dsh-vibe-math 2.8.3 → 2.8.5

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

@@ -32,6 +32,7 @@ process.env.DSH_HOME = EMPTY_ENGINE_ROOT
 process.env.ProgramFiles = EMPTY_ENGINE_ROOT
 process.env['ProgramFiles(x86)'] = EMPTY_ENGINE_ROOT
 process.env.LOCALAPPDATA = EMPTY_ENGINE_ROOT
+process.env.V5_TEX_ROOTS_SANDBOX = EMPTY_ENGINE_ROOT   // task-29: the documented TeX Live roots are sandboxed too (the product's own list now covers D:/C:/texlive)
 
 // V5_PLUGIN lets a sensitivity probe point this suite at a deliberately broken copy.
 // Without it every probe against this suite silently tested the UNMUTATED plugin and was
@@ -1495,6 +1496,40 @@ console.log('\n[27] fake LaTeX compiler: repair path and persistent-failure degr
     '★ [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)')
 }
+// real1004-stall (found on a real host with 2.8.3): the meeting watchdog measured "stall" from the last INPUT and
+// ignored whether an ASKED member was still busy. A member's turn easily outlives the budget on a real host, so
+// meetings were abandoned mid-turn and the speeches could only land as "late notes" (that run: 1067 late notes
+// and 863 refused office finalisations). A meeting must survive while an asked speaker is in flight.
+{
+  const hs = makeHost({ pluginModule })
+  await hs.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await hs.settleSpawns()
+  await hs.callTool('vibe_v5_set', { activityTimeoutMs: 300 })   // soft stall = 600 ms, hard bound = 1800 ms
+  const mts = await hs.callTool('vibe_v5_meeting', { agenda: 'real1004-stall：被点名的成员仍在忙', kind: 'sync' })
+  const midS = mts && mts.meeting
+  assert(!!midS, 'precondition: the meeting is convened (' + JSON.stringify(mts).slice(0, 110) + ')')
+  // acad is asked and NOT answered, so it stays BUSY. Nobody submits an input, so `lastInputAt` keeps ageing.
+  // NOTE (why this is a SOURCE-level invariant): the watchdog branch runs from a heartbeat whose armed delay is
+  // not controllable from this harness (the plugin arms `ctx.timeout` with delays up to tens of seconds), so a
+  // runtime assertion here cannot be made to FAIL on the unfixed code — it would be a vacuous guard. The
+  // behaviour is instead pinned by (a) this source invariant and (b) the real-host re-test, which is the
+  // acceptance criterion of the board task (measured there: 1067 late notes before the fix).
+  const wA = await hs.peekWakeOf('acad', 3000)
+  assert(!!wA, 'precondition: a member was asked to speak (so it is busy) while no input has arrived')
+  await sleep(900)                                             // past the SOFT bound, inside the hard one
+  const mpathS = join(hs.WS, 'VibeMath', 'Projects', 'default', 'Institutes', 'institute', 'Shared', 'Meetings', midS + '.md')
+  assert(!readFileSync(mpathS, 'utf8').includes('因长时间无新发言而被放弃'),
+    'no input has arrived yet, so nothing may close this meeting: ' + readFileSync(mpathS, 'utf8').slice(0, 160))
+  hs.fireEnd(wA.childId, { input: 'real1004-stall：我被点名后仍在忙，现在交回我的发言。', contextPct: 20 })
+  await sleep(300)
+  const docS = readFileSync(mpathS, 'utf8')
+  assert(docS.includes('### acad') && !docS.includes('（会后补记'),
+    'a speech from an ASKED member lands in the NORMAL section (never as a late note)')
+  const v5src = readFileSync(PLUGIN, 'utf8')   // PLUGIN points at the mutated COPY when a family runs this suite
+  assert(v5src.includes('const inFlight = meeting.order.some((id) => busy.has(id))') &&
+    v5src.includes('if (stale >= recoverStallMs() && (!inFlight || stale >= recoverStallMs() * 3)) {'),
+    '★ [real1004-stall] the meeting watchdog exempts ASKED-but-busy speakers up to a hard bound (a real host abandoned meetings mid-turn: 1067 late notes, 863 refused finalisations)')
+}
 
 // ---------- 28. office-only surface + id normalisation ----------
 console.log('\n[28] the final-paper surface is office-only and its directory id cannot escape Paper/')
@@ -1742,9 +1777,9 @@ console.log('\n[35] async Lean: enqueue returns at once, settle(ok) is the only
     'the proof is archived under Verified/Lean/<id>.lean')
   const argv = fake.calls[0].argv
   const root = join(h.WS, 'VibeMath').replace(/\\/g, '/')
-  const si = argv.indexOf('--search-path')
+  const si = argv.indexOf('-R')
   assert(si > 0 && argv[si + 1] === root && si + 2 === argv.length - 1 && /\.lean$/.test(argv[argv.length - 1]),
-    '★ argv carries `--search-path <ABSOLUTE VibeMath root>` right before the file (' + JSON.stringify(argv) + ')')
+    '★ argv carries `-R <ABSOLUTE VibeMath root>` (Lean 4 root; the Lean-3 --search-path made lean fail at parsing) right before the file (' + JSON.stringify(argv) + ')')
   assert(fake.calls.length === 1, 'exactly one compile ran for the settle (got ' + fake.calls.length + ')')
 }
 
