dsh-vibe-math 2.7.2 → 2.8.0

package/tests/audit-installer-policy.test.mjs

@@ -79,7 +79,10 @@ async function applyFrom(pkgDir, home, logs) {
   return logs
 }
 const sha = (buf) => createHash('sha256').update(buf).digest('hex')
-const failedLog = (logs) => logs.filter((l) => l.includes('preset install/update failed'))
+// Label-tolerant on purpose: the installer's user-facing failure line is emitted by the logger helper
+// (not console.*) and its label is localised; the FACT asserted here is that a failure line exists.
+const FAILURE_LINE = /预设安装\/更新失败|preset install\/update failed/
+const failedLog = (logs) => logs.filter((l) => FAILURE_LINE.test(l))
 
 const tmp = mkdtempSync(join(tmpdir(), 'vibe-installer-policy-'))
 const home = join(tmp, 'dshhome')
@@ -153,7 +156,9 @@ let userJsBytes = null
   const state = JSON.parse(readFileSync(stateFile, 'utf8'))
   ok(state.version === '2.0.0', 'the state records the new version', String(state.version))
   ok(logs.some((l) => l.includes('.vibe-math-backup')), 'the log names the backup location')
-  ok(logs.some((l) => l.includes('复制一份')), 'the log tells the user how to customize a preset properly')
+  // A8: the FACT (a version change backs up the user's edit) is asserted structurally above (backup path
+  // + byte equality). What remains is guidance wording, so this is the ONE labelled wording smoke check.
+  ok(logs.some((l) => /复制一份/.test(l)), '[wording smoke check — the ONE allowed] the log tells the user how to customize a preset properly')
   ok(failedLog(logs).length === 0, 'apply() swallowed no failure', failedLog(logs)[0])
 }
 
@@ -164,7 +169,13 @@ console.log('=== 4. a deleted file at the SAME version: restored, nothing backed
   ok(existsSync(at(PRESETS[2], 'agent.cordis.yml')), 'a deleted managed file is restored')
   ok(isCopyOf(PRESETS[2], 'agent.cordis.yml'), 'the restored file is byte-identical to the shipped one')
   ok(!existsSync(join(backupRoot, '2.0.0')), 'a run that replaced nothing creates no backup directory')
-  ok(logs.some((l) => l.includes('restored')), 'the restore is reported')
+  // Guard-sync: the installer's user-visible stream is Chinese by convention (one language, see the
+  // comment at the cleanup line), so these assert the FACT (an action keyword + the count) rather than
+  // an English word.
+  // "one action, one line": the restore must be reported, and reported EXACTLY ONCE (a second report of the
+  // same action is the defect the duplicate-log review was about).
+  const restoreLines = logs.filter((l) => /\u8fd8\u539f\u4e86\s*\d+\s*\u4e2a\u7f3a\u5931\u7684\u9884\u8bbe\u6587\u4ef6/.test(l))
+  ok(restoreLines.length === 1, '★ the restore is reported EXACTLY ONCE (one action, one line)', JSON.stringify(restoreLines.map((l) => l.slice(0, 60))))
   ok(failedLog(logs).length === 0, 'apply() swallowed no failure', failedLog(logs)[0])
 }
 
@@ -182,7 +193,8 @@ console.log('=== 5. a preset this bundle no longer ships ===')
   ok(!existsSync(join(presetRoot, 'vibe-math-v1', 'legacy.js')), 'a package-owned file of a dropped preset is removed')
   ok(!existsSync(join(presetRoot, 'vibe-math-v1')), 'the dropped preset directory is removed once empty')
   ok(existsSync(join(presetRoot, 'vibe-math-v0', 'mine.js')), 'a file recorded as user-owned in a dropped preset is kept')
-  ok(logs.some((l) => l.includes('cleanup')), 'the cleanup is reported')
+  const cleanupLines = logs.filter((l) => /\u9884\u8bbe\u6e05\u7406/.test(l) && /\u5220\u9664\s*\d+\s*\u4e2a\u6587\u4ef6/.test(l))
+  ok(cleanupLines.length === 1, '★ the cleanup is reported EXACTLY ONCE (one action, one line)', JSON.stringify(cleanupLines.map((l) => l.slice(0, 60))))
   ok(failedLog(logs).length === 0, 'apply() swallowed no failure', failedLog(logs)[0])
 }
 
@@ -224,7 +236,9 @@ console.log('=== 7. an unreadable package manifest replaces nothing (and self-he
   ok(readFileSync(at(PRESETS[1], 'preset.yml')).equals(readFileSync(shipped(PRESETS[1], 'preset.yml', pkgC))),
     'the next readable run replaces the drift as a normal version change')
   ok(readFileSync(earlier, 'utf8') === 'EARLIER COPY\n', 'a backup that already exists for that version is kept byte-for-byte (the earliest copy wins)')
-  ok(logs2.every((l) => !l.includes('备份失败')), 'a pre-existing backup for the same version is NOT reported as a failure')
+  // A8: report-level, not wording-level - a pre-existing backup must not be counted as a failure by the
+  // guard's own failure-line helper (which matches the localised failure label, either spelling).
+  ok(failedLog(logs2).length === 0, '★ a pre-existing backup for the same version is NOT reported as a failure (report-level check, not a wording check)', failedLog(logs2)[0])
   ok(JSON.parse(readFileSync(stateFile, 'utf8')).version === '3.0.0', 'the state records the new version')
 
   // the backup directory must never be readable as a preset: DSH only accepts [a-z0-9][a-z0-9-]* ids
@@ -255,7 +269,7 @@ console.log('=== 9. a recorded version NEWER than this package (a downgrade) say
   const backup = join(backupRoot, '3.0.0', PRESETS[0].dst, 'preset.yml')
   ok(existsSync(backup) && readFileSync(backup).equals(newerBytes), 'the replaced newer bytes are backed up under the version they came from')
   ok(JSON.parse(readFileSync(stateFile, 'utf8')).version === '2.0.0', 'the state records the version that is actually installed now')
-  ok(!logs.some((l) => l.includes('preset install/update failed')), 'apply() swallowed no failure')
+  ok(failedLog(logs).length === 0, 'apply() swallowed no failure')
 }
 
