@bongos/core 1.19.39 → 1.19.41

package/.bongos-core.json

@@ -2,22 +2,22 @@
   "artifact": "bongos-core",
   "manifest_schema": 1,
   "generator": "scripts/gds/package-core.js",
-  "core_version": "1.19.39",
-  "core_contract": "1.19.39",
-  "source_commit": "d61c48b44f6f117383794f20d49b5bd7b83d35d7",
+  "core_version": "1.19.41",
+  "core_contract": "1.19.41",
+  "source_commit": "ef65ee2c6121ae010e5ad2c1ac67b349cf9b9161",
   "source_ref": "HEAD",
-  "built_at": "2026-08-11T16:25:51.076Z",
+  "built_at": "2026-08-11T16:42:56.706Z",
   "redaction": {
     "model": "docs-redacted+functional-verbatim",
-    "docs_redacted": 288,
+    "docs_redacted": 289,
     "agent_docs_stubbed": 21,
-    "functional_verbatim": 1212,
+    "functional_verbatim": 1213,
     "rules": 3,
     "gate_literals": 3,
     "gate": "passed"
   },
-  "file_count": 1521,
-  "tree_sha256": "14c889242599811c768effd9bc7ef844e4a6ee4ef0a041c74ff1629fda6f8907",
+  "file_count": 1523,
+  "tree_sha256": "8bdaf758774aaff55446f0ca5ac714795f728937d893a1085e21cf14db853915",
   "files": [
     {
       "path": ".claude/skills/blocker-review/SKILL.md",
@@ -154,6 +154,11 @@
       "mode": "0000644",
       "sha256": "e789b9157c9817bac354eaeed8152b53f3013f677934b062bca1c8e07f1f2dc8"
     },
+    {
+      "path": ".claude/skills/grade-audit/SKILL.md",
+      "mode": "0000644",
+      "sha256": "2b4d33396291f58f1c0d7c01d173b64d587b2a0a483b17328e57eb602a8df9fe"
+    },
     {
       "path": ".claude/skills/grade-recover/SKILL.md",
       "mode": "0000644",
@@ -1362,7 +1367,7 @@
     {
       "path": "docs/file-map.md",
       "mode": "0000644",
-      "sha256": "f51d3978f50ec9b2a34dc12b8fa3b9e8874543e5ab3503f72a09d7d2b06c565b"
+      "sha256": "15801d1c51e7e46ccf4e52dad5d3d34e4adc96b72eed60e3ef2c9ccccf94d1da"
     },
     {
       "path": "docs/handoff-template.md",
@@ -3372,12 +3377,12 @@
     {
       "path": "modules/lifecycle/db.js",
       "mode": "0000644",
-      "sha256": "c7dfc49df888857c8975077a069e8730d16926d8e92776f09489f69a40068f31"
+      "sha256": "57c88852ff0aa9f0b3673fa62b35f352d92cc9070e7872137003e395c42eb026"
     },
     {
       "path": "modules/lifecycle/done-when.js",
       "mode": "0000644",
-      "sha256": "803ce237622eaba321c6fd87db1645b8ba474af3d3486643aa9d52a6b2bd27d1"
+      "sha256": "4e6f6af1cdaf026466155c1b0656a0e89872aa2bb264f36c50997ec4f069e808"
     },
     {
       "path": "modules/lifecycle/github-push.js",
@@ -4022,12 +4027,12 @@
     {
       "path": "package-lock.json",
       "mode": "0000644",
-      "sha256": "12291ed6b1927973c8d8c42c5384d7ba2368779b9a39c566336253069214cfb2"
+      "sha256": "3d5765b40ed913e05bea697ecb878e98fa799e667a5352d4d8f37236a0b671b4"
     },
     {
       "path": "package.json",
       "mode": "0000644",
-      "sha256": "18139931183bcb89c64704ae5167ab99d1e8ae9841728870b277959af8523069"
+      "sha256": "552a8ef2e0d6b1a69377cbf67bdc0b49574312a68b0ac4162bc606db5870afff"
     },
     {
       "path": "public-docs/index.html",
@@ -5647,7 +5652,7 @@
     {
       "path": "src/module-api.js",
       "mode": "0000644",
-      "sha256": "5f51d5f403c88eedf838dab85e4a9ddfa8ff575ec07b5f1189e12999001a335a"
+      "sha256": "f1c39c1f5862092c57f1c5f5232dfce0a9ac41376838899c92c286df3178936f"
     },
     {
       "path": "src/module-loader/loader.js",
@@ -6167,7 +6172,7 @@
     {
       "path": "tests/criterion_progress.mjs",
       "mode": "0000644",
-      "sha256": "3ba6ac8556e9f5f0cd57f724cc136d024d64f684ff1d892bf6825e1b8ec0822d"
+      "sha256": "e5b5ccf87a34658b3b25e4289caf702fbc44d77bacfec5bd823d5f6666ff213b"
     },
     {
       "path": "tests/criterion_proposal.mjs",
@@ -6437,7 +6442,7 @@
     {
       "path": "tests/goal_achievement.mjs",
       "mode": "0000644",
-      "sha256": "0838db3cbd5922e47777a25828b2a1b662284a251c5b8309610da634d448c90a"
+      "sha256": "a6ddb674d4ce1f6da9eb7b653ef92b9f0fcc85387e70bdf5358aff84ef518824"
     },
     {
       "path": "tests/goal_advisory.mjs",
@@ -6477,7 +6482,7 @@
     {
       "path": "tests/goal_routes.mjs",
       "mode": "0000644",
-      "sha256": "cfd76bf4bd4ed22702861833966cd3eb02dee9815d11d0429cb2773ae17a48ad"
+      "sha256": "4fd3ec8ef855b0131e17931908e9aebf16238d45febdf0b8306e2908152e2655"
     },
     {
       "path": "tests/goal_scope_check.mjs",
@@ -7464,6 +7469,11 @@
       "mode": "0000644",
       "sha256": "0eeccae9320f419667cf9ede3c13671a5b03571f7f2f17067bff1857234ddf91"
     },
+    {
+      "path": "tests/skill_grade_audit.mjs",
+      "mode": "0000644",
+      "sha256": "75ad6cd89f6e84e88764a99350e093df31a525eff392e3f84ce99fa32cb952c4"
+    },
     {
       "path": "tests/skill_grade_recover.mjs",
       "mode": "0000644",