@@ -2503,6 +2538,217 @@ console.log('\n[53] v5 L1 leanPathContractOk: the queued note reaches report() d
   assert(rep.indexOf('Lean path contract broken') === -1,
     '★★ [v5 L1] a HEALTHY composition surfaces no Lean path contract note in report() diagnostics')
 }
+// task-24 (real-host R2, D5'): a statement longer than a display cap used to be cut SILENTLY — at the resolve
+// step (1200 chars) and again in the member prompt (800 chars) — and the cut text is what voters read and what
+// the verified record carries. SOURCE-LEVEL invariant: both sites must keep a generous cap AND mark any cut
+// (a runtime scenario would need a >20k-char card, which is impractical here; the cut-marking itself is pinned
+// by the single-site mutation in tests/v5-institute-fixes.mutants.mjs).
+{
+  const src = readFileSync(PLUGIN, 'utf8')
+  assert(src.includes('const cap = 20000') && src.includes('（注意：陈述超过 ' + "' + cap + '" + ' 字符已截断，完整文本见 '),
+    '★ [real1004-visibility] the resolved statement keeps a generous cap (20000) and MARKS any cut in-band')
+  assert(src.includes('（已截断；完整陈述见源卡片 ') && !src.includes("'  陈述：' + String(vs.statement).slice(0, 800)"),
+    '★ [real1004-visibility] the member-prompt statement line marks its 800-char cut instead of hiding the tail')
+}
+// task-26 (real-host R2): `vibe_v5_set` used to coerce out-of-range values and drop unknown keys with NOTHING in
+// the response, and two different capacity limits shared one error code. The response must report both, and the
+// two limits must be distinguishable by a machine-readable field.
+{
+  const hz = makeHost({ pluginModule })
+  await hz.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await hz.settleSpawns()
+  const set = await hz.callTool('vibe_v5_set', { quorumCap: -1, unknownKeyXyz: 5 })
+  assert(set.ok === true && Array.isArray(set.dropped) && set.dropped.indexOf('unknownKeyXyz') !== -1,
+    '★ [real1004-visibility] vibe_v5_set REPORTS the unknown key it dropped (dropped=' + JSON.stringify(set.dropped) + ')')
+  assert(set.adjusted && set.adjusted.quorumCap && Number(set.adjusted.quorumCap.from) === -1 && Number(set.adjusted.quorumCap.to) !== -1,
+    '★ [real1004-visibility] vibe_v5_set reports `adjusted` so a coerced value is visible (adjusted=' + JSON.stringify(set.adjusted) + ')')
+  // The per-member vs institute-wide distinction is pinned as a SOURCE invariant instead of by driving two
+  // refusals: `vibe_v5_hire` demands a full brief (`purpose` + `initial_task` + …) before it reaches the cap
+  // check, and assembling that brief here would only test the brief. The mutation in
+  // tests/v5-institute-fixes.mutants.mjs collapses the two scopes and reddens this assertion by name.
+  const src26 = readFileSync(PLUGIN, 'utf8')
+  assert(src26.includes("scope: 'per-member'") && src26.includes("scope: 'institute'"),
+    '★ [real1004-visibility] the two capacity refusals carry DISTINCT machine-readable scopes (both literals present)')
+}
+// task-25 (real-host R2/R3): `status().fieldScopes` must cover the fields the payload actually carries; an
+// incomplete electorate must NAME the missing voters; and v4 must name the residents that did not vote true.
+{
+  const hf = makeHost({ pluginModule })
+  await hf.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await hf.settleSpawns()
+  const stf = await hf.callTool('vibe_v5_status', {})
+  const fs = stf.fieldScopes || {}
+  assert(stf.diagnostics !== undefined && stf.backend !== undefined,
+    'precondition: the payload really carries diagnostics/backend (' + JSON.stringify({ d: typeof stf.diagnostics, b: typeof stf.backend }) + ')')
+  assert(Array.isArray(fs.session) && fs.session.indexOf('diagnostics') !== -1 && fs.session.indexOf('backend') !== -1,
+    '★ [real1004-visibility] fieldScopes.session covers the payload fields it used to omit (diagnostics/backend)')
+  assert(Array.isArray(fs.durable) && fs.durable.indexOf('verdicts') !== -1 && fs.durable.indexOf('solve') !== -1,
+    '★ [real1004-visibility] fieldScopes.durable declares verdicts/solve (a reader can tell durable facts apart)')
+  const src25 = readFileSync(PLUGIN, 'utf8')
+  assert(src25.includes('pendingVoters: need.filter((id) => !votes[id])'),
+    '★ [real1004-visibility] an incomplete electorate NAMES the missing voters (pendingVoters)')
+  const v4src = readFileSync(new URL('../vibe-math-v4/vibe-math-v4.js', import.meta.url), 'utf8')
+  assert(v4src.includes('；未投 true：'),
+    '★ [real1004-visibility] v4 names the residents that did not vote solved=true (no silent "no unanimous vote")')
+}
+// task-27 (real-host A round + local reproduction): Lean 4 accepts -R/--root (or LEAN_PATH); the Lean-3 spelling
+// `--search-path` makes `lean` fail at ARGUMENT PARSING (rc=1) before it ever reads the file — every compile failed.
+{
+  const src27 = readFileSync(PLUGIN, 'utf8')
+  assert(src27.includes("const LEAN_SEARCH_PATH_FLAG = '-R'"),
+    '★ [real1004-lean-flag] the injected Lean search root uses -R (Lean 4), not the Lean-3 --search-path')
+  assert(!src27.includes('--search-path <VibeMath root>') && !/registerTool\(\s*'[^']+'\s*,\s*'[^']*--search-path/.test(src27),
+    '★ [real1004-lean-flag] no agent-facing tool description still teaches --search-path')
+}
+// Iron-rules P4: the LaTeX-missing block was pushed into EVERY wake (founding/verify/meeting included) with no
+// statement of when it applies, so an agent that has nothing to do with the paper met unexplained noise.
+{
+  const srcP4 = readFileSync(PLUGIN, 'utf8')
+  assert(srcP4.includes('仅在你参与论文写作或编译时适用'),