 // ============================================================================================
@@ -302,7 +316,7 @@ console.log('=== 10. preset writes are ATOMIC: a failed write cannot touch the d
   const logs = await applyFrom(pkgB, H.home, [])
   ok(readText(H.at(target.p, target.f)) === userBytes.toString(),
     'a failed preset write leaves the destination byte-for-byte intact (tmp+rename, never in place)')
-  ok(logs.some((l) => l.includes('preset install/update failed')),
+  ok(failedLog(logs).length > 0,
     '...and the failure is reported instead of being swallowed', logs.filter((l) => l.includes('failed')).join(' | ').slice(0, 200))
   ok((readJson(H.state) || {}).version === '1.0.0',
     'the state still describes the bytes that are actually on disk, so the next boot self-heals')
@@ -414,7 +428,7 @@ console.log('=== 15. a failed state write is reported (not swallowed), and does
   const logs = await applyFrom(pkgA, H.home, [])
   ok(logs.some((l) => l.includes('状态文件写入失败')),
     'a failed state write is REPORTED (it used to fail silently)', logs.filter((l) => l.includes('写入失败')).join(' | ').slice(0, 160))
-  ok(!logs.some((l) => l.includes('preset install/update failed')), '...without aborting the preset step')
+  ok(failedLog(logs).length === 0, '...without aborting the preset step')
   ok((readJson(H.state) || {}).version === '1.0.0',
     'the previous state file is still parseable (tmp+rename left no half-written file)')
   rmSync(H.state + '.' + process.pid + '.0.tmp', { recursive: true, force: true })
@@ -478,6 +492,25 @@ console.log('=== 16b. a stale file whose BACKUP FAILS is kept, not deleted unbac
   rmSync(join(H.root, '.vibe-math-backup'), { force: true })
 }
 
+console.log('=== 16. one action, one language: no user-facing line carries an English-only label ===')
+{
+  // The convention (docs/AUDIT-CHECKLIST.md): installer user-facing output is Chinese; an English-only
+  // LABEL is a defect (an English domain word inside a Chinese label, like `preset 声明方式`, is fine).
+  // This is a CONVENTION sweep, not a wording pin: any CJK-bearing label passes.
+  const src = readFileSync(INSTALLER_SRC, 'utf8')
+  const labelOf = (l) => { const m = /\[dsh-vibe-math\]\s*([^:：]*)/.exec(l); return m ? m[1].trim() : '' }
+  const asciiOnly = (lab) => lab.length > 0 && /^[\x20-\x7e]+$/.test(lab)
+  const sites = src.split(/\r?\n/).filter((l) => /logger\?\.(info|warn|error)\?\.\(/.test(l))
+  ok(sites.length > 0, 'the installer has user-facing log sites to sweep (found ' + sites.length + ')')
+  const badSites = sites.filter((l) => asciiOnly(labelOf(l)))
+  ok(badSites.length === 0, '★ no user-facing log SITE carries an English-only label (installer output is Chinese)',
+    JSON.stringify(badSites.map((l) => l.trim().slice(0, 70))))
+  const runtimeLogs = await applyFrom(pkgB, home, [])
+  const badRuntime = runtimeLogs.filter((l) => asciiOnly(labelOf(l)))
+  ok(badRuntime.length === 0, '★ ...and no CAPTURED line does either (code and convention agree at runtime)',
+    JSON.stringify(badRuntime.slice(0, 2)))
+}
+
 rmSync(tmp, { recursive: true, force: true })
 
 console.log('')