package/.claude/skills/grade-audit/SKILL.md

@@ -0,0 +1,77 @@
+---
+name: grade-audit
+description: Manual-cadence accountability sweep over recent grades — the override ledger (what shipped despite a failing grade, and was it accounted for), dropped blocker/major findings on passing grades, a false-pass spot-check for runtime-behavior blindness, and an outage roll-up. Triggers when the user says "/grade-audit", "audit recent grades", "run the grade audit", "what shipped past the grader", or "weekly grader accountability sweep". Describes and queues follow-up work ONLY — it NEVER confirms a task, flips a status, or re-grades anything (ADR 0162; verdicts are advisory, never self-executing). Manually invoked; do not depend on scheduled routines.
+---
+
+You are auditing the last week of grades for accountability gaps: work that shipped past a failing grade unaccounted for, real findings that rode passing grades into production and evaporated, and passes the panel structurally could not have judged. You **describe and queue**. You never confirm, flip, re-grade, or add a gate — every follow-up lands as a queued idea or a handoff to another skill.
+
+## The sweep window
+
+```bash
+node scripts/gds/api.js GET "/api/gds/public/recent-shipped?days=7"
+```
+
+For each shipped task in the window, read its persisted grade:
+
+```bash
+node scripts/gds/api.js GET "/api/gds/tasks/<id>?include=grade"
+```
+
+Bucket by grade shape: `passed`, failed, `signals.panel_outcome='unavailable'` (outage — no quality signal existed), or no grade row at all.
+
+## Leg 1 — the override ledger
+
+For every shipped task whose grade **failed** (or is absent), the land needed an override. Cross-reference the ledger (Archon — `override_request.decide`):
+
+```bash
+node scripts/gds/api.js GET "/api/gds/override-requests?status=approved"
+```
+
+- Split **infra vs real** first: a fail that is actually `panel_outcome='unavailable'` (or the errored-worker zero-issue shape) is an outage artifact, not an overridden quality verdict — route those to the Leg 4 roll-up.
+- A shipped task with a real failing grade and **no approved override-request row** landed via the raw confirm — the un-audited path. Name it in the report and print its unaddressed `issues[]` verbatim. That set is the ledger's debt.
+- For each such task, queue the debt (see the queueing rule below) — do not chase the lander, do not un-ship anything.
+
+## Leg 2 — dropped findings on passing grades
+
+A pass with `blocker`/`major` entries in `issues[]` shipped real findings that no process ever picks up (the evaluation counted 274 findings on passing ships with zero conversion paths). For each:
+
+- **Worker-attribution filter:** findings attributed to the advisory **Narc** are the known noise class — triage them by hand (read the finding, decide), never auto-trust them into the queue. Findings from the gating workers (Quality, Hacker, Efficiency) or the deterministic pre-passes are higher-confidence — queue them unless plainly stale.
+- **Queueing rule (idempotent):** first read the open inbox (`node scripts/gds/api.js GET /api/gds/inbox`) and skip any finding already queued — the marker is the deterministic title prefix. Then:
+
+```bash
+node scripts/gds/capture.js "[grade-audit] task <id>: <finding summary>" --kind bug
+```
+
+One idea per surviving finding, titled exactly `[grade-audit] task <id>: …` so a re-run of this audit files nothing twice.
+
+## Leg 3 — false-pass spot-check (runtime-behavior blindness)
+
+The re-review found 2/8 material misses, both the same root cause: **the panel grades diff text and never reasons about runtime** — a CI workflow whose default shallow checkout broke on first run scored functional_fidelity 10/10, and an SSRF guard bypassable via HTTP redirect passed at 9.5. For the riskiest ships in the window — security surfaces, CI/workflow files, the largest diffs, and 9.5s with suspiciously few findings — re-read the merged diff asking one question: *what does this code do at runtime that the diff text does not show?* (defaults it inherits, redirects it follows, first-run behavior, tool/CI defaults). A suspicion is queued via the Leg 2 rule, prefixed the same way — never a re-grade, never a flipped verdict.
+
+## Leg 4 — outage roll-up
+
+Count the window's `panel_outcome='unavailable'` rows. One-off outages just get the `--regrade` pointer in the report; a streak (2+ consecutive, or clustered on one builder) is a health problem — hand off to **/grader-health**, which owns outage diagnosis. Do not diagnose here.
+
+## Report shape
+
+End with the ledger, most severe first:
+
+1. **Un-accounted overrides** — task, findings printed verbatim, queued-idea ref. A task that landed with no approved override-request row is the headline.
+2. **Queued findings** — what was filed this run, what was skipped as already-queued, what was held back by the Narc filter (and why).
+3. **Spot-check suspicions** — task, the runtime question the panel couldn't answer, queued-idea ref.
+4. **Outage roll-up** — count + the /grader-health handoff if a streak.
+
+A clean week is four one-liners.
+
+## Constraints
+
+- **Never confirm, never flip, never re-grade (ADR 0162).** Verdicts are advisory and never self-executing here — this skill's entire output is a report plus queued `idea_inbox` rows. If a leg tempts you to "just fix it", the fix is a queued idea or a claimed task, not an action inside the audit.
+- **Manual cadence.** Invoke by hand (weekly is the intended rhythm). Do not wire it to cron or depend on `.claude/scheduled-tasks/`.
+- **Idempotent re-runs.** The `[grade-audit] task <id>:` title marker + the inbox pre-check are what make running it twice harmless — keep both.
+- **Archon surfaces degrade gracefully.** Without `override_request.decide`, Leg 1 can only report "ledger unreadable at this rank" — say so rather than skipping silently.
+
+## Files this skill touches
+
+- Reads: `GET /api/gds/public/recent-shipped`, `GET /api/gds/tasks/:id?include=grade`, `GET /api/gds/override-requests?status=approved` (Archon), `GET /api/gds/inbox`.
+- Writes: `idea_inbox` rows via `node scripts/gds/capture.js` (queued follow-ups only).
+- Never calls: any confirm, status, grade, or regrade endpoint.

