dsh-vibe-math 2.7.2 → 2.7.3

package/tests/audit-status-report-fields.mutants.mjs

@@ -0,0 +1,64 @@
+#!/usr/bin/env node
+/**
+ * Mutant harness for tests/audit-status-report-fields.mjs (ships: `tests/*.mutants.mjs`).
+ *
+ * ONE MUTANT PER DIRECTION (the doc↔source comparison must bite both ways):
+ *   A) doc→source: delete one v2 row from a COPY of docs/status-report-fields.md -> the "every source
+ *      field has a doc row" assertion reddens by name
+ *   B) source→doc: add a field to a COPY of vibe-math-v2.js's `fieldScopes.session` -> the "every doc row
+ *      names a real source field" / missing-row assertion reddens by name
+ *
+ * Run: node tests/audit-status-report-fields.mutants.mjs
+ */
+import { mkdtempSync, readFileSync, writeFileSync, 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, '..')
+const GUARD = join(HERE, 'audit-status-report-fields.mjs')
+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 run = (env) => {
+  const r = spawnSync(process.execPath, [GUARD], { cwd: REPO, encoding: 'utf8', env: Object.assign({}, process.env, env || {}) })
+  const out = String(r.stdout || '') + String(r.stderr || '')
+  return { status: r.status, failLines: out.split('\n').filter((l) => /^  - /.test(l)), tail: out.trim().split('\n').filter(Boolean).slice(-1)[0] || '' }
+}
+const baseline = run(null)
+ok(baseline.status === 0, 'baseline: the field-table guard is green', baseline.tail.slice(0, 80))
+
+const dir = mkdtempSync(join(tmpdir(), 'fieldtable-mut-'))
+// A) doc -> source
+const docPath = join(REPO, 'docs/status-report-fields.md')
+const docLines = readFileSync(docPath, 'utf8').split(/\r?\n/)
+const v2Head = docLines.findIndex((l) => l.startsWith('### v2 的字段与作用域'))
+const firstRow = docLines.findIndex((l, i) => i > v2Head && /^\|\s*`[^`]+`\s*\|/.test(l))
+if (v2Head === -1 || firstRow === -1) ok(false, 'A: the v2 section and its first row exist (harness and guard agree on the shape)', 'ANCHOR MISS')
+else {
+  const removed = docLines[firstRow]
+  const tampered = docLines.slice(0, firstRow).concat(docLines.slice(firstRow + 1)).join('\n')
+  writeFileSync(join(dir, 'status-report-fields.md'), tampered)
+  const a = run({ FIELD_DOC: join(dir, 'status-report-fields.md') })
+  const named = a.failLines.some((l) => l.indexOf('v2: every source field has a doc row') !== -1)
+  ok(a.status !== 0 && named, '★ A (doc→source): deleting a v2 row reddens the guard by name', 'removed=' + removed.trim().slice(0, 60) + ' exit=' + a.status)
+  if (!named) for (const l of a.failLines.slice(0, 3)) console.log('      guard failure: ' + l.trim().slice(0, 130))
+}
+// B) source -> doc
+const v2Path = join(REPO, 'vibe-math-v2/vibe-math-v2.js')
+const v2src = readFileSync(v2Path, 'utf8')
+const m = /(\n\s*session:\s*\[)([\s\S]*?)(\])/.exec(v2src)
+if (!m) ok(false, 'B: the v2 fieldScopes session array exists (anchor check)', 'ANCHOR MISS')
+else {
+  const injected = m[1] + m[2].replace(/\s*$/, '') + ", 'ghostFieldFromMutant'" + m[3]
+  writeFileSync(join(dir, 'vibe-math-v2.js'), v2src.replace(m[0], injected))
+  const b = run({ FIELD_V2_JS: join(dir, 'vibe-math-v2.js') })
+  const named = b.failLines.some((l) => l.indexOf('v2: every source field has a doc row') !== -1 || l.indexOf('v2: every doc row names a real source field') !== -1)
+  ok(b.status !== 0 && named, '★ B (source→doc): adding a field to the source reddens the guard by name', 'exit=' + b.status)
+  if (!named) for (const l of b.failLines.slice(0, 3)) console.log('      guard failure: ' + l.trim().slice(0, 130))
+}
+rmSync(dir, { recursive: true, force: true })
+console.log('')
+console.log('=== STATUS FIELD TABLE MUTANTS: ' + passed + ' passed, ' + failed + ' failed ===')
+process.exit(failed === 0 ? 0 : 1)

package/tests/audit-v3-registration-parity.mjs

@@ -0,0 +1,156 @@
+/**
+ * Probe for prompt-audit M3 (v3): the two registerTool paths must not drift.
+ *
+ * v3 declares each tool twice — the session handler table inside makeSession()
+ * (registerTool(name, description, parameters, executeFn)) and the real
+ * `tools.register` at apply scope (registerTool(name, description, parameters,
+ * handlerName)). The parameter schemas are guarded by audit-prompt-invariants
+ * (I13/I14); the DESCRIPTIONS were not, and five of them had already drifted.
+ *
+ * Checks per tool (33 of them):
+ *   - MISSING      : registered on only one path
+ *   - DESC-DRIFT   : the description argument text differs between the two paths
+ *   - SCHEMA-DRIFT : the parameter-schema argument text differs
+ *   - NO-SOURCE    : a description is not read from the shared module-level table
+ *                    (the "one source of truth" claim would be nominal only)
+ *   - BAD-REF      : a TOOL_DESC.<name> reference has no matching entry in the table
+ *
+ * Read-only: parses the source text, never executes the plugin.
