料非确定性。
+
+## 变更
+
+### 1. 数据安全与"运行不卡死"
+
+- **v5 状态文件不再被覆盖**：状态后端改为**按路径锁存**，读取未成功即**拒绝写入**；对已存在的研究所改为**采纳**而非重置。此前会话里第一次工具调用会锁存默认路径，随后指定研究所的配置会把它覆盖成空研究所。
+- **v2 路径逃逸**：验证日志文件名改用 `safeId` 归一化，`q.id="x/../../../../pwn"` 这类输入不再能把文件写出项目目录。
+- **v4 调度器停摆**：邮箱投递失败的两条路径（立即投递失败、brainstorm 阶段无处重投）都会重排队并在后续 pass 重试，不再出现"心跳已清、永不重试"的静止状态。
+- **v5 队列竞态**：验证队列的追加与取头改为在事件折叠内完成，同一 tick 内两个成员同时提议不再丢提案。
+
+### 2. 目标功能真正可达
+
+- **v4 的 `/compact`**：按 `subagent/start` 载荷认领活动 Agent，真实压缩调用（`compactNow`）首次真正生效。
+- **v2 的"判断命题"链**：为判断问题补种可验证的解法入口，使"问题 → 判断问题 → 验证 → 收口"主线真正跑通。
+- **v3 的空转**：空闲时不再每个 tick 派生 planner（实测由 41 次降到 13 次，并首次出现正常终止）。
+- **v3 的项目锁**：改为**定时器续租**（此前的续租写在 tick 内，单次慢 tick 就会让别的会话无 override 夺锁）。
+- **v3 的最小票数**：要求达到承诺的验证者数量、且**至少 2 票**才裁决，单票不再能把命题写成"已验证·真"。
+
+### 3. 语义与表述对齐（含提示词面）
+
+- **v4 验证契约**：人设与工具描述明确"**只有精确的 1 / 0 是绝对真/绝对假**；介于 0 与 1 之间是模型认为的概率，只有全 1 或全 0 才算全体一致，否则取**平均概率**更新布尔估计"（代码分支未改，此前示例与说明不足）。
+- **v5 的 m-vote 表述**：文档/人设/README 改为与计票代码一致——"**至少 m 张布尔票且没有任何一张反对票**，默认名册下退化为全体一致"，并删除"比全体一致更易收敛"这一**不实**说法。
+- **四套同级**：v2/v3 改称"经典"、v4/v5 为"实验性"，npm 描述、选择器描述、人设文本、v3 设计文档、安装段顺序全部对齐（此前只有 README 改了）。
+- **v2 的侧信道**：推送汇报的消息来源由未声明的 `kind:'plugin'` 改为已声明的 `kind:'user'`；`solver/verifierMaxToolCalls` 改为如实标注为"仅提示、不强制"。
+- **v5 的调用者授权**：`isOffice` 不再把空身份当所办；无法识别身份的成员调用一律拒绝（此前一个幽灵身份可以改优先级、甚至删掉真实任务）。
+- **v5 的工具描述**：补齐会议 `kind` 枚举与 `target`、说明 `hire.term`。
+
+### 4. 安装器、声明与守卫
+
+- **形态判定改能力差**：旧线上"组合行 registry 包"根本不存在（已用 npm 元数据与包内容证实：0.1.6-alpha.2 无 `@deepseek-ai/dsh-agent-preset`，0.1.7-rc.2 才有），判定改用 `register` 能力，并新增可证伪守卫（把该行改回 `list` 只有新守卫会红）。
+- **声明行补宿主 group 标记**：8 处嵌套 `!!js` 现在以表达式形式保留给 preset 作用域，而不是在声明域被就地求值。
+- **安装器更安全**：preset 文件改**原子写**；删除 preset 前**先备份**用户改动过的文件；**读不出来**的文件不再被无备份删除；状态文件内容不变则跳过写、写失败会告警、`listBundles` 有超时兜底。
+- **守卫扩面**：`run-tests` 现在收集**全部**测试与探针（此前有 2 套红探针被静默漏收），修正了一处测试抖动（v4 会议用例）与一处语料非确定性（v3），并新增安装器/声明/命令失败路径/宿主失败路径等守卫。
+
+## 兼容性
+
+- **DSH 支撑窗口不变**：`0.1.2-alpha.4 … 0.2.0-rc.2`；组合行（≥0.1.7）与目录（≤0.1.6）两条线都保留，且新增能力差判定后旧线不会再被误判。
+- **数据与目录结构不变**：`State/<研究所>.v5state.json`、项目布局、镜像文件位置均照旧；无需迁移。
+- **提示词**：除"v4 验证契约/闭包规则说明、v5 m-vote 表述、v2/v3/v4 人设补一行一致的闭包规则、v5 工具描述补字段"之外未改动；语料已按当前四套重新生成。
+- **参数默认值**：未变（v2 每个待验证对象最多 `max(2, min(verifierCount, maxParallelThreshold-2))` 个验证者，默认配置下为 2；工具调用上限仍为提示性）。
+
+## 升级
+
+```sh
+npm i dsh-vibe-math@latest
+```
+
+桌面版在「设置 → 插件」里更新；命令行 profile 用 `dsh plugin --profile <你的 profile> add dsh-vibe-math@latest`，随后重启 DSH。
+
+---
+
+# dsh-vibe-math 2.5.0 — Release Notes (English)
+
+> Previous: 2.4.1. This release is the outcome of a deep, independently cross-checked audit: the four
+> presets' prompts, tool surface, composition declarations and data formats have no breaking change, but