+    '★ [real1004-prompt-scope] the LaTeX block says WHEN it applies instead of appearing as unexplained noise')
+  assert(/function paperPushLine\(L\)[\s\S]{0,600}仅在你参与论文写作或编译时适用/.test(srcP4),
+    '★ [real1004-prompt-scope] the scope note is what the shared tail actually pushes (not a dangling sentence)')
+}
+// ---------- 54. task-30: the methodology/collaboration feedback library (Shared/Feedback/) ----------
+console.log('\n[54] vibe_v5_feedback: 用途＋三路由提示；add→list→状态流转；off 不注入＋具名拒绝＋不写盘；计数与库一致')
+{
+  const INSDIR = join('VibeMath', 'Projects', 'default', 'Institutes', 'institute')
+  const h = makeHost({ pluginModule })
+  await h.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await h.settleSpawns()
+  const r1 = h.childAgent(h.childOf('r-1'))
+  const acad = h.childAgent(h.childOf('acad'))
+  const fbPath = (...p) => join(h.WS, INSDIR, 'Shared', 'Feedback', ...p)
+  const wake = async (member) => {
+    await h.callTool('vibe_v5_say', { to: member, text: '请继续推进。' }, acad)
+    const w = await h.peekWakeOf(member, 3000)
+    if (w) h.fireEnd(w.childId, { progress: '收到。', solved: false, contextPct: 10 })
+    await sleep(25)
+    return w ? w.text : ''
+  }
+  const countsLine = (t) => (String(t).split('\n').find((l) => l.indexOf('条目 ') !== -1 && l.indexOf('未闭环 ') !== -1) || '')
+
+  // (1) 每一轮都可见：为什么记／记什么／三条路由／记了之后会怎样
+  const p1 = await wake('r-1')
+  assert(p1.indexOf('【工作经验／流程反馈') !== -1, '★ [task-30] the per-round prompt carries the feedback hint')
+  assert(p1.indexOf('为什么记') !== -1 && p1.indexOf('记什么') !== -1 && p1.indexOf('记了之后') !== -1,
+    '★ [task-30] the hint explains WHY / WHAT / what happens after')
+  assert(p1.indexOf('self＝') !== -1 && p1.indexOf('team＝') !== -1 && p1.indexOf('interpersonal＝') !== -1,
+    '★ [task-30] the hint distinguishes the THREE routes')
+  assert(p1.indexOf('自己调整，无需谁采纳') !== -1 && p1.indexOf('同样无需审批') !== -1,
+    '★★ [task-30] self/team routes are self-regulation — the hint never asks for an approval')
+  assert(p1.indexOf('缜密评估') !== -1 && p1.indexOf('事后回填验证') !== -1,
+    '★ [task-30] only the interpersonal route demands a careful assessment + written-back verification')
+  assert(p1.indexOf('Progress/') !== -1 && p1.indexOf('Verified/') !== -1,
+    '★ [task-30] the boundary against Progress//Verified/ is stated in the prompt')
+  assert(p1.indexOf('向用户问一次') === -1 && p1.indexOf('ask the user once') === -1,
+    'the hint never demands an action a member cannot perform (I16)')
+
+  // (2) add → list 可见；self/team 的状态流转无需任何审批
+  const add = await h.callTool('vibe_v5_feedback', { op: 'add', category: 'process', route: 'self', context: '每轮都在等别人回复', phenomenon: '轮次空转', impact: '进展慢', action: '我自己改成先写进度再发问' }, r1)
+  assert(add.ok === true && add.entry && add.entry.id === 'fb-1' && add.entry.status === 'open',
+    '★ [task-30] add records an entry (open, id=fb-1): ' + JSON.stringify(add.entry && add.entry.id))
+  assert(add.mirror === true && existsSync(fbPath('process.md')),
+    '★ [task-30] the human-readable mirror is really written (Shared/Feedback/process.md)')
+  const list1 = await h.callTool('vibe_v5_feedback', { op: 'list' }, r1)
+  assert(list1.count === 1 && list1.entries[0].id === 'fb-1' && list1.openOnly === true,
+    '★ [task-30] list defaults to the NOT-yet-closed entries')
+  const upd = await h.callTool('vibe_v5_feedback', { op: 'update', id: 'fb-1', status: 'adjusted', outcome: '工作流改完，空转少了' }, r1)
+  assert(upd.ok === true && upd.entry.status === 'adjusted' && /空转少了/.test(upd.entry.outcome),
+    '★ [task-30] status flows open→adjusted WITHOUT any approval step (self-regulation)')
+  const addTeam = await h.callTool('vibe_v5_feedback', { op: 'add', category: 'cooperation', route: 'team', phenomenon: '任务分配不均', impact: '有人闲置', action: '分派前先看负载' }, acad)
+  assert(addTeam.ok === true, 'the academician can record a team-route entry (organisation / workflow)')
+  const forbidden = await h.callTool('vibe_v5_feedback', { op: 'update', id: addTeam.entry.id, status: 'closed' }, r1)
+  assert(forbidden.ok === false && forbidden.code === 'V5_FEEDBACK_FORBIDDEN',
+    '★ [task-30] a member cannot update somebody else\'s entry (named refusal)')
+
+  // (3) interpersonal：必须带评估；闭环必须回填 outcome（事后验证）
+  const needA = await h.callTool('vibe_v5_feedback', { op: 'add', category: 'conflict', route: 'interpersonal', phenomenon: 'r-2 不回消息', impact: '我卡住', action: '等他回' }, r1)
+  assert(needA.ok === false && needA.code === 'V5_FEEDBACK_NEEDS_ASSESSMENT',
+    '★★ [task-30] the interpersonal route refuses to be recorded WITHOUT the assessment')
+  const inter = await h.callTool('vibe_v5_feedback', { op: 'add', category: 'conflict', route: 'interpersonal', phenomenon: 'r-2 不回消息', impact: '我卡住', action: '约会议当面问', assessment: '让他改习惯是否真更优？先约会议成本更低' }, r1)
+  assert(inter.ok === true && inter.entry.route === 'interpersonal', 'the interpersonal entry is accepted WITH the assessment')
+  const noOut = await h.callTool('vibe_v5_feedback', { op: 'update', id: inter.entry.id, status: 'closed' }, r1)
+  assert(noOut.ok === false && noOut.code === 'V5_FEEDBACK_NEEDS_OUTCOME',
+    '★★ [task-30] closing an interpersonal entry WITHOUT the written-back verification is refused')
+  const officeFix = await h.callTool('vibe_v5_feedback', { op: 'update', id: inter.entry.id, status: 'closed', outcome: '会上确认了回复习惯，之后未再发生' })
+  assert(officeFix.ok === true && officeFix.entry.status === 'closed',
+    '★ [task-30] the OFFICE may resolve ANY entry and write the result back')
+
+  // (4) off：提示词不注入 ＋ 每个 op 都具名拒绝 ＋ 一个字都不写
+  const rpBefore = (await h.callTool('vibe_v5_report', {})).report
