dsh-vibe-math 2.5.0 → 2.6.0

package/tests/audit-persona-sensitivity.mjs

@@ -148,27 +148,80 @@ const probes = [
     name: 'v4-usage-advertises-phantom-subcommand',
     preset: 'vibe-math-v4', file: 'vibe-math-v4.js', nth: 'first', expect: 1,
     guarantee: 'the unknown-subcommand usage string must not advertise a subcommand that no branch implements',
-    // The anchor must be the CURRENT literal: Round A changed both the usage string and the
+    // The anchor must be the CURRENT literal. Round A changed both the usage string and the
     // hint to advertise `start [problem]` / `set <k=v>...` (vibe-math-v4.js:2492). While the
     // anchor kept the pre-fix literal it matched 0 times, so this probe reported SETUP-FAIL
     // every run and the F-section agreement between usage and branches had NO sensitivity
     // proof at all (the hint probe at :155 was updated, this one was not).
-    from: "usage:'configure|start [problem]|resume|pause|abort|status|report|message <to|all> <content>|meeting|members|add|remove|set <k=v>...'",
-    to: "usage:'configure|start [problem]|resume|pause|abort|status|report|message <to|all> <content>|meeting|members|add|remove|scan|set <k=v>...'",
+    // Round H (final paper) inserted `paper [...]` into both the usage and the hint; the anchor
+    // follows the new literal again, so the probe keeps proving the usage<->branches agreement.
+    from: "usage:'configure|start [problem]|resume|pause|abort|status|report|message <to|all> <content>|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|members|add|remove|set <k=v>...'",
+    to: "usage:'configure|start [problem]|resume|pause|abort|status|report|message <to|all> <content>|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|members|add|remove|scan|set <k=v>...'",
   },
   {
     name: 'v4-hint-advertises-unimplemented-subcommand',
     preset: 'vibe-math-v4', file: 'vibe-math-v4.js', nth: 'first', expect: 1,
     guarantee: 'the typing hint must not advertise a subcommand the handler does not implement',
-    from: "message <to|all> <content>|meeting|members|add|remove|set]'",
-    to: "message <to|all> <content>|meeting|add|remove|set]'",
+    // Same reason as above: the hint now carries the `paper [...]` entry between `meeting` and
+    // `members`, so the anchor names the current literal and drops an IMPLEMENTED subcommand
+    // (`members`) from the mutated hint — the suite's implemented-subset-of-hint check must fail.
+    from: "message <to|all> <content>|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|members|add|remove|set]'",
+    to: "message <to|all> <content>|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|add|remove|set]'",
   },
   {
     name: 'v5-persona-slash-list-drops-add-remove',
     preset: 'vibe-math-v5', file: 'agent.cordis.yml', nth: 'all', expect: 2,
     guarantee: "the persona's /v5 list must match the command hint (a human reading the persona must see every subcommand)",
-    from: 'meeting|hire|fire|add|remove|set)',
-    to: 'meeting|hire|fire|set)',
+    // Round H appended `|paper` to the persona's /v5 list (and to the command hint), so the
+    // anchor carries the new tail; the mutation still drops the two subcommands whose absence
+    // the Round-G persona actually had (`add|remove`), which the hint<->persona equality check
+    // must catch.
+    from: 'meeting|hire|fire|add|remove|set|paper)',
+    to: 'meeting|hire|fire|set|paper)',
+  },
+  // ── final-paper surface (Round H) ──────────────────────────────────────────
+  // These two pin the surface the final-paper feature ADDED, so a later edit cannot quietly
+  // drop it: the v4 hint must keep advertising the subcommand its handler implements, and the
+  // v5 persona must keep naming the tools the plugin actually registers.
+  {
+    name: 'v4-hint-stops-advertising-paper',
+    preset: 'vibe-math-v4', file: 'vibe-math-v4.js', nth: 'first', expect: 1,
+    guarantee: 'the /v4 hint must keep advertising the final-paper subcommand the handler implements',
+    from: "|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|members|add|remove|set]'",
+    to: "|meeting|members|add|remove|set]'",
+  },
+  {
+    name: 'v5-paper-tool-renamed-in-plugin',
+    preset: 'vibe-math-v5', file: 'vibe-math-v5.js', nth: 'first', expect: 1,
+    guarantee: 'renaming the final-paper tool must be caught (the persona still advertises vibe_v5_paper)',
+    from: "  registerTool('vibe_v5_paper',",
+    to: "  registerTool('vibe_v5_paperX',",
+  },
+  // ── the persona↔hint list itself (the gap that hid `paper` in v2/v3/v4) ────────────────────
+  // The surface suite used to SKIP this comparison whenever the persona list contained `...`
+  // (v3's trailing ellipsis) or an argument placeholder with `...` (v4's `set <k=v>...`), and v2
+  // never enumerated at all. Each probe below drops ONE implemented subcommand from the persona
+  // list and requires the suite to go RED, so a future omission cannot hide behind an ellipsis.
+  {
+    name: 'v2-persona-slash-list-drops-paper',
+    preset: 'vibe-math-v2', file: 'agent.cordis.yml', nth: 'all', expect: 2,
+    guarantee: 'the v2 persona must name EVERY implemented /vibe subcommand explicitly (it never enumerated before)',
+    from: '|decisions|agents|paper [lang=zh|en] [format=both|md|tex] [force]).',
+    to: '|decisions|agents).',
+  },
+  {
+    name: 'v3-persona-slash-list-drops-paper',
+    preset: 'vibe-math-v3', file: 'agent.cordis.yml', nth: 'all', expect: 2,
+    guarantee: "the v3 persona's /vibe list must name every implemented subcommand (the old trailing '...' skipped the check)",
+    from: '|decisions|agents|paper [lang=zh|en] [format=both|md|tex] [force]).',
+    to: '|decisions|agents).',
+  },
+  {
+    name: 'v4-persona-slash-list-drops-paper',
+    preset: 'vibe-math-v4', file: 'agent.cordis.yml', nth: 'all', expect: 2,
+    guarantee: "a dropped v4 persona entry must be caught even though the list contains `set <k=v>...` (the old ellipsis escape)",
+    from: '|meeting|paper [lang=zh|en] [format=md|tex|both] [editor=office|resident:<id>] [force]|members',
+    to: '|meeting|members',
   },
 ]
 