+> several defects that "looked like they worked while the intended function did not" are fixed — three
+> of them could lose data or stall a run — and the guard suite is substantially stronger.
+
+---
+
+## Overview
+
+- Multiple audit rounds with independent falsification covered the four presets, the installer and every
+  guard. Verdict: **every host interface, tool, command and service call exists on DSH 0.2.0-rc.2 with
+  the correct call shape**; the defects were in goals and internal logic.
+- Three fixes prevent real damage: v5's state file being overwritten by an empty institute, v2's
+  verification-log filename escaping the project directory, and v4's scheduler stalling forever after a
+  failed mailbox delivery.
+- Also fixed: v4's "real /compact" never firing, v3 spawning a planner on every idle tick, v3 finalising
+  a proposition on a single vote, and v5 losing a proposal when two members proposed in the same tick.
+- Verified on a **real DSH 0.2.0-rc.2 host**: the four presets register (orders 20–23) with zero
+  activation warnings, and the row line no longer writes the legacy `.agent-presets` directory.
+- Guards grew from 30 to **38 suites + 18 probes**, all collected by the gate (several probes used to be
+  silently skipped); one flaky test and one non-deterministic corpus were fixed.
+
+## Changes
+
+### 1. Data safety and "no run can stall"
+
+- **v5 state file can no longer be overwritten**: the backend latches **per path** and **refuses to
+  write** when the read never succeeded; an existing institute is **adopted** instead of reset. The
+  first tool call of a session used to latch the default path, after which configuring a named institute
+  overwrote it with an empty one.
+- **v2 path escape**: the verification-log filename is normalised with `safeId`, so an id such as
+  `x/../../../../pwn` can no longer write outside the project directory.
+- **v4 scheduler stall**: both failed-mailbox paths (immediate delivery rejection, and the brainstorm
+  phase where nothing retried) now re-queue and retry on a later pass.
+- **v5 queue race**: queue append and head-take now happen inside the event fold, so two members
+  proposing in the same tick can no longer lose a proposal.
+
+### 2. The intended function actually happens
+
+- **v4 `/compact`**: the live Agent is claimed from the `subagent/start` payload, so the real
+  `compactNow` finally fires.
+- **v2's judge-problem chain**: the judge problem now carries a verifiable solution entry, so
+  problem → judge problem → verification → closure really runs.