package/tests/audit-math-computation-contract.mjs

@@ -108,14 +108,48 @@ for (const P of PRESETS) {
     ? "registerTool('vibe_math_set_params'"
     : (P.dir === 'vibe-math-v4' ? "registerTool('vibe_v4_set'" : "registerTool('vibe_v5_set'")
   {
-    const i = js.indexOf(SET_TOOL)
-    const callText = i === -1 ? '' : js.slice(i, js.indexOf('registerTool(', i + 1) === -1 ? i + 4000 : js.indexOf('registerTool(', i + 1))
-    const oi = callText.lastIndexOf('objParams(')
-    const setSchema = oi === -1 ? '' : callText.slice(oi)
-    ok(setSchema.length > 0, tag + 'set tool takes an objParams({...}) schema argument')
-    for (const k of SIX) {
-      ok(new RegExp('(^|[{,\\s])' + k + '\\s*:').test(setSchema), tag + 'set schema has the property ' + k + ' (property-level, not prose)', 'schema len ' + setSchema.length)
-    }
+    // §9.7 ㉓/㉜: EVERY registration site is checked, not just the first. This preset class is defined by
+    // having TWO independent registration layers (session `makeSession` + `apply`, each with its own
+    // `objParams`), so an audit that inspects only the first cannot see the second diverge — the F-B defect
+    // class. If a site ever legitimately carries a narrower schema, record that difference ON THAT SITE
+    // explicitly instead of narrowing the loop back to "first site only".
+    const sites = []
+    for (let at = js.indexOf(SET_TOOL); at !== -1; at = js.indexOf(SET_TOOL, at + 1)) sites.push(at)
+    ok(sites.length >= 1, tag + 'the set tool is registered', 'sites=' + sites.length)
+    // F-B (docs/parameter-schema.md): a schema argument has two shapes —
+    //   objParams({ ...literal... })   the historical hand-written table
+    //   objParams(paramProps())        derived from the SINGLE source PARAM_SCHEMA; the machine key set is
+    //                                  pinned by PARAM_PROPS_KEYS (+ PARAM_PROPS_EXTRA)
+    // The derived form has NO `{` after `objParams(`, so a naive slice grabbed unrelated text (331-char
+    // fragment). Resolve it to the single-source key list, searched FILE-WIDE (that table is module scope).
+    const derivedKeys = (() => {
+      const km = /const PARAM_PROPS_KEYS = \[([^\]]*)\]/.exec(js)
+      if (!km) return null
+      const keys = km[1].split(',').map((s) => s.trim().replace(/^['"]|['"]$/g, '')).filter(Boolean)
+      const em = /const PARAM_PROPS_EXTRA = \{([\s\S]*?)\n\s*\}/.exec(js)
+      const extra = em ? [...em[1].matchAll(/^\s*([A-Za-z_$][\w$]*)\s*:/gm)].map((m) => m[1]) : []
+      return [...new Set(keys.concat(extra))]
+    })()
+    sites.forEach((idx, n) => {
+      const next = js.indexOf('registerTool(', idx + 1)
+      const callText = js.slice(idx, next === -1 ? idx + 4000 : next)
+      const oi = callText.lastIndexOf('objParams(')
+      const setSchema = oi === -1 ? '' : callText.slice(oi)
+      const at = oi === -1 ? -1 : oi + 'objParams('.length
+      const open = at === -1 ? -1 : callText.indexOf('{', at)
+      const close = at === -1 ? -1 : callText.indexOf(')', at)
+      const literalForm = open !== -1 && (close === -1 || open < close)
+      ok(literalForm ? setSchema.length > 0 : !!derivedKeys,
+        tag + 'site#' + n + ' takes an objParams(...) schema argument (literal table or single-source derivation)',
+        literalForm ? 'schema len ' + setSchema.length : 'derived keys ' + (derivedKeys ? derivedKeys.length : 0))
+      for (const k of SIX) {
+        const present = literalForm
+          ? new RegExp('(^|[{,\\s])' + k + '\\s*:').test(setSchema)
+          : !!derivedKeys && derivedKeys.includes(k)
+        ok(present, tag + 'site#' + n + ' set schema has the property ' + k + ' (property-level, not prose)',
+          literalForm ? 'schema len ' + setSchema.length : 'derived keys ' + (derivedKeys ? derivedKeys.length : 0))
+      }
+    })
   }
   if (P.dir === 'vibe-math-v2' || P.dir === 'vibe-math-v3') {
     const schema = arrayBlock(js, 'const PARAM_SCHEMA')
@@ -238,6 +272,28 @@ for (const f of ['README.md', 'README.en.md']) {
   ok(doc.indexOf('Octave / Julia 未安装') !== -1 && doc.indexOf('VERIFY') !== -1 && doc.indexOf('无法强制禁网') !== -1, 'docs §6 keeps the still-unverified set (no Octave/Julia, commercial VERIFY, no network/write enforcement)')
   ok(doc.indexOf('只通过假 subprocess seam 验证') === -1, 'docs §6 no longer claims engine execution was verified ONLY through the fake seam (stale claim removed)')
   ok(doc.indexOf('串行') !== -1 && doc.indexOf('跨进程') !== -1, 'docs §6 states the in-process serialisation and the cross-process limitation')
+  // ── P1/6.1: the v4 frozen participant set is documented as the ATOMIC triple, and the deprecated
+  // alias is described as an alias - both in the doc AND verifiable in the preset source.
+  {
+    const v4 = existsSync(join(ROOT, 'vibe-math-v4/vibe-math-v4.js')) ? read('vibe-math-v4/vibe-math-v4.js') : ''
+    const atomic = /frozen:\s*\(function\(sn\)\{\s*return \{version: sn\?sn\.rosterVersion:null, participants: sn\?sn\.rosterSnapshot:null, kind:/.test(v4)
+    ok(atomic, 'v4 exposes the frozen participant set as ONE atomic object {version, participants, kind}')
+    const alias = /frozenParticipants:\s*\(activeRosterSnapshot\(\)\|\|\{\}\)\.rosterSnapshot\|\|null/.test(v4) && /deprecated/.test(v4)
+    ok(alias, 'v4 keeps `frozenParticipants` only as a deprecated alias of frozen.participants')
+    ok(doc.indexOf('frozen') !== -1 && doc.indexOf('deprecated') !== -1 && doc.indexOf('原子') !== -1,
+      'docs document the atomic `frozen{version,participants,kind}` and mark the alias deprecated')
+  }
+  // ── P1/6.2: the evidence-field list cannot drift from the code again - every timeout evidence field
+  // the module produces must be named in the doc, and vice versa for the two receipt-only ones.
+  {
+    const mod = readFileSync(join(ROOT, 'vibe-math-v2/math-computation.js'), 'utf8')
+    const timeoutFields = ['partialStdout', 'partialStderr', 'stdout', 'stderr', 'exit', 'timedOut']
+    const docMissing = timeoutFields.filter((f) => doc.indexOf(f) === -1)
+    ok(docMissing.length === 0, 'docs name every MATH_TIMEOUT evidence field', 'missing in doc: ' + JSON.stringify(docMissing))
+    const codeMissing = ['partialStdout', 'partialStderr'].filter((f) => mod.indexOf(f) === -1)
+    ok(codeMissing.length === 0, 'the doc-sync check is anchored in the code too (partialStdout/partialStderr exist in the module)',
+      'missing in code: ' + JSON.stringify(codeMissing))
+  }
 }
 
 console.log('')