+  const idxBefore = readFileSync(fbPath('index.json'), 'utf8')
+  await h.callTool('vibe_v5_set', { feedback: 'off' })
+  const p2 = await wake('r-1')
+  assert(p2.indexOf('【工作经验／流程反馈') === -1, '★★ [task-30] feedback=off: the per-round hint is NOT injected')
+  const offAdd = await h.callTool('vibe_v5_feedback', { op: 'add', category: 'process', route: 'self', phenomenon: 'x', impact: 'y', action: 'z' }, r1)
+  assert(offAdd.ok === false && offAdd.code === 'V5_FEEDBACK_DISABLED' && /已关闭/.test(offAdd.message),
+    '★★ [task-30] off refuses the write BY NAME (never a silent drop)')
+  const offList = await h.callTool('vibe_v5_feedback', { op: 'list', all: true }, r1)
+  assert(offList.ok === false && offList.code === 'V5_FEEDBACK_DISABLED',
+    'off refuses EVERY op, so nothing can look normal while the feature is disabled')
+  const rpAfter = (await h.callTool('vibe_v5_report', {})).report
+  assert(countsLine(rpAfter) === countsLine(rpBefore) && idxBefore === readFileSync(fbPath('index.json'), 'utf8'),
+    '★★ [task-30] off wrote NOTHING (report counts and the on-disk index are unchanged)')
+  await h.callTool('vibe_v5_set', { feedback: 'on' })
+
+  // (5) 计数与库内容一致（工具 summary ／ report ／ overview 三个面）
+  const summary = (await h.callTool('vibe_v5_feedback', { op: 'summary' }, r1)).counts
+  const allEntries = (await h.callTool('vibe_v5_feedback', { op: 'list', all: true }, r1)).entries
+  assert(summary.total === allEntries.length && summary.open + summary.closed === summary.total,
+    '★★ [task-30] summary counts match the library contents exactly (' + JSON.stringify(summary) + ')')
+  const byCat = allEntries.reduce((m, e) => { m[e.category] = (m[e.category] || 0) + 1; return m }, {})
+  assert(Object.keys(summary.byCategory).every((k) => summary.byCategory[k] === (byCat[k] || 0)),
+    '★★ [task-30] the by-category counts match the entries')
+  const rp = await h.callTool('vibe_v5_report', {})
+  assert(rp.report.indexOf('## 反馈库') !== -1 && rp.report.indexOf('条目 ' + summary.total + '｜未闭环 ' + summary.open) !== -1,
+    '★★ [task-30] report() shows the SAME counts as the library')
+  const ov = await h.callTool('vibe_v5_overview', {}, acad)
+  assert(ov.overview.indexOf('## 反馈库') !== -1 && ov.overview.indexOf('条目 ' + summary.total) !== -1 && ov.overview.indexOf('未闭环 ' + summary.open) !== -1,
+    '★★ [task-30] overview shows the same counts (the academician\'s periodic summary reads this)')
+}
+// ---------- 55. task-31: three prompt-review follow-ups (P1 receipt background / P3 empty engine table / P4 frame) ----
+console.log('\n[55] task-31: P1 首次提回执处给背景＋字段最小注解；P3 空引擎表"可用引擎 无"；P4 面向代理不再自称"框架"')
+{
+  // P3/P1 at the SOURCE of the line: the shared module builds it, so call it for real with an empty probe.
+  // The sibling is resolved THROUGH the same seam as the plugin, so a deliberately broken copy is the
+  // module actually under test (otherwise a mutation here could never redden this guard).
+  const mc = await import(new URL('./math-computation.js', PLUGIN))
+  const emptyZh = mc.mathAvailabilityLine({ engines: [] }, 'zh', 'typed+shell')
+  const emptyEn = mc.mathAvailabilityLine({ engines: [] }, 'en', 'typed+shell')
+  // P1: the archive→edit→re-run section explains WHAT a receipt is at its FIRST mention, and the bare
+  // field name carries its own minimal context (only the necessary ones - the section must stay short).
+  assert(emptyZh.indexOf('回执＝一次 math_computation 调用的 JSON 结果') !== -1,
+    '★★ [task-31] P1: 归档→编辑→重跑 段首次提到回执处解释了"回执是什么"')
+  assert(emptyZh.indexOf('（脚本内容相对上一份回执变过）') !== -1,
+    '★ [task-31] P1: 裸字段名 scriptChanged 自带最小上下文')
+  assert(emptyEn.indexOf('a RECEIPT is the JSON result of ONE math_computation call') !== -1
+    && emptyEn.indexOf('(the script content differs from the previous receipt)') !== -1,
+    '★ [task-31] P1: en 段同步给出回执背景与字段上下文')
+  // P3: an empty engine table must not read as "可用的东西：无".
+  assert(emptyZh.indexOf('math_computation：可用引擎 无') !== -1 && emptyZh.indexOf('本机可用') === -1,
+    '★★ [task-31] P3: 空引擎表渲染为「可用引擎 无」: ' + JSON.stringify((/math_computation：[^\n]{0,30}/.exec(emptyZh) || ['<none>'])[0]))
+  assert(emptyEn.indexOf('math_computation: available engines none') !== -1,
+    '★ [task-31] P3: en 空表同步为「available engines none」')
+  // …and the SAME text really reaches a member (end-to-end, not only at the module boundary).
+  const h = makeHost({ pluginModule })
+  await h.callTool('vibe_v5_start', { problem: PROBLEM, researcherCount: 1 })
+  await h.settleSpawns()
+  await h.callTool('vibe_v5_say', { to: 'r-1', text: '请继续推进。' }, h.childAgent(h.childOf('acad')))
+  const w = await h.peekWakeOf('r-1', 3000)
+  const p = w ? w.text : ''
+  if (w) h.fireEnd(w.childId, { progress: '收到。', solved: false, contextPct: 10 })
+  assert(p.indexOf('回执＝一次 math_computation 调用的 JSON 结果') !== -1,
+    '★★ [task-31] P1: 成员那一轮真正读到的提示词就带这段背景')
+  assert(p.indexOf('math_computation：可用引擎 无') !== -1 && p.indexOf('本机可用') === -1,
+    '★★ [task-31] P3: 成员提示词里是「可用引擎 无」')
+  // P4: agent-facing text must not call the plugin "框架" (ambiguous from an agent's point of view:
+  // plugin? host? DSH?). The one class fixed here is the notice frame.
+  const src31 = readFileSync(PLUGIN, 'utf8')
+  assert(src31.indexOf('【框架提示】') === -1, '★★ [task-31] P4: 面向代理文本里不再出现【框架提示】')
+  assert(src31.indexOf("return '【研究所提示】' + m.text") !== -1,
+    '★ [task-31] P4: 研究所回执帧改用【研究所提示】')
+}
 console.log('')
 console.log('passed=' + passed + ' failed=' + failed)
 if (failed) { console.error('FAILURES:'); for (const f of failures) console.error('  - ' + f); process.exit(1) }