+- **v3 idle spawning**: no planner is spawned on every idle tick any more (measured 41 → 13) and a
+  normal termination was observed for the first time.
+- **v3 project lock**: renewal moved to a timer (it used to live inside `tick()`, so one slow tick let
+  another session steal the lock without an override).
+- **v3 minimum votes**: the promised verifier count and at least two votes are now required before a
+  verdict; a single vote can no longer write `Verified/`.
+
+### 3. Semantics and wording aligned (including prompt-facing text)
+
+- **v4 verdict contract**: the persona and tool description now state that **only an exact 1 / 0 is
+  absolutely true/false**; anything strictly between is the model's probability of truth, only an all-1
+  or all-0 set is unanimity, and otherwise the **mean** updates the boolean estimate (the code path is
+  unchanged; the example and explanation were insufficient).
+- **v5 m-vote**: docs, persona and README now match the tally — "**at least m boolean votes with no
+  opposing vote**, which degenerates to unanimity on the default roster" — and the false claim that it
+  "converges more easily than unanimity" is gone.
+- **Four peers**: v2/v3 are "classic", v4/v5 "experimental" — the npm description, picker descriptions,
+  persona text, v3's design doc and the install-section order were all brought in line (previously only
+  the READMEs).
+- **v2 side channels**: the push report's message source is the declared `kind:'user'`; the
+  `solver/verifierMaxToolCalls` settings are now honestly described as advisory.
+- **v5 caller authorisation**: `isOffice` no longer treats an empty identity as the office, and tool
+  calls from an unidentifiable caller are refused (a ghost identity could previously change priorities
+  and even delete a real task).
+- **v5 tool descriptions**: the meeting `kind` enum plus `target`, and `hire.term`, are now documented.
+
+### 4. Installer, declarations and guards
+
+- **Mechanism detection by capability**: the row registry package does not exist on the old line
+  (proven from npm metadata and package contents: 0.1.6-alpha.2 has no
+  `@deepseek-ai/dsh-agent-preset`, 0.1.7-rc.2 adds it), so the decision now tests the `register`
+  capability, with a falsifiable guard (reverting that one line turns only the new guard red).
+- **Declaration rows carry the host group marker**: the eight nested `!!js` expressions are preserved
+  for the preset scope instead of being evaluated in the declaring realm.
+- **Safer installer**: preset files are written atomically; a user-edited file is backed up before its
+  preset is dropped; an unreadable file is no longer deleted without a backup; the state file is
+  skipped when unchanged, write failures warn, and `listBundles` has a timeout fallback.
+- **Wider guards**: `run-tests` now collects **every** test and probe (two red probes used to be
+  silently skipped), one flaky test (v4 meetings) and one non-deterministic corpus (v3) were fixed, and
+  new guards cover the installer, the declaration row, command failure paths and host failure paths.
+
+## Compatibility
+
+- **Supported DSH window unchanged**: `0.1.2-alpha.4 … 0.2.0-rc.2`; both the composition-row (≥ 0.1.7)
+  and directory (≤ 0.1.6) lines are kept, and the capability-based verdict means the old line can no
+  longer be misclassified.
+- **No data or layout change**: `State/<institute>.v5state.json`, the project layout and the mirror file
+  locations are unchanged; no migration needed.
+- **Prompts**: unchanged except the v4 verdict-contract/clojure-rule wording, the v5 m-vote wording, one
+  consistent closure-rule line added to the v2/v3/v4 personas, and the added v5 tool-description fields;
+  corpora were regenerated for the current four presets.
+- **Defaults**: unchanged (v2 caps reviewers per target at
+  `max(2, min(verifierCount, maxParallelThreshold-2))`, i.e. 2 on the default configuration; the tool-call
+  limits remain advisory).
+
+## Upgrade
+
+```sh
+npm i dsh-vibe-math@latest
+```
+
+Desktop updates through Settings → Plugins; a CLI profile uses
+`dsh plugin --profile <your profile> add dsh-vibe-math@latest`, then restart DSH.