+ *
+ * Sensitivity probe (the suite must be falsifiable):
+ *   node tests/audit-v3-registration-parity.mjs --self-probe '["<from>","<to>"]'
+ *   The pair is applied to the source text, so a first-occurrence mutation makes the
+ *   second path disagree and the probe must exit 1.
+ */
+import { readFileSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import { dirname, join } from 'node:path';
+
+const HERE = dirname(fileURLToPath(import.meta.url));
+let SRC = readFileSync(join(HERE, '..', 'vibe-math-v3', 'vibe-math-v3.js'), 'utf8');
+
+const SELF_PROBE = process.argv.indexOf('--self-probe');
+if (SELF_PROBE !== -1 && process.argv[SELF_PROBE + 1]) {
+  const [from, to] = JSON.parse(process.argv[SELF_PROBE + 1]);
+  const before = SRC;
+  SRC = SRC.replace(from, to);
+  if (SRC === before) { console.log('MUTATION DID NOT APPLY: ' + from); process.exit(2); }
+}
+
+/** Skip a string literal (any quote) starting at i; returns the index after it. */
+function skipString(s, i) {
+  const q = s[i];
+  i++;
+  while (i < s.length) {
+    if (s[i] === '\\') { i += 2; continue; }
+    if (s[i] === q) return i + 1;
+    i++;
+  }
+  return i;
+}
+
+// Normalise line endings: this file is checked out with CRLF on Windows, and a stray \r
+// between an entry's `,` and its key would make the table parser under-count (it did).
+SRC = SRC.replace(/\r/g, '');
+
+/** Collect a top-level-comma-separated argument list starting at `open` (an index of `(`). */
+function argsOf(s, open) {
+  let depth = 0;
+  const args = [];
+  let cur = '';
+  for (let i = open; i < s.length; i++) {
+    const c = s[i];
+    if (c === '"' || c === "'" || c === '`') { const j = skipString(s, i); cur += s.slice(i, j); i = j - 1; continue; }
+    if (c === '(' || c === '[' || c === '{') { depth++; cur += c; continue; }
+    if (c === ')' || c === ']' || c === '}') {
+      depth--;
+      if (depth === 0) { args.push(cur.trim()); return args; }
+      cur += c; continue;
+    }
+    if (c === ',' && depth === 1) { args.push(cur.trim()); cur = ''; continue; }
+    if (depth >= 1) cur += c;
+  }
+  return args;
+}
+
+const calls = [];
+for (const m of SRC.matchAll(/registerTool\(\s*'([A-Za-z0-9_]+)'/g)) {
+  const open = SRC.indexOf('(', m.index + 'registerTool'.length);
+  calls.push({ name: m[1], at: m.index, args: argsOf(SRC, open) });
+}
+
+const byName = new Map();
+for (const c of calls) {
+  if (!byName.has(c.name)) byName.set(c.name, []);
+  byName.get(c.name).push(c);
+}
+
+// The shared description table (module-level `const TOOL_DESC = { … }`).
+//
+// Keys are collected with an ANCHORED PER-LINE regex inside the table's own text slice —
+// string values are blanked out first (keeping every newline) so the `{`/`}` braces inside
+// the sync_meta documentation text cannot shift a brace-depth scanner (that mis-parse is
+// exactly how an earlier version of this probe silently saw 15 of 33 entries).
+const tableAt = SRC.search(/const\s+TOOL_DESC\s*=\s*\{/);
+const tableKeys = new Set();
+if (tableAt !== -1) {
+  const open = SRC.indexOf('{', tableAt);
+  // slice to the matching close brace with strings blanked (newlines preserved)
+  let depth = 0;
+  let end = -1;
+  let blanked = '';
+  for (let i = open; i < SRC.length; i++) {
+    const c = SRC[i];
+    if (c === '"' || c === "'" || c === '`') {
+      const stop = skipString(SRC, i);
+      for (let k = i; k < stop; k++) blanked += (SRC[k] === '\n' ? '\n' : ' ');
+      i = stop - 1;
+      continue;
+    }
+    blanked += c;
+    if (c === '{') depth++;
+    else if (c === '}') { depth--; if (depth === 0) { end = i; break; } }
+  }
+  const tableText = blanked.slice(0, end === -1 ? blanked.length : end - open + 1);
+  for (const m of tableText.matchAll(/^[ \t]*([A-Za-z_$][\w$]*)\s*:/gm)) tableKeys.add(m[1]);
+}
+
+let descDrift = 0;
+let schemaDrift = 0;
+let noSource = 0;
+let badRef = 0;
+const missing = [];