package/tests/formal-verify-v2.test.mjs

@@ -1362,9 +1362,10 @@ section('15d Lean 增量 + 异步（spec §1–§4 + 修订 §1–§5）：参
   assert(JSON.stringify(planDefault.paths) === JSON.stringify(['/root/vm']) && planDefault.inject.length === 2, '★ default plan injects exactly the automatic VibeMath root')
   const planExtra = H.leanSearchPathPlan(['env', 'lean'], ['/a', '/b'], '/root/vm')
   assert(JSON.stringify(planExtra.paths) === JSON.stringify(['/a', '/b', '/root/vm']), '★★ user leanSearchPaths come FIRST, then the automatic root')
-  assert(JSON.stringify(planExtra.inject) === JSON.stringify(['--search-path', '/a', '--search-path', '/b', '--search-path', '/root/vm']), 'and each path is injected with its own --search-path flag')
-  assert(H.leanSearchPathPlan(['--search-path', '/x'], ['/a'], '/root/vm').inject.length === 0, '★ an explicit --search-path in leanArgs suppresses all injection (user wins)')
+  assert(JSON.stringify(planExtra.inject) === JSON.stringify(['-R', '/a', '-R', '/b', '-R', '/root/vm']), 'and each path is injected with its own -R (Lean 4) flag')
+  assert(H.leanSearchPathPlan(['--search-path', '/x'], ['/a'], '/root/vm').inject.length === 0, '★ an explicit legacy Lean-3 --search-path in leanArgs suppresses all injection (user wins)')
   assert(H.leanSearchPathPlan(['-R', '/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--root', '/x'], ['/a'], '/root/vm').inject.length === 0, '★ -R / --root are honoured too')
+  assert(H.leanSearchPathPlan(['-R=/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--search-path=/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--root=/x'], ['/a'], '/root/vm').inject.length === 0, '★ the = forms of -R / legacy --search-path / --root are honoured too')
 
   // ── argv：注入位置在用户参数之后、文件名之前 ──────────────────────────────────────
   const b = await makeCase('lean-argv')
@@ -1377,15 +1378,15 @@ section('15d Lean 增量 + 异步（spec §1–§4 + 修订 §1–§5）：参
   const argv1 = leanRuns[leanRuns.length - 1].argv
   assert(leanRuns.length === runsBefore + 1, 'the sync run really reached the compiler')
   assert(JSON.stringify(argv1.slice(0, 2)) === JSON.stringify(['lean', 'env']), 'the user leanArgs are preserved in order')
-  assert(argv1[argv1.length - 2] === '--search-path' && SAME(argv1[argv1.length - 1], vibeRoot(b)), '★★ --search-path + the VibeMath root are injected AFTER the user args and immediately BEFORE the file (argv=' + JSON.stringify(argv1) + ')')
+  assert(argv1[argv1.length - 2] === '-R' && SAME(argv1[argv1.length - 1], vibeRoot(b)), '★★ -R + the VibeMath root are injected AFTER the user args and immediately BEFORE the file (argv=' + JSON.stringify(argv1) + ')')
   await b.call('vibe_math_set_params', { leanArgs: ['env', 'lean', '--search-path', '/explicit'] })
   await b.call('vibe_math_lean_run', { file: 'Formal/srch.lean' })
   const argv2 = leanRuns[leanRuns.length - 1].argv
-  assert(argv2.indexOf('--search-path') === 3 && argv2[4] === '/explicit' && argv2.filter((x) => x === '--search-path').length === 1, '★ an explicit --search-path is respected verbatim (no second injection) — got ' + JSON.stringify(argv2))
+  assert(argv2.indexOf('--search-path') === 3 && argv2[4] === '/explicit' && argv2.filter((x) => x === '--search-path').length === 1 && argv2.indexOf('-R') === -1, '★ an explicit legacy --search-path is respected verbatim: nothing is injected (no -R) — got ' + JSON.stringify(argv2))
   await b.call('vibe_math_set_params', { leanArgs: ['env', 'lean'], leanSearchPaths: ['/extra1', '/extra2'] })
   await b.call('vibe_math_lean_run', { file: 'Formal/srch.lean' })
   const argv3 = leanRuns[leanRuns.length - 1].argv
-  assert(JSON.stringify(argv3.slice(-6)) === JSON.stringify(['--search-path', '/extra1', '--search-path', '/extra2', '--search-path', slash(vibeRoot(b))]), '★★ leanSearchPaths are injected first, then the automatic root (got ' + JSON.stringify(argv3.slice(-6)) + ')')
+  assert(JSON.stringify(argv3.slice(-6)) === JSON.stringify(['-R', '/extra1', '-R', '/extra2', '-R', slash(vibeRoot(b))]), '★★ leanSearchPaths are injected first, then the automatic root (got ' + JSON.stringify(argv3.slice(-6)) + ')')
 