package/tests/audit-math-computation-contract.mutants.mjs

@@ -0,0 +1,106 @@
+// Registered mutant harness for tests/audit-math-computation-contract.mjs — the audit carries its own proof.
+// Defect shapes (all real, from this project's history):
+//   (a) the SINGLE SOURCE (PARAM_PROPS_KEYS) drops a math param  ⇒ the derived schema loses it
+//   (b) the FIRST (session-layer) registration reverts to a hand-written literal without the math keys
+//   (c) the APPLY-LAYER registration does the same — only visible because the audit now checks EVERY site
+// Each family mutates a COPY of the tree (the audit's MC_CONTRACT_ROOT seam) and must make the audit redden
+// BY NAME. Child runs are bounded (a timeout is a HANG); skipped families are printed and counted.
+// Exit criterion: every family reddens by name; ALL MUTANTS RED AS REQUIRED + exit 0 only then.
+// Run: node tests/audit-math-computation-contract.mutants.mjs
+import { readFileSync, writeFileSync, mkdirSync, copyFileSync, rmSync, readdirSync } from 'node:fs'
+import { execFileSync, spawnSync } from 'node:child_process'
+import { dirname, join } from 'node:path'
+import { tmpdir } from 'node:os'
+import { fileURLToPath } from 'node:url'
+const HERE = dirname(fileURLToPath(import.meta.url))
+const REPO = join(HERE, '..')
+const AUDIT = 'tests/audit-math-computation-contract.mjs'
+// Bound every child run: run-tests has no per-suite timeout, so an unbounded child hangs the gate.
+// A timeout is reported as HANG, counts as a failure for that family, and is listed in hangs=[].
+const CHILD_TIMEOUT_MS = Number(process.env.MUTANT_CHILD_TIMEOUT_MS || 120000)
+const TIMES = []
+const hangs = []
+const skipped = []
+function copyTree(from, to) {
+  mkdirSync(to, { recursive: true })
+  for (const e of readdirSync(from, { withFileTypes: true })) {
+    if (e.name === '.git' || e.name === 'node_modules' || e.name === '_oneoff') continue
+    if (e.isDirectory()) copyTree(join(from, e.name), join(to, e.name))
+    else copyFileSync(join(from, e.name), join(to, e.name))
+  }
+}
+function runFamily(f) {
+  const dest = join(tmpdir(), 'mc-contract-mut-' + Math.random().toString(36).slice(2, 10))
+  copyTree(REPO, dest)
+  const target = join(dest, f.rel)
+  const before = readFileSync(target, 'utf8')
+  const n = before.split(f.from).length - 1
+  if (f.firstOnly ? n < 1 : n !== 1) {
+    skipped.push(f.name + ' (anchor x' + n + ')')
+    console.error('  SKIP - ' + f.name + ' (anchor x' + n + ')')
+    rmSync(dest, { recursive: true, force: true })
+    return false
+  }
+  const mutated = f.firstOnly
+    ? (() => { const i = before.indexOf(f.from); return before.slice(0, i) + f.to + before.slice(i + f.from.length) })()
+    : before.split(f.from).join(f.to)
+  writeFileSync(target, mutated, 'utf8')
+  try { execFileSync(process.execPath, ['--check', target], { stdio: 'pipe' }) }
+  catch (e) {
+    writeFileSync(target, before, 'utf8')
+    skipped.push(f.name + ' (mutation did not compile)')
+    console.error('  SKIP - ' + f.name + ' (mutation did not compile; restored)')
+    rmSync(dest, { recursive: true, force: true })
+    return false
+  }
+  const t0 = Date.now()
+  const r = spawnSync(process.execPath, [AUDIT], {
+    cwd: REPO, encoding: 'utf8', timeout: CHILD_TIMEOUT_MS, killSignal: 'SIGKILL',
+    env: Object.assign({}, process.env, { MC_CONTRACT_ROOT: dest }),
+  })
+  const ms = Date.now() - t0
+  TIMES.push([f.name, ms])
+  const hang = !!(r.error && (r.error.code === 'ETIMEDOUT' || r.signal === 'SIGKILL'))
+  if (hang) hangs.push(f.name + '(' + Math.round(ms / 1000) + 's)')
+  const out = String(r.stdout || '') + String(r.stderr || '')
+  // this audit prints failures as `  - <label> — <detail>` (not `FAIL - ...`), so match the label itself
+  const named = out.split(/\r?\n/).filter((l) => f.expect.test(l)).map((l) => l.trim())
+  const ok = !hang && r.status === 1 && named.length > 0
+  console.log((hang ? '  HANG - ' : (ok ? '  ok - ' : '  FAIL - ')) + f.name + ' [' + ms + 'ms]' + (named.length ? ' :: ' + named[0].slice(0, 150) : ' :: no named red (exit=' + r.status + ')'))
+  rmSync(dest, { recursive: true, force: true })
+  return ok
+}
+const WITHOUT_MATH_KEYS = "objParams({ mode: { type: 'string', enum: ['manual', 'auto'] } })"
+const FAMILIES = [
+  {
+    name: 'the SINGLE SOURCE drops mathMode (derived key table)',
+    rel: 'vibe-math-v3/vibe-math-v3.js',
+    from: "'mathMode', 'mathEngines'",
+    to: "'mathEngines'",
+    expect: /set schema has the property mathMode/,
+  },
+  {
+    name: 'the FIRST (session-layer) registration reverts to a literal without the math keys',
+    rel: 'vibe-math-v3/vibe-math-v3.js',
+    from: 'objParams(paramProps()), async function',
+    to: WITHOUT_MATH_KEYS + ', async function',
+    firstOnly: true,
+    expect: /site#0 set schema has the property mathComputation/,
+  },
+  {
+    name: 'the APPLY-LAYER registration reverts to a literal without the math keys',
+    rel: 'vibe-math-v3/vibe-math-v3.js',
+    from: "objParams(paramProps()), 'vibe_math_set_params')",
+    to: WITHOUT_MATH_KEYS + ", 'vibe_math_set_params')",
+    expect: /site#1 set schema has the property mathComputation/,
+  },
+]
+let red = 0
+for (const f of FAMILIES) { const ok = runFamily(f); if (ok) red++ }
+console.log('')
+console.log('mutant families reddening the math-contract audit by name: ' + red + '/' + FAMILIES.length)
+console.log('timings: ' + TIMES.map((t) => String(t[0]).split('(')[0].trim().slice(0, 18) + '=' + t[1] + 'ms').join('  '))
+console.log('hangs=[' + hangs.join(' | ') + ']')
+console.log('skipped=[' + skipped.join(' | ') + ']')
+if (red !== FAMILIES.length || skipped.length || hangs.length) process.exit(1)
+console.log('ALL MUTANTS RED AS REQUIRED')

package/tests/audit-math-computation-parity.mjs

@@ -17,7 +17,7 @@
 import { createHash } from 'node:crypto'
 import { existsSync, readFileSync } from 'node:fs'
 import { join, resolve } from 'node:path'
-import { fileURLToPath } from 'node:url'
+import { fileURLToPath, pathToFileURL } from 'node:url'
 
 const HERE = fileURLToPath(new URL('./', import.meta.url))
 const REPO = resolve(HERE, '..')
@@ -37,6 +37,101 @@ function ok(cond, label, detail) {
 const sha = (p) => createHash('sha256').update(readFileSync(p)).digest('hex')
 const read = (rel) => readFileSync(join(REPO, rel), 'utf8')
 