+for (const [name, list] of byName) {
+  if (list.length !== 2) { missing.push(name + ' (' + list.length + ' registration(s))'); continue; }
+  const [a, b] = list.map((c) => ({ desc: c.args[1], schema: c.args[2] }));
+  if (a.desc !== b.desc) {
+    descDrift++;
+    console.log('DESC-DRIFT   ' + name);
+    console.log('   path#1: ' + String(a.desc).slice(0, 140));
+    console.log('   path#2: ' + String(b.desc).slice(0, 140));
+  }
+  if (a.schema !== b.schema) {
+    schemaDrift++;
+    console.log('SCHEMA-DRIFT ' + name);
+    console.log('   path#1: ' + String(a.schema).slice(0, 140));
+    console.log('   path#2: ' + String(b.schema).slice(0, 140));
+  }
+  for (const d of [a.desc, b.desc]) {
+    const want = 'TOOL_DESC.' + name;
+    if (d === want) continue;
+    if (/^TOOL_DESC\./.test(d)) { badRef++; console.log('BAD-REF      ' + name + ' → ' + d); continue; }
+    noSource++;
+    console.log('NO-SOURCE    ' + name + ' → ' + String(d).slice(0, 140));
+  }
+}
+
+console.log('shared TOOL_DESC entries: ' + tableKeys.size);
+console.log('tools registered twice: ' + [...byName.values()].filter((v) => v.length === 2).length + ' / ' + byName.size);
+console.log('DESC-DRIFT count: ' + descDrift);
+console.log('SCHEMA-DRIFT count: ' + schemaDrift);
+console.log('descriptions not from the shared table: ' + noSource);
+console.log('bad TOOL_DESC references: ' + badRef);
+console.log('one-path-only tools: ' + (missing.length ? missing.join(', ') : 'none'));
+const ok = descDrift === 0 && schemaDrift === 0 && noSource === 0 && badRef === 0 && missing.length === 0 && tableKeys.size === byName.size;
+console.log(ok ? '=== DESC PARITY OK ===' : '=== DESC PARITY FAILED ===');
+process.exit(ok ? 0 : 1);

package/tests/audit-v5-integrity.mjs

@@ -9,13 +9,125 @@
 //   5. leftover development markers / TODO scaffolding
 // Run: node tests/audit-v5-integrity.mjs   (exit 1 on any finding)
 // ============================================================
-import { readFileSync, existsSync, mkdtempSync, writeFileSync, rmSync } from 'node:fs'
+import { readFileSync, existsSync, mkdtempSync, writeFileSync, rmSync, readdirSync } from 'node:fs'
 import { spawnSync } from 'node:child_process'
 import { tmpdir } from 'node:os'
 import { join } from 'node:path'
+import { fileURLToPath } from 'node:url'
 
-const FILE = new URL('../vibe-math-v5/vibe-math-v5.js', import.meta.url)
-const raw = readFileSync(FILE, 'utf8')
+const HERE = new URL('../', import.meta.url)
+/**
+ * `V5_INTEGRITY_MUTATE` carries a JSON `[rel, from, to]` triple: that ONE repo-relative file is mutated
+ * IN MEMORY for one child run. Combined with `--self-probe` this makes every guard below reproducible
+ * from OUTSIDE — a verifier no longer has to copy this audit and rewrite its URLs to point it at a
+ * mutated copy (R16's request). Same shape as `audit-prompt-invariants.mjs` and the
+ * `math-computation-shared --self-probe` pattern.
+ *
+ * SEAM COVERAGE (R17 measured, both directions): the seam reaches a **plugin** path (a plugin mutation
+ * produces 1 finding) AND the **corpus** path (a corpus mutation produces 2 findings) — i.e. it is not
+ * limited to the plugin file, and a corpus-only drift is observable from outside too. The corpus is read
+ * through the same `readRaw()`, so any new corpus check inherits the seam.
+ */
+const MUT = process.env.V5_INTEGRITY_MUTATE
+function readRaw(rel) {
+  let text = readFileSync(new URL(rel, HERE), 'utf8')
+  if (MUT) {
+    try {
+      const [r, from, to] = JSON.parse(MUT)
+      if (r === rel && from) text = text.replace(from, to)
+    } catch (e) { /* a malformed mutation is a harness error, not an audit failure */ }
+  }
+  return text
+}
+
+/**
+ * Each mutation is a REAL defect shape for one of this audit's guards. `expect` is a substring that MUST
+ * appear in the failing run; `control: true` marks the unmutated run, which must stay green.