   // ── 异步状态机：入队立即返回 → 心跳排空 → 只有 settled(ok) 才是通过 ────────────────
   const c = await makeCase('lean-async')

package/tests/formal-verify-v3.test.mjs

@@ -564,7 +564,7 @@ writeFileSync(join(toolProj, 'Formal', 'hang.lean'), '-- HANG\ntheorem t : 1 = 1
   assert(/ok/.test(runGood.stdout), 'the compiler stdout is returned')
   const last = leanRuns[leanRuns.length - 1]
   assert(last.cwd.replace(/\\/g, '/') === toolProj.replace(/\\/g, '/'), 'the toolchain runs with the PROJECT root as cwd')
-  assert(last.argv.join(' ') === 'lean --search-path ' + VIBE.replace(/\\/g, '/'), 'by default the argv is the executable + the injected search path (got ' + last.argv.join(' ') + ')')
+  assert(last.argv.join(' ') === 'lean -R ' + VIBE.replace(/\\/g, '/'), 'by default the argv is the executable + the injected -R (Lean 4) search root (got ' + last.argv.join(' ') + ')')
   assert(last.graceMs === 120000, 'the default leanTimeoutMs is passed down as graceMs (' + last.graceMs + ')')
 }
 {
@@ -645,7 +645,7 @@ try {
   await callTool('vibe_math_set_params', { leanCommand: 'lake', leanArgs: ['env', 'lean'] }, RE)
   const runLake = await callTool('vibe_math_lean_run', { file: 'Formal/good.lean' }, RE)
   assert(runLake.ok === true, 'lake env lean works through leanCommand/leanArgs')
-  assert(leanRuns[leanRuns.length - 1].argv.join(' ') === 'lake env lean --search-path ' + VIBE.replace(/\\/g, '/'), 'the argv is [exe, ...leanArgs, injected --search-path] (got ' + leanRuns[leanRuns.length - 1].argv.join(' ') + ')')
+  assert(leanRuns[leanRuns.length - 1].argv.join(' ') === 'lake env lean -R ' + VIBE.replace(/\\/g, '/'), 'the argv is [exe, ...leanArgs, injected -R (Lean 4)] (got ' + leanRuns[leanRuns.length - 1].argv.join(' ') + ')')
   await callTool('vibe_math_set_params', { leanCommand: 'lean', leanArgs: [] }, RE)
 }
 
@@ -1324,6 +1324,7 @@ section('9b Lean 增量 + 异步：参数/argv/队列/去重/只读面/主动性
   const planExtra = H.leanSearchPathPlan(['env', 'lean'], ['/a', '/b'], '/root/vm')
   assert(JSON.stringify(planExtra.paths) === JSON.stringify(['/a', '/b', '/root/vm']), '★★ user leanSearchPaths come FIRST, then the automatic root')
   assert(H.leanSearchPathPlan(['--search-path', '/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['-R', '/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--root', '/x'], ['/a'], '/root/vm').inject.length === 0, '★ an explicit --search-path/-R/--root in leanArgs suppresses all injection (user wins)')
+  assert(H.leanSearchPathPlan(['-R=/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--search-path=/x'], ['/a'], '/root/vm').inject.length === 0 && H.leanSearchPathPlan(['--root=/x'], ['/a'], '/root/vm').inject.length === 0, '★ the = forms of -R / legacy --search-path / --root are honoured too')
   const ctxA = H.leanBuildContext('lean', ['a'], ['/p'])
   const ctxB = H.leanBuildContext('lean', ['b'], ['/p'])
   assert(H.leanJobFingerprint('x\n', ctxA) !== H.leanJobFingerprint('x\n', ctxB), '★ the job fingerprint covers the build context (engine/args/search paths)')
@@ -1339,15 +1340,15 @@ section('9b Lean 增量 + 异步：参数/argv/队列/去重/只读面/主动性
   await callTool('vibe_math_lean_run', { file: 'Formal/srch.lean' }, ra)
   const argv1 = leanRuns[leanRuns.length - 1].argv
   assert(leanRuns.length === runs0 + 1 && JSON.stringify(argv1.slice(0, 3)) === JSON.stringify(['lake', 'env', 'lean']), 'the user leanArgs are preserved in order (lake env lean)')
-  assert(argv1[argv1.length - 2] === '--search-path' && SAME(argv1[argv1.length - 1], VIBE), '★★ --search-path + the VibeMath root are injected AFTER the user args and immediately BEFORE the file (' + JSON.stringify(argv1) + ')')
+  assert(argv1[argv1.length - 2] === '-R' && SAME(argv1[argv1.length - 1], VIBE), '★★ -R + the VibeMath root are injected AFTER the user args and immediately BEFORE the file (' + JSON.stringify(argv1) + ')')
   await callTool('vibe_math_set_params', { leanArgs: ['env', 'lean', '--search-path', '/explicit'] }, ra)
   await callTool('vibe_math_lean_run', { file: 'Formal/srch.lean' }, ra)
   const argv2 = leanRuns[leanRuns.length - 1].argv
-  assert(argv2.filter((x) => x === '--search-path').length === 1 && argv2[4] === '/explicit', '★ an explicit --search-path is respected verbatim (no second injection) — got ' + JSON.stringify(argv2))
+  assert(argv2.filter((x) => x === '--search-path').length === 1 && argv2[4] === '/explicit' && argv2.indexOf('-R') === -1, '★ an explicit legacy --search-path is respected verbatim: nothing is injected (no -R) — got ' + JSON.stringify(argv2))
   await callTool('vibe_math_set_params', { leanArgs: ['env', 'lean'], leanSearchPaths: ['/extra1', '/extra2'] }, ra)
   await callTool('vibe_math_lean_run', { file: 'Formal/srch.lean' }, ra)
   const argv3 = leanRuns[leanRuns.length - 1].argv
-  assert(JSON.stringify(argv3.slice(-6)) === JSON.stringify(['--search-path', '/extra1', '--search-path', '/extra2', '--search-path', slash(VIBE)]), '★★ leanSearchPaths are injected first, then the automatic root (got ' + JSON.stringify(argv3.slice(-6)) + ')')
+  assert(JSON.stringify(argv3.slice(-6)) === JSON.stringify(['-R', '/extra1', '-R', '/extra2', '-R', slash(VIBE)]), '★★ leanSearchPaths are injected first, then the automatic root (got ' + JSON.stringify(argv3.slice(-6)) + ')')
 