+// ---- extracted predicates (shared by the checks below AND by --self-probe) -----------------------
+// §7/§8 used to compute these inline; extracting them is what lets the self-probe feed the SAME
+// predicates deliberately broken strings. Keep the two users in sync by construction: both call these.
+const SUPPORTED_PLACEHOLDERS = {
+  '<script>': "split('<script>').join(",
+  '<expr>': "split('<expr>').join(",
+  '<pkgs>': "a === '<pkgs>'",
+  '<pkgs...>': "a === '<pkgs...>'",
+  '<probeCode>': "a === '<probeCode>'",
+  '<exe>': ".replace(/<exe>/g, exe)",
+  '<pkg>': ".replace(/<pkg>/g, pkg)",
+  '__PKGS__': ".replace(/__PKGS__/g, plain)",
+  '__QPKGS__': ".replace(/__QPKGS__/g, quoted)",
+  '__PKG__': ".replace(/__PKG__/g, pkg)",
+}
+const stringsOfDescriptor = (d) => [].concat(
+  Object.values(d.scriptArgv || {}).filter((x) => typeof x === 'string'),
+  Object.values(d.evalArgv || {}).filter((x) => typeof x === 'string'),
+  ((d.packageProbe && d.packageProbe.argv) || []).filter((x) => typeof x === 'string'),
+  typeof d.probeCode === 'string' ? [d.probeCode] : [],
+  (d.install ? [].concat(d.install.userArgv || [], d.install.systemArgv || [], d.install.uninstallArgv || []) : []).filter((x) => typeof x === 'string'),
+)
+/** Placeholders used by descriptors that the module does not KNOW about. Unit: placeholder occurrences. */
+function unknownPlaceholders(src, all) {
+  const out = []
+  for (const name of Object.keys(all)) {
+    for (const str of stringsOfDescriptor(all[name])) {
+      for (const m of (str.match(/<[a-z.]+>|__[A-Z]+__/g) || [])) if (!(m in SUPPORTED_PLACEHOLDERS)) out.push(name + ':' + m)
+    }
+  }
+  return out
+}
+/** Known placeholders whose implementation is missing from the module. Unit: placeholder occurrences. */
+function unimplementedPlaceholders(src, all) {
+  const out = []
+  for (const name of Object.keys(all)) {
+    for (const str of stringsOfDescriptor(all[name])) {
+      for (const m of (str.match(/<[a-z.]+>|__[A-Z]+__/g) || [])) {
+        if ((m in SUPPORTED_PLACEHOLDERS) && SUPPORTED_PLACEHOLDERS[m] && src.indexOf(SUPPORTED_PLACEHOLDERS[m]) === -1) out.push(name + ':' + m)
+      }
+    }
+  }
+  return out
+}
+/** The substitution must be STRING-level so embedded forms (matlab run('<script>')) are covered. */
+const stringLevelSubstitution = (src) => /String\(a\)\.split\('<script>'\)\.join\(/.test(src) && /String\(a\)\.split\('<expr>'\)\.join\(/.test(src)
+/** §8: the descriptor declares the policy / the module consumes it / nothing is hard-coded. */
+const cliDeclaresPolicy = (eng) => /policy:\s*\{\s*requiresMathMode:\s*'typed\+shell',\s*requiresEngineInList:\s*true\s*\}/.test(eng)
+const cliConsumesPolicy = (mod) => (mod.match(/CLI_POLICY\.requiresMathMode/g) || []).length >= 2 && (mod.match(/CLI_POLICY\.requiresEngineInList/g) || []).length >= 1
+const cliNoHardCoded = (mod) => !/mathMode !== 'typed\+shell'/.test(mod)
+/** §8: the shared refusal shape must surface the descriptor value (the OLD literal had only reason). */
+const refusalSurfacesPolicy = (n) => !!n && n.reason === 'policy' && n.requiresMathMode === 'typed+shell'
+
+// ---- `--self-probe`: the SAME predicates against deliberately broken inputs (shipped form of the
+// was-dev-only proofs `descriptor-sweep-proof.mjs` + `cli-policy-proof.mjs`). No repo mutation.
+async function selfProbe() {
+  const mod = read('vibe-math-v2/math-computation.js')
+  const eng = read('vibe-math-v2/math-engines.js')
+  const { MATH_ENGINES, MATH_P2_ENGINES } = await import(pathToFileURL(join(REPO, 'vibe-math-v2', 'math-engines.js')).href)
+  const all = Object.assign({}, MATH_ENGINES, MATH_P2_ENGINES)
+  const LIT = /<script>/
+  const matlabTemplate = ['-batch', "run('<script>')"]
+  const preFix = (tmpl, abs) => tmpl.map((a) => (a === '<script>' ? abs : a))
+  const postFix = (tmpl, abs) => tmpl.map((a) => String(a).split('<script>').join(abs))
+  const cases = [
+    ['§7 substitution is STRING-level (matlab embedding covered)', stringLevelSubstitution(mod),
+      stringLevelSubstitution(mod.replace("String(a).split('<script>').join(payload.scriptAbs)", "a === '<script>' ? payload.scriptAbs : a"))],
+    ['§7 descriptors use only KNOWN placeholders', unknownPlaceholders(mod, all).length === 0,
+      unknownPlaceholders(mod, Object.assign({}, all, { probe: { scriptArgv: { x: '<mysterySlot>' } } })).length > 0],
+    ['§7 known placeholders are IMPLEMENTED in the module', unimplementedPlaceholders(mod, all).length === 0,
+      unimplementedPlaceholders('/* no implementation */', all).length === 0],
+    ['§7 the guard predicate catches what the OLD substitution produced', postFix(matlabTemplate, 'ABS').every((a) => !LIT.test(a)),
+      preFix(matlabTemplate, 'ABS').every((a) => !LIT.test(a))],
+    ['§8 the module CONSUMES the descriptor policy', cliConsumesPolicy(mod),
+      cliConsumesPolicy(mod.replace(/CLI_POLICY\./g, 'HARDCODED.'))],
+    ['§8 no hard-coded policy comparison remains', cliNoHardCoded(mod),
+      cliNoHardCoded(mod + "\\n  if (params.mathMode !== 'typed+shell') { /* hard-coded revert */ }")],
+    ['§8 descriptor DECLARES the policy', cliDeclaresPolicy(eng),
+      cliDeclaresPolicy(eng.replace("policy: { requiresMathMode: 'typed+shell', requiresEngineInList: true }", 'policy: null'))],
+    ['§8 the shared refusal surfaces the descriptor value (OLD literal caught)',
+      refusalSurfacesPolicy({ reason: 'policy', requiresMathMode: 'typed+shell' }),
+      refusalSurfacesPolicy({ reason: 'policy' })],
+  ]
+  let bad = 0
+  for (const [name, greenNow, redWhenBroken] of cases) {
+    const good = greenNow === true && redWhenBroken === false
+    if (!good) bad++
+    console.log((good ? '  ok   ' : '  FAIL ') + name + ' :: green-now=' + greenNow + ' broken-goes-red=' + !redWhenBroken)
+  }
+  console.log('=== MATH PARITY SELF-PROBE: ' + (cases.length - bad) + '/' + cases.length + ' as required ===')
+  return bad === 0
+}
+if (process.argv.includes('--self-probe')) process.exit(await selfProbe() ? 0 : 1)
+
+
 console.log('-- math_computation parity --')
 