+ */
+const SELF_PROBE_MUTATIONS = [
+  { name: 'control (no mutation)', rel: '', from: '', to: '', expect: '', control: true },
+  {
+    name: 'F3: a status() top-level key is renamed (the frozen field surface must redden)',
+    rel: 'vibe-math-v5/vibe-math-v5.js',
+    from: 'leanNoticesScope: ',
+    to: 'leanNoticesScopeRenamed: ',
+    expect: 'status() field surface changed',
+  },
+  {
+    name: 'F1: the README resume claim loses its in-instance case (the pairing gate must redden)',
+    rel: 'README.md',
+    from: '**同实例重建**（宿主丢了子会话、插件实例还在）**继续**原编号',
+    to: '**同实例重建**从 1 重新开始',
+    expect: 'README resume-claim side missing',
+  },
+  {
+    name: 'F1: a code anchor the README cites is removed (its philosophy gate must redden)',
+    rel: 'vibe-math-v5/vibe-math-v5.js',
+    from: 'rounds.set(member.id, startRound)',
+    to: 'void 0 /* anchor removed (self-probe) */',
+    expect: 'philosophy gate missing from the implementation: the founding round is applied only AFTER a successful start',
+  },
+  {
+    name: 'F2/G1: the diagram is mutated back to the pre-G1 inbox order (the pairing gate must redden)',
+    rel: 'vibe-math-v5/架构图.md',
+    from: '唯一的推进驱动',
+    to: '唯一的推进驱动（先 ack 再构造）',
+    expect: '架构图 still states the pre-G1 order',
+  },
+  {
+    name: 'F2/G1: the plan is mutated back to the explicit pre-G1 inbox phrasing (the pairing gate must redden)',
+    rel: 'vibe-math-v5/实现方案.md',
+    from: '**未 ack ⇒ 仍可重投**',
+    to: '**先 ack 邮件、再构造**；**未 ack ⇒ 仍可重投**',
+    expect: '实现方案.md still states the pre-G1 order',
+  },
+  {
+    name: 'F2/G1: a docs file is mutated back to the named pre-G1 phrasing (the docs-wide gate must redden)',
+    rel: 'docs/AUDIT-CHECKLIST.md',
+    from: '身份一律显式传递，绝不猜测。',
+    to: '身份一律显式传递，绝不猜测。先 ack 再构造',
+    expect: 'still states the pre-G1 order',
+  },
+  {
+    name: 'F2/G1: the plan is mutated to re-introduce the pre-G1 re-send contract (the named check must redden)',
+    rel: 'vibe-math-v5/实现方案.md',
+    from: '`makeFileBackend`',
+    to: '`makeFileBackend`（queued 永不重发）',
+    expect: 'still states the pre-G1 re-send contract',
+  },
+]
+
+if (process.argv.includes('--self-probe')) {
+  const bad = []
+  for (const mut of SELF_PROBE_MUTATIONS) {
+    const r = spawnSync(process.execPath, [fileURLToPath(import.meta.url), '--json'], {
+      cwd: fileURLToPath(HERE),
+      env: Object.assign({}, process.env, mut.control ? {} : { V5_INTEGRITY_MUTATE: JSON.stringify([mut.rel, mut.from, mut.to]) }),
+      encoding: 'utf8',
+    })
+    let parsed = null
+    try { parsed = JSON.parse(r.stdout) } catch (e) { /* fall through to the diagnostic below */ }
+    const text = (r.stdout || '') + (r.stderr || '')
+    if (mut.control) {
+      const ok = r.status === 0 && parsed && parsed.failed === 0
+      console.log((ok ? 'PASS ' : 'FAIL ') + 'control: the unmutated run stays green (' + (parsed ? parsed.failed + ' findings, ' + parsed.gates + ' gates' : 'unparsable output') + ')')
+      if (!ok) bad.push('control run was not green')
+      continue
+    }
+    const hit = text.includes(mut.expect)
+    const red = r.status === 1 && parsed && parsed.failed > 0
+    const ok = red && hit
+    console.log((ok ? 'PASS ' : 'FAIL ') + mut.name + (ok ? '' : '  → exit=' + r.status + ' hit=' + hit + ' failed=' + (parsed && parsed.failed)))
+    if (!ok) bad.push(mut.name + ' (exit ' + r.status + ', expected ' + JSON.stringify(mut.expect) + ')')
+  }
+  console.log('')
+  console.log('V5 INTEGRITY SELF-PROBE: ' + (SELF_PROBE_MUTATIONS.length - bad.length) + '/' + SELF_PROBE_MUTATIONS.length + ' as required')
+  for (const b of bad) console.error('  FAIL ' + b)
+  process.exit(bad.length === 0 ? 0 : 1)
+}
+
+const FILE_REL = 'vibe-math-v5/vibe-math-v5.js'
+const raw = readRaw(FILE_REL)
 // Strip comments, string literals AND regex literals before any identifier scan. Without this the
 // heuristic matches English words inside comments that merely precede a '(' (e.g.
 // "// per unit (" becomes a phantom call to unit()), drowning the real findings.
@@ -79,6 +191,8 @@ function stripNoise(s) {
 const src = stripNoise(raw)
 const findings = []
 const notes = []
+// How many philosophy gates the run checked (module scope so `--json` can report it).
+let gateCount = 0
 
 const lineOf = (idx) => src.slice(0, idx).split('\n').length
 
@@ -213,7 +327,7 @@ else {
 const raised = new Set()
 for (const m of raw.matchAll(/v5err\(\s*'(V5_[A-Z_]+)'/g)) raised.add(m[1])
 for (const m of raw.matchAll(/code:\s*'(V5_[A-Z_]+)'/g)) raised.add(m[1])
-const docs = readFileSync(new URL('../vibe-math-v5/实现方案.md', import.meta.url), 'utf8')
+const docs = readRaw('vibe-math-v5/实现方案.md')
 const documented = new Set()
 for (const m of docs.matchAll(/`(V5_[A-Z_]+)`/g)) documented.add(m[1])
 const onlyDocs = new Set()
@@ -262,8 +376,8 @@ function packageRows(yaml) {
   }
   return rows
 }
-const v5yaml = readFileSync(new URL('../vibe-math-v5/agent.cordis.yml', import.meta.url), 'utf8')
-const v4yaml = readFileSync(new URL('../vibe-math-v4/agent.cordis.yml', import.meta.url), 'utf8')
+const v5yaml = readRaw('vibe-math-v5/agent.cordis.yml')
+const v4yaml = readRaw('vibe-math-v4/agent.cordis.yml')
 const v5rows = packageRows(v5yaml)
 const v4names = new Set(packageRows(v4yaml).map(r => r.name))
 const RELATIVE_OK = new Set(['./vibe-math-v5.js'])
@@ -293,7 +407,7 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
 // against the implementation: a name in the plan but not the code is an unimplemented
 // requirement, and a name in the code but not the plan is undocumented surface.