   // ── 异步状态机：入队立即返回 → 心跳排空 → 只有 settled(ok) 才是通过 ───────────────
   const rq = makeRoot()

package/tests/formal-verify-v4.test.mjs

@@ -1507,14 +1507,14 @@ section('22 leanSearchPaths/leanJobsMaxParallel/leanInitiative + build-context j
   writeIn(N, 'Projects/default/Formal/a-sp.lean', 'theorem a_sp : True := by trivial\n')
   await N.callTool('vibe_v4_lean_run', { file: 'Formal/a-sp.lean' }, N.resAgent(N.childOf('r-1')))
   let last = leanRuns[leanRuns.length - 1]
-  const spIdx = last.argv.indexOf('--search-path')
-  assert(last.argv.filter(a => a === '--search-path').length === 3, '★ one --search-path per DISTINCT root (user paths first, then the automatic VibeMath root)')
+  const spIdx = last.argv.indexOf('-R')
+  assert(last.argv.filter(a => a === '-R').length === 3, '★ one -R (Lean 4 root) per DISTINCT root (user paths first, then the automatic VibeMath root)')
   assert(last.argv[spIdx + 1] === '/extra/one' && last.argv[spIdx + 3] === '/extra/two' && last.argv[spIdx + 5] === N.vibeRoot.replace(/\\/g, '/'), '★ order + de-duplication: user paths in order, automatic root last')
-  assert(last.argv.indexOf('--search-path') < last.argv.length - 1 && /\.lean$/.test(String(last.argv[last.argv.length - 1])), 'the roots stay BEFORE the file name')
+  assert(last.argv.indexOf('-R') < last.argv.length - 1 && /\.lean$/.test(String(last.argv[last.argv.length - 1])), 'the roots stay BEFORE the file name')
   await N.callTool('vibe_v4_set', { leanArgs: ['--search-path', '/user/own'] })
   await N.callTool('vibe_v4_lean_run', { file: 'Formal/a-sp.lean' }, N.resAgent(N.childOf('r-1')))
   last = leanRuns[leanRuns.length - 1]
-  assert(last.argv.filter(a => a === '--search-path').length === 1 && last.argv[last.argv.indexOf('--search-path') + 1] === '/user/own', '★ an explicit --search-path in leanArgs wins outright: nothing is injected')
+  assert(last.argv.filter(a => a === '--search-path').length === 1 && last.argv[last.argv.indexOf('--search-path') + 1] === '/user/own' && last.argv.indexOf('-R') === -1, '★ an explicit legacy --search-path in leanArgs wins outright: nothing is injected (no -R)')
   await N.callTool('vibe_v4_set', { leanArgs: ['-R', '/user/root'] })
   await N.callTool('vibe_v4_lean_run', { file: 'Formal/a-sp.lean' }, N.resAgent(N.childOf('r-1')))
   assert(leanRuns[leanRuns.length - 1].argv.indexOf('--search-path') === -1, '★ -R/--root count as a user-stated root too')

package/tests/prompt-v5-integrity.test.mjs

@@ -639,19 +639,19 @@ const wakeR1With = async (kind, reply) => {
   await callTool('vibe_v5_say', { to: 'r-1', text: '请就当前状态给个结论。' }, childAgent(rd2))
   await settle(); await drainWakes(20, RD)
   const handled = delivered.slice()
-  for (const w of handled) if (/【框架提示】/.test(w.prompt)) recordAndCheck(kind, w.owner, w.prompt).length
+  for (const w of handled) if (/【研究所提示】/.test(w.prompt)) recordAndCheck(kind, w.owner, w.prompt).length
   return handled.map(w => w.prompt).join('\n')
 }
 const badVerdictText = await wakeR1With('notice', { verdict: { target: 'p-ghost', verdict: 'not-a-number', reason: 'x' }, contextPct: 20 })
-assert(/【框架提示】/.test(badVerdictText), 'a malformed verdict produces a real 【框架提示】 delivery')
+assert(/【研究所提示】/.test(badVerdictText), 'a malformed verdict produces a real 【研究所提示】 delivery')
 assert(/verdict 必须是 0-1 的数值/.test(badVerdictText), 'the notice says what was wrong')
-assert(/【框架提示】[^\n]*verdict/.test(badVerdictText), 'the notice is framed by the framework, not by the member itself')
+assert(/【研究所提示】[^\n]*verdict/.test(badVerdictText), 'the notice is framed by the institute, not by the member itself')
 const badClaimText = await wakeR1With('notice-claim', { task_claim: 't-999', contextPct: 20 })
-assert(/【框架提示】/.test(badClaimText) && /没有任务 t-999/.test(badClaimText), 'a refused claim is reported back to the claimer')
+assert(/【研究所提示】/.test(badClaimText) && /没有任务 t-999/.test(badClaimText), 'a refused claim is reported back to the claimer')
 // A CAS refusal reported through the JSON reply must also come back.
 const t = await callTool('vibe_v5_task_create', { subject: '一个没人认领的任务' }, childAgent(rd1))
 const staleText = await wakeR1With('notice-task', { task_update: { task_id: t.task.id, expected_revision: 99, action: 'claim' }, contextPct: 20 })
-assert(/【框架提示】/.test(staleText) && /V5_TASK_STALE_REVISION/.test(staleText), 'a stale CAS reported through the reply channel is echoed back')
+assert(/【研究所提示】/.test(staleText) && /V5_TASK_STALE_REVISION/.test(staleText), 'a stale CAS reported through the reply channel is echoed back')
 const selfFramed = corpus.filter(c => new RegExp('【研究所·私信 from ' + c.owner + '】').test(c.prompt))
 assert(selfFramed.length === 0, 'no member ever receives a message framed as coming from itself')
 await endCase(RD)
@@ -1278,7 +1278,7 @@ const corpusMd = readFileSync(mdPath, 'utf8')
 assert(corpusMd.length > 30000, 'the corpus is substantial (' + corpusMd.length + ' chars) — the real prompt text is preserved')
 assert(corpusMd.indexOf('[状态] 你是 acad（院士）') !== -1, 'a human can verify the academician brief verbatim')
 assert(corpusMd.indexOf('[状态] 你是 r-2') !== -1, 'a human can verify a researcher brief naming itself')
-assert(corpusMd.indexOf('【框架提示】') !== -1, 'the corpus contains the framework-feedback interaction')
+assert(corpusMd.indexOf('【研究所提示】') !== -1, 'the corpus contains the institute-feedback interaction')
 assert(corpusMd.indexOf('【会话重建 —— ') !== -1, 'the corpus contains a resume brief')
 assert(corpusMd.indexOf('【所办分派】') !== -1, 'the corpus contains an office assignment')
 assert(!/你是 \?/.test(corpusMd), 'the corpus contains NO wrong-identity "?" brief')

package/tests/run-tests.mjs

@@ -37,8 +37,8 @@
  *   node tests/run-tests.mjs --no-temp-hygiene    # skip the sweep and the roomier-drive temp root
  */
 import { spawn } from 'node:child_process'
-import { existsSync, readdirSync } from 'node:fs'
-import { cpus } from 'node:os'
+import { existsSync, readdirSync, mkdtempSync, rmSync } from 'node:fs'
+import { cpus, tmpdir } from 'node:os'
 import { join } from 'node:path'
 import { fileURLToPath } from 'node:url'
 