 // ---- 1/2. the four copies are identical, and match the canonical source ----
@@ -106,7 +201,12 @@ for (const f of MODULES) {
     : new URL('../vibe-math-v2/math-engines.js', import.meta.url).href)
   const order = E.MATH_ENGINE_ORDER
   ok(JSON.stringify(order) === JSON.stringify(['python', 'r', 'octave', 'julia', 'matlab', 'maple', 'wolfram', 'cli']), 'MATH_ENGINE_ORDER is exactly the P1 set')
-  ok(E.MATH_ENGINES.cli.defaultOn === true, 'cli is flagged defaultOn')
+  // Item-3 ruling: `defaultOn` was declared but never consumed - cli's default-on property is
+  // already pinned above through the REAL single source (MATH_PARAM_DEFAULTS.mathEngines).
+  const zombie = ['winPrefix', 'stdinArgv', 'defaultOn']
+    .filter((f) => JSON.stringify(E.MATH_ENGINES).indexOf('"' + f + '"') !== -1 || JSON.stringify(E.MATH_P2_ENGINES).indexOf('"' + f + '"') !== -1)
+  ok(zombie.length === 0, 'no zombie descriptor field is declared (winPrefix/stdinArgv/defaultOn were removed)', JSON.stringify(zombie))
+  ok(!/<cliArgv/.test(readFileSync(join(REPO, 'vibe-math-v2', 'math-engines.js'), 'utf8')), 'the decorative <cliArgv...> marker is gone from the descriptor table')
   ok(E.MATH_ENGINES.cli.policy && E.MATH_ENGINES.cli.policy.requiresMathMode === 'typed+shell' && E.MATH_ENGINES.cli.policy.requiresEngineInList === true, 'cli policy gate declared')
   ok(E.MATH_ENGINES.maple.verify === true && !!E.MATH_ENGINES.maple.verifyReason, 'Maple template is flagged VERIFY with a reason')
   for (const name of ['matlab', 'maple', 'wolfram']) {
@@ -192,7 +292,68 @@ for (const f of MODULES) {
   for (const r of reasons) ok(doc.indexOf('`' + r + '`') !== -1, 'docs/math-computation.md names next.reason ' + r)
 }
 