 {
-  const plan = readFileSync(new URL('../vibe-math-v5/实现方案.md', import.meta.url), 'utf8')
+  const plan = readRaw('vibe-math-v5/实现方案.md')
   const planTools = new Set()
   for (const m of plan.matchAll(/\bvibe_v5_[a-z_]+/g)) planTools.add(m[0])
   const codeTools = new Set()
@@ -325,6 +439,158 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
   }
   notes.push('plan §16 parameter rows: ' + paramRows.size)
 
+  // F3 (docs-vs-behaviour sweep): `status()`'s machine surface. OWNERSHIP (agreed division, so the two
+  // guards never both claim the same job):
+  //   • FIELD-LEVEL reconciliation (field ↔ doc row, BOTH directions) is owned by
+  //     `tests/audit-status-report-fields.mjs`, which carries the table and prints the
+  //     `frozen / doc-rows / exceptions` counts;
+  //   • THIS audit keeps ONLY this frozen-key drift check: the extracted 43 keys below, the printed
+  //     `captured: N`, and the `<30` anti-vacuity floor. Any addition/removal of a top-level key must be
+  //     an explicit decision here (and in the owning guard's reconciliation).
+  // No "doc half" is promised or planned here — the note printed at the end points at the owner instead.
+  const stIdx = raw.indexOf('function status() {')
+  const statusKeys = []
+  if (stIdx === -1) findings.push('could not locate status() — the machine-surface extraction has no anchor')
+  else {
+    const open = raw.indexOf('return {', stIdx)
+    const close = raw.indexOf('\n      }', open)
+    // Tokenise the returned literal: every key directly inside it — including SHORTHAND keys that share a
+    // line (`institute: instituteName, project, key, phase,`) — while SKIPPING each value (so an
+    // identifier used as a value can never be mistaken for a key).
+    const body = open === -1 ? '' : raw.slice(open + 'return {'.length, close === -1 ? open + 9000 : close)
+    // Comments must be skipped IN PLACE: the payload's comments contain prose with apostrophes, parens and
+    // commas (`…delivered ONCE…`, `phase:'failed'`, `(installer review)`), which otherwise corrupt both the
+    // depth tracking and the value skipping.
+    const skipComment = (s, at) => {
+      if (s[at + 1] === '/') { let j = at + 2; while (j < s.length && s[j] !== '\n') j += 1; return j }
+      if (s[at + 1] === '*') { const j = s.indexOf('*/', at + 2); return j === -1 ? s.length : j + 2 }
+      return -1
+    }
+    let i = 0, depth = 0
+    while (i < body.length) {
+      const ch = body[i]
+      if (ch === '/') { const j = skipComment(body, i); if (j !== -1) { i = j; continue } }
+      if (ch === '"' || ch === "'") { const q = ch; i += 1; while (i < body.length && body[i] !== q) { if (body[i] === '\\') i += 1; i += 1 } i += 1; continue }
+      if (ch === '{' || ch === '[' || ch === '(') { depth += 1; i += 1; continue }
+      if (ch === '}' || ch === ']' || ch === ')') { if (depth === 0) break; depth -= 1; i += 1; continue }
+      if (depth === 0 && /[A-Za-z_$]/.test(ch)) {
+        const m = /^[A-Za-z_$][\w$]*/.exec(body.slice(i))
+        const after = body.slice(i + m[0].length).replace(/^[\s]+/, '')
+        if (after[0] === ':') {
+          statusKeys.push(m[0])
+          i += m[0].length
+          let d = 0
+          while (i < body.length) {                       // skip the VALUE up to the next top-level comma
+            const c = body[i]
+            if (c === '/') { const j = skipComment(body, i); if (j !== -1) { i = j; continue } }
+            if (c === '"' || c === "'") { const q = c; i += 1; while (i < body.length && body[i] !== q) { if (body[i] === '\\') i += 1; i += 1 } i += 1; continue }
+            if (c === '{' || c === '[' || c === '(') { d += 1; i += 1; continue }
+            if (c === '}' || c === ']' || c === ')') { if (d === 0) break; d -= 1; i += 1; continue }
+            if (c === ',' && d === 0) { i += 1; break }
+            i += 1
+          }
+          continue
+        }
+        if (after[0] === ',') { statusKeys.push(m[0]); i += m[0].length; continue }
+        i += m[0].length; continue
+      }
+      i += 1
+    }
+  }
+  notes.push('status() top-level keys captured: ' + statusKeys.length)
+  if (statusKeys.length < 30) findings.push('status() key extraction captured only ' + statusKeys.length + ' keys (anchor or payload broke — this check must never pass vacuously)')
+  for (const must of ['phase', 'members', 'quorum', 'chat', 'persistence']) {
+    if (statusKeys.indexOf(must) === -1) findings.push('status() no longer exposes `' + must + '` (either the extraction anchor or the payload broke)')
+  }
+  const EXPECTED_STATUS_KEYS = ['ok', 'institute', 'project', 'key', 'phase', 'running', 'autoDone', 'runId', 'leanNotices', 'leanNoticesScope', 'fieldScopes', 'backend', 'diagnostics', 'debug', 'quorum', 'members', 'tasks', 'failedMembers', 'failedMembersNote', 'pendingSpawns', 'pendingSpawnsNote', 'chat', 'chatScope', 'officeRequests', 'officeRequestsShown', 'officeRequestsCap', 'officeRequestsDropped', 'officeRequestsTruncated', 'persistence', 'meeting', 'parkedMeeting', 'verify', 'verifyQueue', 'verified', 'verifiedTrue', 'concludedFalse', 'verifiedNote', 'undecided', 'solveVotes', 'formal', 'paper', 'lastProgressAt', 'params']
+  const statusMissing = EXPECTED_STATUS_KEYS.filter((k) => statusKeys.indexOf(k) === -1)
+  const statusAdded = statusKeys.filter((k) => EXPECTED_STATUS_KEYS.indexOf(k) === -1)
+  if (statusMissing.length || statusAdded.length) {
+    findings.push('status() field surface changed — missing=' + JSON.stringify(statusMissing) + ' added=' + JSON.stringify(statusAdded) +
+      ': record the decision in this freeze AND let the field-level owner reconcile the documents (tests/audit-status-report-fields.mjs)')
+  }
+  notes.push('status() field-surface drift: frozen=' + statusKeys.length + ' keys (this audit owns ONLY this freeze + the <30 anti-vacuity floor); ' +
+    'field-level reconciliation (field ↔ doc row, both directions) is OWNED BY tests/audit-status-report-fields.mjs, which prints frozen / doc-rows / exceptions')
+
+  // F1 (docs-vs-behaviour): the README's TWO-CASE `resume` claim must stay paired with the behavioural
+  // anchor that pins the in-instance case (the two CODE anchors are gates in the GATES table below). Only
+  // the prose side is checked here, and every side reports how much it matched, so this cannot pass vacuously.