package/tests/audit-persona-surface.test.mjs

@@ -170,9 +170,11 @@ const PRESETS = [
     dir: 'vibe-math-v5',
     js: 'vibe-math-v5.js',
     prefix: 'vibe_v5_',
-    tools: 35,
+    tools: 37,
     // member- and academician-facing tools; the office (main agent) holds only the
-    // institute-level controls plus the hiring authority.
+    // institute-level controls plus the hiring authority. The final-paper pair is an
+    // OFFICE control and IS named in the persona (vibe_v5_paper / vibe_v5_finalize_paper),
+    // so it must not appear in this snapshot.
     undocumented: [
       'vibe_v5_wait', 'vibe_v5_record_progress', 'vibe_v5_record_proposition',
       'vibe_v5_record_method', 'vibe_v5_record_subproblem', 'vibe_v5_read_library',
@@ -258,12 +260,19 @@ for (const P of PRESETS) {
   if (!hintOpen) eq(implemented.filter((c) => !hinted.includes(c)).sort(), [], `${P.dir}: implemented slash subcommand(s) are missing from the hint`)
   eq(used.sort(), implemented, `${P.dir}: the usage string and the implemented slash subcommands disagree`)
   const personaEnum = /slash command mirrors[^(]*\(([\s\S]*?)\)/.exec(blocks.prefix.join('\n') + '\n' + blocks.text.join('\n'))
-  if (personaEnum && personaEnum[1].includes('|')) {
+  // An ellipsis must NOT switch this check off (that is how a dropped `paper` entry survived here),
+  // and `...` inside an argument placeholder (`set <k=v>...`) is not an abbreviation at all. So the
+  // persona is compared against the BRANCHES (the ground truth), and every implemented subcommand
+  // must be named EXPLICITLY. Only a HINT that is itself abbreviated makes the reverse direction
+  // (persona ⊆ hint) unprovable, so only that direction is relaxed then.
+  if (!personaEnum) {
+    ok(false, `${P.dir}: the persona never enumerates its /vN subcommands (list them so a human can see every one)`)
+  } else {
     const personaCmds = cmdsFrom(personaEnum[1].replace(/^\s*\/[a-z0-9]+\s+/, ''))
-    if (personaCmds.length) {
-      eq(personaCmds.filter((c) => !hinted.includes(c)).sort(), [], `${P.dir}: the persona enumerates slash subcommand(s) the hint does not know`)
-      if (!/\.\.\./.test(personaEnum[1]) && !hintOpen) eq(personaCmds.slice().sort(), hinted.slice().sort(), `${P.dir}: the persona's /vN list and the command hint disagree`)
-    }
+    ok(personaCmds.length > 0, `${P.dir}: no slash subcommand could be parsed from the persona's /vN list`)
+    eq(personaCmds.filter((c) => !implemented.includes(c)).sort(), [], `${P.dir}: the persona enumerates slash subcommand(s) the handler does not implement`)
+    eq(implemented.filter((c) => !personaCmds.includes(c)).sort(), [], `${P.dir}: the persona's /vN list omits subcommand(s) the handler implements — list them explicitly, do not hide behind "..."`)
+    if (!hintOpen) eq(personaCmds.slice().sort(), implemented.slice().sort(), `${P.dir}: the persona's /vN list and the command hint disagree`)
   }
 
   // ---- E. the Lean feature's prompt surface ----------------------------