-console.log('')
+// ── 7. descriptor sweep: every argv placeholder dialect must be one the module IMPLEMENTS
+// Sweep finding (HIGH, fixed): substitution was exact-element (`a === '<script>'`), so matlab's embedded
+// `run('<script>')` reached the engine literally. The fix is string-level; this guard keeps the two
+// sides in sync: no descriptor may use a placeholder the module does not substitute.
+{
+  const E = await import(pathToFileURL(join(REPO, 'vibe-math-v2', 'math-engines.js')).href)
+  const modPath = join(REPO, 'vibe-math-v2', 'math-computation.js')
+  const src = existsSync(modPath) ? readFileSync(modPath, 'utf8') : ''
+  const all = Object.assign({}, E.MATH_ENGINES, E.MATH_P2_ENGINES)
+  const unknown = unknownPlaceholders(src, all)
+  const unimplemented = unimplementedPlaceholders(src, all)
+  ok(stringLevelSubstitution(src),
+    'substitution is STRING-level for <script>/<expr> (embedded forms like matlab run(\'<script>\') are covered)')
+}
+
+// ── 8. the cli policy is declared ONCE (descriptor) and consumed, never hard-coded
+// Sweep ruling: `cli.policy.{requiresMathMode,requiresEngineInList}` were declared but never consumed,
+// while the same rules were hard-coded twice in the module (so they could drift). The module now reads
+// the descriptor; this guard makes a hard-coded revert impossible to land.
+{
+  const engPath = join(REPO, 'vibe-math-v2', 'math-engines.js')
+  const modPath = join(REPO, 'vibe-math-v2', 'math-computation.js')
+  const eng = existsSync(engPath) ? readFileSync(engPath, 'utf8') : ''
+  const mod = existsSync(modPath) ? readFileSync(modPath, 'utf8') : ''
+  ok(cliDeclaresPolicy(eng),
+    'the cli descriptor DECLARES its policy (requiresMathMode + requiresEngineInList)')
+  ok(cliConsumesPolicy(mod),
+    'the module CONSUMES the descriptor policy (both refusal sites read CLI_POLICY)')
+  ok(cliNoHardCoded(mod),
+    'no hard-coded policy comparison is left in the module (the rule lives only in the descriptor)')
+}
+
+// ── 9. doc-sync: the v5 editor closed set and v2's single-source status/report helper are documented
+{
+  const v5 = existsSync(join(REPO, 'vibe-math-v5', 'vibe-math-v5.js')) ? readFileSync(join(REPO, 'vibe-math-v5', 'vibe-math-v5.js'), 'utf8') : ''
+  const fp = existsSync(join(REPO, 'docs', 'final-paper.md')) ? readFileSync(join(REPO, 'docs', 'final-paper.md'), 'utf8') : ''
+  // The closed set appears twice in v5 (the coerce call and the tool schema enum); the separator may be
+  // written with or without a space, so accept both rather than pinning one spelling.
+  ok(/\[\s*'office'\s*,\s*'academician'\s*\]/.test(v5) && /V5_INVALID_ARGUMENT/.test(v5), 'v5 validates the paper editor against the closed set and returns V5_INVALID_ARGUMENT')
+  ok(fp.indexOf('office|academician') !== -1 && fp.indexOf('V5_INVALID_ARGUMENT') !== -1, 'docs/final-paper.md documents the v5 editor closed set + its error code')
+  const v2src = existsSync(join(REPO, 'vibe-math-v2', 'vibe-math-v2.js')) ? readFileSync(join(REPO, 'vibe-math-v2', 'vibe-math-v2.js'), 'utf8') : ''
+  const v2doc = existsSync(join(REPO, 'vibe-math-v2', '实现方案.md')) ? readFileSync(join(REPO, 'vibe-math-v2', '实现方案.md'), 'utf8') : ''
+  ok((v2src.match(/verifyTasksView\(/g) || []).length >= 3, 'v2 produces status+report through ONE verifyTasksView helper')
+  ok(v2doc.indexOf('verifyTasksView') !== -1, 'the v2 design doc states the single-source status/report helper')
+}
+
+// S4/S5 (install-failure discrimination) is asserted in tests/math-computation-shared.test.mjs §33, NOT
+// here: parity reads the shipped copy unconditionally, so a mutant cannot reach what it reads (the
+// assertion would exist but could never redden). What parity keeps is the canonical<->copy byte identity,
+// which makes a canonical-only change impossible to ship unnoticed - a stronger defence than a text grep.
+// ── 11. S6: the version policy has ONE wording (constant + helper), not one per path.
+{
+  const mc = readFileSync(join(REPO, 'vibe-math-v2', 'math-computation.js'), 'utf8')
+  ok((mc.match(/version-constraints-are-existence-only: /g) || []).length === 1, '★ S6: the policy wording appears exactly ONCE (in MATH_VERSION_POLICY)')
+  ok(/function mathVersionPolicy\(/.test(mc), 'S6: the shared helper mathVersionPolicy() exists')
+  ok(mc.indexOf('版本求解交给包管理器（本工具只检查是否已安装') === -1, '★ S6: the old divergent install-plan wording is gone')
+  ok(/versionPolicy: mathVersionPolicy\(args\.packages/.test(mc), 'S6: the install plan builds its wording through the helper')
+  ok((mc.match(/只检查是否存在；版本求解交给包管理器/g) || []).length === 1, '★ S6: the policy clause literal appears exactly ONCE (MATH_VERSION_POLICY_CLAUSE)')
+  ok(/const MATH_VERSION_POLICY = .*MATH_VERSION_POLICY_CLAUSE/.test(mc), 'S6: the long policy sentence is DERIVED from the clause')
+  ok(/MATH_MISSING_PACKAGES[\s\S]{0,120}MATH_VERSION_POLICY_CLAUSE/.test(mc), 'S6: the missing-package detail derives the clause')
+}
+
 console.log('=== MATH COMPUTATION PARITY: ' + passed + ' passed, ' + failed + ' failed ===')
 if (failures.length) for (const f of failures) console.error('  - ' + f)
 process.exit(failed === 0 ? 0 : 1)

package/tests/audit-math-computation-sensitivity.mjs

@@ -34,8 +34,10 @@ const PROBES = [
   {
     name: 'cli-policy-ignored',
     file: 'math-computation.js',
-    from: "if (p.mathMode !== 'typed+shell') return bad('engine=cli is disabled while mathMode=' + p.mathMode, 'MATH_REFUSED', next('reason', { reason: 'policy' }))",
-    to: 'if (false) return bad(\'noop\')',
+    // Re-anchored after the single-source ruling: the policy now comes from the DESCRIPTOR (CLI_POLICY),
+    // so the probe disables the descriptor-driven guard. Ignoring it must turn the shared suite RED.
+    from: 'if (CLI_POLICY.requiresMathMode && p.mathMode !== CLI_POLICY.requiresMathMode)',
+    to: 'if (false && CLI_POLICY.requiresMathMode)',
     expect: 'cli refused while mathMode=typed',
   },
   {
@@ -208,7 +210,7 @@ function runSuite(modulePath) {
 for (const p of PROBES) {
   const dir = mkdtempSync(join(tmpdir(), 'mc-sens-' + p.name + '-'))
   copyFileSync(join(SRC, 'math-engines.js'), join(dir, 'math-engines.js'))
-  const src = readFileSync(join(SRC, p.file), 'utf8')
+  const src = readFileSync(join(SRC, p.file), 'utf8').replace(/\r\n/g, '\n')  // anchors are LF: normalize CRLF checkouts
   if (src.indexOf(p.from) === -1) {
     bad++
     console.log('  FAIL ' + p.name + ' — mutation anchor no longer applies (drift): ' + JSON.stringify(p.from.slice(0, 60)))