+  {
+    const readme = readRaw('README.md')
+    const suite = readRaw('tests/e2e-v5-round2.test.mjs')
+    const readmeClaims = [
+      ['the two-case framing', /`resume`\s*后轮次计数的两种情形/],
+      ['case 1 = a cross-process reload counts from 1', /跨进程重载[^\n]*从\s*\*\*1\*\*\s*重新计/],
+      ['case 2 = an in-instance rebuild CONTINUES the numbering (轮次 2)', /同实例重建[^\n]*继续[^\n]*轮次 2/],
+      ['a failed/invalid start no longer erases the existing count', /失败\/无效的启动\*\*不再抹掉\*\*已有计数/],
+      ['the claim is pinned to a revision', /口径复核于\s*`?[0-9a-f]{7,40}`?/],
+    ]
+    let readmeHits = 0
+    for (const [label, re] of readmeClaims) {
+      if (re.test(readme)) readmeHits += 1
+      else findings.push('README resume-claim side missing: ' + label)
+    }
+    const suiteClaims = [
+      ['the in-instance rebuild is asserted as 轮次 2', /a FAILED resume must not move the round number backwards/],
+      ['the founding round is asserted as 轮次 1', /the founding prompt starts the member at 轮次 1/],
+    ]
+    let suiteHits = 0
+    for (const [label, re] of suiteClaims) {
+      if (re.test(suite)) suiteHits += 1
+      else findings.push('README resume-claim has no behavioural anchor: ' + label)
+    }
+    // The README cites exact code/test lines PINNED to a revision ("口径复核于 <sha>"). Those drift by
+    // design, so the drift is a measured NOTE carrying the current lines — never a failure.
+    const lineOf = (needle) => { const at = raw.indexOf(needle); return at === -1 ? -1 : raw.slice(0, at).split('\n').length }
+    const startLine = lineOf('const startRound = (rounds.get(member.id) || 0) + 1')
+    const applyLine = lineOf('rounds.set(member.id, startRound)')
+    if (startLine === -1 || applyLine === -1) findings.push('the spawnMember round-count anchors are gone (startRound=' + startLine + ', applied=' + applyLine + ')')
+    notes.push('F1 resume-claim pairing: README ' + readmeHits + '/' + readmeClaims.length + ' claims, behavioural anchors ' + suiteHits + '/' + suiteClaims.length + '; cited-line drift (README pins a revision): startRound now :' + startLine + ', applied now :' + applyLine)
+  }
+
+  // F2/G1 pairing (prose ↔ code): the v5 diagram must not state the PRE-G1 order, and must describe the
+  // implemented one. Going through `readRaw` gives this block the same `V5_INTEGRITY_MUTATE` seam as every
+  // other file, which is what the `--self-probe` case below relies on.
+  {
+    const diagram = readRaw('vibe-math-v5/架构图.md')
+    const FORBID = /先\s*ack[^\n]{0,8}再构造/
+    const CLAUSES = /构造[^\n]{0,12}(后|再)[^\n]{0,12}ack|ack[^\n]{0,12}(仅|只)[^\n]{0,12}(成功|送达)/
+    if (FORBID.test(diagram)) findings.push('架构图 still states the pre-G1 order: ack before the prompt is built (F2/G1)')
+    if (!CLAUSES.test(diagram)) findings.push('架构图 does not describe the implemented inbox order (construct first / ack only on success)')
+    if (!raw.includes('if (ok && prompt.pending.length) await ackPending(')) findings.push('the ack is no longer gated by a successful wake (F2/G1 anchor missing)')
+    notes.push('架构图 inbox pairing: forbid-hits=' + (FORBID.test(diagram) ? 1 : 0) + ', clauses=' + (CLAUSES.test(diagram) ? 1 : 0))
+    // The PLAN uses the explicit pre-G1 phrases only: a sentence that legitimately QUOTES or FORBIDS the old
+    // order must not trip this, so the pattern names the phrases instead of the bare `先 ack`.