@@ -92,6 +92,22 @@ if (!has('no-temp-hygiene') && !has('self-check') && !has('counts')) {
     }
   } catch (e) { console.error('run-tests: temp hygiene skipped (' + ((e && e.message) || e) + ')') }
 }
+// task-29: HOST-INDEPENDENT GATE. Whether THIS machine has TeX Live installed must not decide whether the
+// suites that assert the "no LaTeX engine here" degradation are green: on this dev box `D:\texlive\...`
+// really exists and the product now FINDS it (that is the task-28 fix), so those suites would redden for a
+// reason that has nothing to do with the code under test. Point every child suite (they inherit
+// `process.env` — see runSuite's spawn, which passes no `env`) at ONE empty scratch directory through the
+// four documented-root sandbox gates: the product still probes every documented TeX root, one by one,
+// inside that empty directory, finds nothing, and `triedPaths` still NAMES what it looked at. Real-machine
+// discovery is covered by the SLV live runs, not by the gate; `audit-engine-faces` FACE 3 drives an
+// explicit `paperLatexCommand`, which the sandbox never touches (only the known-install stage is rebased).
+// Nothing else about the sweep changes; the scratch dir is removed when the runner exits.
+const TEX_ROOTS_SANDBOX_DIR = mkdtempSync(join(tmpdir(), 'vibe-gate-tex-sandbox-'))
+process.env.V2_TEX_ROOTS_SANDBOX = TEX_ROOTS_SANDBOX_DIR
+process.env.V3_TEX_ROOTS_SANDBOX = TEX_ROOTS_SANDBOX_DIR
+process.env.V4_TEX_ROOTS_SANDBOX = TEX_ROOTS_SANDBOX_DIR
+process.env.V5_TEX_ROOTS_SANDBOX = TEX_ROOTS_SANDBOX_DIR
+process.on('exit', () => { try { rmSync(TEX_ROOTS_SANDBOX_DIR, { recursive: true, force: true }) } catch (e) { /* best effort: a leftover empty scratch dir must never fail the gate */ } })
 const asJson = has('json')
 // task-13 (MEASURED): the default used to be min(4, cpus) = 4 on this 4-core box, and the heavy jobs then
 // ran 1.3x-4.5x SLOWER than standalone - the gate's own "slowest" lines recorded v2-fix-probes.mutants

package/tests/v4-final-paper.mutants.mjs

@@ -103,6 +103,20 @@ const FAMILIES = [
     to: "zh?''",
     expect: /\[deliverable\/v4\]/,
   },
+  {
+    // task-23: a meeting speech missing from the end block must be recorded with a NAMED placeholder.
+    name: 'real1004-minutes: a silent meeting speech is stored as an empty body again',
+    from: "'（本轮结束块未含 input/summary：框架按空发言记录，见 docs/final-paper.md）'",
+    to: "''",
+    expect: /\[real1004-minutes\] a meeting speech missing from the end block is recorded with a NAMED placeholder/,
+  },
+  {
+    // task-27: v4 must inject Lean 4's -R, not the Lean-3 --search-path.
+    name: 'real1004-lean-flag: v4 falls back to the Lean-3 --search-path',
+    from: "const LEAN_SEARCH_PATH_FLAG = '-R'",
+    to: "const LEAN_SEARCH_PATH_FLAG = '--search-path'",
+    expect: /\[real1004-lean-flag\] v4 injects -R/,
+  },
 ]
 let red = 0
 for (const f of FAMILIES) { const t0 = Date.now(); const ok = runFamily(f); const ms = Date.now() - t0; if (ok) red++ }

package/tests/v4-final-paper.test.mjs

@@ -621,6 +621,24 @@ section('G3 a paper whose REQUIRED artifact was not written is FAILED, never don
   await finish(m)
 }
 
+// task-23 (real-host R3): all four meetings recorded correct speaker headers with EMPTY bodies, because a
+// resident's end block carried no `input`/`summary` and the collector stored '' silently. A NAMED placeholder is
+// required instead. SOURCE-LEVEL invariant: driving a full meeting with a deliberately silent resident would
+// exercise the harness more than this one line.
+{
+  const v4src = readFileSync(PLUGIN, 'utf8')   // PLUGIN points at the mutated COPY when a family runs this suite
+  assert(v4src.includes('（本轮结束块未含 input/summary：框架按空发言记录') && !v4src.includes("{input:parsed.input||parsed.summary||'',voteSolved:"),
+    '★ [real1004-minutes] a meeting speech missing from the end block is recorded with a NAMED placeholder, never as a silent empty body')
+}
+// task-27: the same Lean-4 flag correction as v5 — the injected search root must be `-R`, and no agent-facing
+// tool description may still teach the Lean-3 `--search-path` (which makes `lean` fail at argument parsing).
+{
+  const v4s = readFileSync(PLUGIN, 'utf8')
+  assert(v4s.includes("const LEAN_SEARCH_PATH_FLAG = '-R'"),
+    '★ [real1004-lean-flag] v4 injects -R (Lean 4 root), not the Lean-3 --search-path')
+  assert(!v4s.includes('--search-path <VibeMath root>') && !/registerTool\(\s*'[^']+'\s*,\s*'[^']*--search-path/.test(v4s),
+    '★ [real1004-lean-flag] no v4 tool description still teaches --search-path')
+}
 console.log('passed=' + passed + ' failed=' + failed)
 if (failed) { console.error('FAILURES:'); for (const f of failures) console.error('  - ' + f) }
 if (failed) process.exit(1)