package/tests/audit-package-membership.mjs

@@ -0,0 +1,57 @@
+#!/usr/bin/env node
+/**
+ * PACKAGE MEMBERSHIP - every guard/harness meant to ship is listed in package.json#files.
+ *
+ * DEFECT CLASS: "shipped guard whose proving harness is not in the tarball" (the ship rule: a guard
+ * ships together with its harness). A package reader could not reproduce the proof.
+ *
+ * The intended shipped set is stated EXPLICITLY, not globbed blindly: every tests/*.mutants.mjs on
+ * disk must be listed, and every tests/*.mjs cited by the guard index must be listed unless it is in
+ * EXCLUDED (documented, deliberate exclusions - e.g. audit-market-metadata.test.mjs asserts its own
+ * absence because it checks repo-only image artifacts).
+ *
+ * Run: node tests/audit-package-membership.mjs
+ * Env: PACKAGE_JSON=<path>  (used by the mutants harness to point at a mutated copy)
+ */
+import { readFileSync, readdirSync, existsSync } from 'node:fs'
+import { dirname, join } from 'node:path'
+import { fileURLToPath } from 'node:url'
+const HERE = dirname(fileURLToPath(import.meta.url))
+const REPO = join(HERE, '..')
+const PKG = process.env.PACKAGE_JSON || join(REPO, 'package.json')
+const EXCLUDED = new Set(['tests/audit-market-metadata.test.mjs', 'tests/audit-diagram-assets.test.mjs'])
+let passed = 0, failed = 0
+const failures = []
+const ok = (cond, label, detail) => { if (cond) { passed++ } else { failed++; failures.push(label + (detail ? ' — ' + detail : '')) } }
+const pkg = JSON.parse(readFileSync(PKG, 'utf8'))
+const listed = new Set(pkg.files)
+const disk = readdirSync(join(REPO, 'tests')).filter((f) => f.endsWith('.mjs'))
+const mutants = disk.filter((f) => /\.mutants\.mjs$/.test(f)).map((f) => 'tests/' + f).sort()
+const unlistedMutants = mutants.filter((f) => !listed.has(f) && !EXCLUDED.has(f))
+ok(mutants.length > 0, 'there is at least one proving harness on disk')
+ok(unlistedMutants.length === 0, 'every tests/*.mutants.mjs on disk is listed in package.json#files', JSON.stringify(unlistedMutants))
+// guard-index citations: tests/*.mjs paths mentioned in the checklist
+const idx = readFileSync(join(REPO, 'docs', 'AUDIT-CHECKLIST.md'), 'utf8')
+const cited = [...new Set((idx.match(/tests\/[A-Za-z0-9._-]+\.mjs/g) || []))]
+const unlistedCited = cited.filter((f) => !listed.has(f) && !EXCLUDED.has(f))
+ok(cited.length > 0, 'the guard index cites at least one test file')
+ok(unlistedCited.length === 0, 'every guard-index-cited test file is listed (or explicitly excluded)', JSON.stringify(unlistedCited))
+ok(listed.has('tests/audit-package-membership.mjs'), 'this audit itself ships')
+ok(!listed.has('tests/audit-market-metadata.test.mjs'), 'the self-excluded repo-only guard is NOT listed (its own assertion)')
+const missingPaths = [...listed].filter((f) => !existsSync(join(REPO, f)))
+// docs sibling rule: a doc cited by a SHIPPED file must ship too (docs/** was previously uncovered).
+const citers = ['README.md', 'README.en.md', 'docs/AUDIT-CHECKLIST.md', 'docs/COMPAT-AUDIT-ROUND2.md']
+const docRefs = new Set()
+for (const c of citers) {
+  const cp = join(REPO, c)
+  if (!existsSync(cp)) continue
+  for (const m of readFileSync(cp, 'utf8').matchAll(/docs\/[A-Za-z0-9._-]+\.md/g)) docRefs.add(m[0])
+}
+ok(docRefs.size > 0, 'shipped files cite at least one docs/*.md')
+const unlistedDocs = [...docRefs].filter((f) => !listed.has(f))
+ok(unlistedDocs.length === 0, 'every docs/*.md cited by a shipped file is listed in package.json#files', JSON.stringify(unlistedDocs))
+ok(missingPaths.length === 0, 'every listed path exists', JSON.stringify(missingPaths))
+console.log('')
+console.log('=== PACKAGE MEMBERSHIP: ' + passed + ' passed, ' + failed + ' failed ===')
+for (const f of failures) console.error('  - ' + f)
+process.exit(failed === 0 ? 0 : 1)

package/tests/audit-package-membership.mutants.mjs

@@ -0,0 +1,28 @@
+#!/usr/bin/env node
+/**
+ * Mutant harness for the package-membership audit: drop one listed entry from a COPY of package.json
+ * and require the audit to go named-red. Run: node tests/audit-package-membership.mutants.mjs
+ */
+import { readFileSync, writeFileSync, mkdtempSync, rmSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { spawnSync } from 'node:child_process'
+import { fileURLToPath } from 'node:url'
+const HERE = fileURLToPath(new URL('./', import.meta.url))
+const REPO = join(HERE, '..')
+let passed = 0, failed = 0
+const ok = (c, l, d) => { if (c) { passed++; console.log('  ok   ' + l) } else { failed++; console.log('  FAIL ' + l + (d ? ' — ' + d : '')) } }
+const base = spawnSync(process.execPath, [join(HERE, 'audit-package-membership.mjs')], { cwd: REPO, encoding: 'utf8' })
+ok(base.status === 0, 'baseline: the membership audit is green', String(base.stdout || '').trim().split('\n').slice(-1)[0])
+const dir = mkdtempSync(join(tmpdir(), 'pkgmem-'))
+const pkg = JSON.parse(readFileSync(join(REPO, 'package.json'), 'utf8'))
+const victim = 'tests/audit-package-membership.mjs'
+pkg.files = pkg.files.filter((f) => f !== victim)
+writeFileSync(join(dir, 'package.json'), JSON.stringify(pkg, null, 2))
+const mut = spawnSync(process.execPath, [join(HERE, 'audit-package-membership.mjs')], { cwd: REPO, encoding: 'utf8', env: Object.assign({}, process.env, { PACKAGE_JSON: join(dir, 'package.json') }) })
+const out = String(mut.stdout || '') + String(mut.stderr || '')
+ok(mut.status !== 0 && /this audit itself ships/.test(out), '★ removing a listed entry makes the membership audit RED (named assertion)', 'exit=' + mut.status)
+rmSync(dir, { recursive: true, force: true })
+console.log('')
+console.log('=== PACKAGE MEMBERSHIP MUTANTS: ' + passed + ' passed, ' + failed + ' failed ===')
+process.exit(failed === 0 ? 0 : 1)