+    const planDoc = readRaw('vibe-math-v5/实现方案.md')
+    const PLAN_FORBID = /先\s*ack\s*邮件[、,，]?\s*再构造|先\s*ack[、,，]?\s*后构造/
+    if (PLAN_FORBID.test(planDoc)) findings.push('实现方案.md still states the pre-G1 order: ack the mail before building the prompt (F2/G1)')
+    notes.push('实现方案.md inbox pairing: forbid-hits=' + (PLAN_FORBID.test(planDoc) ? 1 : 0))
+    // The SAME class produced five instances, the last one inside the guard checklist itself, so the scan
+    // covers every top-level docs/*.md. The phrase is NAMED (never the bare `先 ack`) so a sentence that
+    // legitimately QUOTES or FORBIDS the old order cannot trip it. NOTE (applier): `HERE` is the REPO ROOT
+    // (see `new URL(rel, HERE)` above), so the docs directory is `new URL('docs/', HERE)` — NOT `../docs/`.
+    const DOC_FORBID = /先\s*ack[^\n]{0,8}再构造/
+    let docHits = 0, docFiles = 0
+    for (const f of readdirSync(new URL('docs/', HERE)).filter((n) => n.endsWith('.md'))) {
+      docFiles += 1
+      if (DOC_FORBID.test(readRaw('docs/' + f))) { docHits += 1; findings.push('docs/' + f + ' still states the pre-G1 order: ack the mail before building the prompt (F2/G1)') }
+    }
+    // The scan covers LIVE docs only. `docs/release-notes/**` is a FROZEN published record: the pre-G1
+    // wording there is history, not a stale claim, so it must stay OUT of the scan — and the exemption is
+    // ASSERTED (a missing/empty directory fails) so a future change to recursive scanning cannot silently
+    // start scanning frozen history. Same convention as audit-readme-counts.mjs's frozen TOTAL-65 notes.
+    let frozenNotes = []
+    try { frozenNotes = readdirSync(new URL('docs/release-notes/', HERE)).filter((n) => n.endsWith('.md')) } catch (e) { frozenNotes = [] }
+    if (!frozenNotes.length) findings.push('the frozen release-notes exemption is EMPTY or missing (docs/release-notes/*.md) — re-decide the scan scope explicitly, never widen it silently')
+    notes.push('docs/** inbox pairing: files=' + docFiles + ', forbid-hits=' + docHits + '; frozen docs/release-notes/*.md exempted=' + frozenNotes.length)
+    // The plan ALSO carried the pre-G1 re-send contract ("queued 不重发/永不重发") in two sketch lines while
+    // its own M2 row already said "未 ack ⇒ 仍可重投". NAMED phrase only, scoped to the plan file, so a
+    // sentence that legitimately quotes or forbids the old wording cannot trip it.
+    const QUEUED_FORBID = /queued\s*(永不|不)重发/
+    if (QUEUED_FORBID.test(planDoc)) findings.push('实现方案.md still states the pre-G1 re-send contract (queued 不重发/永不重发): unacked mail is re-deliverable (F2/G1)')
+    notes.push('实现方案.md queued-no-resend phrasing: hits=' + (QUEUED_FORBID.test(planDoc) ? 1 : 0))
+  }
+
   // The plan's philosophy is enforced by concrete gates; assert the load-bearing ones
   // still exist so a future edit cannot quietly drop a guard.
   const GATES = [
@@ -336,24 +602,36 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
     ['assignment objections are broadcast', 'if (p.reject_assign && typeof p.reject_assign === \'object\' && params.memberMayRejectAssign)'],
     ['the academician has no extra vote weight', 'const E = voters().map((m) => m.id)'],
     ['members may reject an assignment', "memberMayRejectAssign: true,"],
-    ['the charter states the progress definition', "'  · Progress/<你>/progress.md —— **你的研究日志**（叙述体，可追加）。'"],
+    ['the charter states the progress definition', "'  · Members/<你>/Progress/progress.md —— **你的研究日志**（叙述体，可追加）。'"],
     ['the charter describes the academician as organizer', "'  【四、你的组织职责与边界（院士）】'"],
     ['the framework never assigns on its own', "agenda: '本所较长时间没有新进展。请你们自行讨论：现在最该推进的是什么？谁来做？是否需要发起验证？'"],
 
     // ── PROMPT / INTERACTION CORRECTNESS GATES ─────────────────────────────
+    // F1 (README two-case resume claim): the two code anchors the prose cites.
+    ['the founding round is computed from the stored count (never reset)', 'const startRound = (rounds.get(member.id) || 0) + 1'],
+    ['the founding round is applied only AFTER a successful start', 'rounds.set(member.id, startRound)'],
     // The 2026-09 field test shipped a framework whose every member brief named the WRONG
     // member. These gates keep the structural fixes in place, and the companion
     // prompt-v5-integrity.test.mjs asserts the TEXT those fixes produce.
-    ['the status block takes the member it describes', 'function briefBlock(member) {'],
+    ['the status block takes the member it describes', 'function briefBlock(member, roundNo) {'],
     ['an unknown identity fails loudly instead of being guessed', "throw v5err('V5_INTERNAL', 'briefBlock: a member is required"],
-    ['the status block is built from that member, never a global', 'function stateBlock(member) {\n      return briefBlock(member)\n    }'],
+    ['the status block is built from that member, never a global', 'function stateBlock(member, roundNo) {\n      return briefBlock(member, roundNo)\n    }'],
+    ['the founding prompt is told the round it is starting (never a stored 0)', 'const prompt = initialPrompt(member, initialTask, mode, startRound)'],
     ['the joiner is committed to the roster BEFORE its brief is built', "member.phase = 'active'\n      member.childId = ''\n      await putMember(member)"],
     ['the charter is frozen at hire and reused on resume', 'const persona = member.persona || memberPersona(member)'],
     ['a rebuilt session is framed as a rebuild', "const resume = mode === 'resume'"],
     ['leadership text follows the live roster', 'function academicianId() {'],
     ['no charter invents a leader when there is none', "本所当前**没有在册院士**"],
     ['the office resolves as the office, never as a guessed member', "try { if (rootOf(agent) === agent) return 'office' } catch (e) { /* fall through */ }"],
-    ['the mailbox is acked BEFORE the round prompt is built', 'if (pending.length) await ackPending(pending)\n      const base = typeof baseFn === \'function\' ? baseFn() : baseFn'],
+    ['the mailbox is acked only AFTER the wake actually succeeded', "if (ok && prompt.pending.length) await ackPending(prompt.pending)"],
+    // F2: the three REAL mailbox invariants, each at its own source anchor. `inFlightMessages` is a DSH
+    // mailbox internal that v5 does NOT have (dedupe is `delivered` + the per-round injection marks), so
+    // no gate may reference it.
+    ['the acknowledgement is gated by the wake result AND a non-empty pending list', 'const ok = await wakeMember(member, prompt.text, kind)'],
+    ['the delivered ledger is CAPPED (oldest evicted, never unbounded)', 'const capped = delivered.length > DELIVERED_CAP ? delivered.slice(delivered.length - DELIVERED_CAP) : delivered'],
+    ['a prompt prepends only messages not already prepended in this round', 'const fresh = pending.filter((p) => !seen.has(p.id))'],
+    ['the per-round injection mark is cleared with every concluded attempt', 'inboxInjected.delete(member.id)'],
+    ['a prepended inbox suppresses the base block (exactly one inbox section per prompt)', "const pending = inboxSuppressed.has(member.id) ? [] : pendingFor(member.id)"],
     ['framework feedback has its own sender (never a self-message)', "return await say('framework', { to: memberId, kind: 'notice', text: String(text) })"],
     ['an assignment is framed by its true origin', "if (m.kind === 'assign') return (m.from === 'office' ? '【所办分派】' : '【院士分派】') + m.text"],
     ['a nudge is framed as supervision, not as an assignment', "to, kind: 'nudge',"],
@@ -388,6 +666,7 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
     if (!raw.includes(needle)) findings.push('philosophy gate missing from the implementation: ' + label)
   }
   notes.push('philosophy gates checked: ' + GATES.length)
+  gateCount = GATES.length
 
   // The prompt corpus is a SHIPPED deliverable, not a build artifact: a human must be able
   // to read the exact text every member receives without decoding session logs.
@@ -411,7 +690,7 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
   // never contain a wrong-identity brief.
   try {
     const corpusPath = new URL('prompt-corpus-v5/prompt-corpus-v5.json', REPO_ROOT)
-    const c = JSON.parse(readFileSync(corpusPath, 'utf8'))
+    const c = JSON.parse(readRaw('prompt-corpus-v5/prompt-corpus-v5.json'))
     const kinds = new Set((c.prompts || []).map(p => p.kind))
     const need = ['founding', 'founding-temp', 'founding-leaderless', 'resume', 'normal', 'checkpoint',
       'verify', 'verify-debate', 'meeting', 'meeting-proposal', 'inbox-dm', 'inbox-voters', 'inbox-chat',
@@ -440,7 +719,7 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
   // it is the only prompt the MAIN agent receives, and no e2e suite loads the YAML, so a
   // human reviewer needs the corpus on disk (generated by audit-persona-surface.test.mjs).
   try {
-    const pc = JSON.parse(readFileSync(new URL('prompt-corpus-persona/persona-corpus.json', REPO_ROOT), 'utf8'))
+    const pc = JSON.parse(readRaw('prompt-corpus-persona/persona-corpus.json'))
     for (const d of ['vibe-math-v2', 'vibe-math-v3', 'vibe-math-v4', 'vibe-math-v5']) {
       if (!(pc.presets || []).some((p) => p.preset === d)) findings.push('the persona corpus is missing preset ' + d)
     }
@@ -495,6 +774,14 @@ notes.push('composition rows: ' + v5rows.length + '; non-v4 package rows: ' + v5
 }
 
 // ---- report ------------------------------------------------------------
+// `--json` is the machine-readable contract the `--self-probe` children parse (same shape as
+// audit-prompt-invariants.mjs): { gates, failed, findings, notes }. It is emitted BEFORE the human
+// output so a child run's stdout is parseable in one piece.
+if (process.argv.includes('--json')) {
+  console.log(JSON.stringify({ gates: gateCount, failed: findings.length, findings, notes }))
+  for (const f of findings) console.error('  FINDING: ' + f)
+  process.exit(findings.length ? 1 : 0)
+}
 console.log('-- V5 integrity audit --')
 for (const n of notes) console.log('  note: ' + n)
 console.log('')