package/docs/file-map.md

@@ -350,6 +350,7 @@ tests/
 ├── fix-task/SKILL.md              ← /fix-task N — claim, fix (test-first), verify, and ship a single GDS task by id end to end; defers to /builder-claim + /builder-ship for their own mechanics
 ├── goal-create/SKILL.md           ← /goal-create — plan and create ONE goal (the module-scoped workspace between version and criteria, ADR 0086) in one session: <redacted>, scope wall, 2-4 criteria, seed tasks; also the auto-pickup when a task is created with no goal to live in (goal_advisory); Metic+ (Archon for a protected-module scope)
 ├── goal-review/SKILL.md           ← /goal-review — walk the criterion review queue: each auto-flagged 'met — pending review' criterion gets confirm / reject / defer; confirming a goal's last criterion auto-achieves it (ADR 0086 §6); Metic+
+├── grade-audit/SKILL.md           ← /grade-audit — manual-cadence accountability sweep: override ledger (shipped-despite-fail, infra vs real, un-requested confirms named), dropped blocker/major findings on passes queued via capture.js with the advisory-Narc filter, runtime-blindness false-pass spot-check, outage roll-up → /grader-health; describes and queues only, never confirms (ADR 0162, task 1002657)
 ├── grade-recover/SKILL.md         ← /grade-recover N — the 'my grade failed, now what' loop: diagnose outage vs fabricated fail vs genuine fail from the persisted grade, then run the right recovery (--regrade or the auditable override-request); adds no ship-path gate (ADR 0162, task 1002655)
 ├── grader-health/SKILL.md         ← /grader-health — read-only grader health check: outage runs (n_unavailable streaks, outage-vs-rejection confirmation), pass-rate drift vs baseline, per-grade cost anomalies, stale model pins; emits exact --regrade remediation lines; optional --probe live calibration (Metic+, paid) (task 1002656)
 ├── idea-triage/SKILL.md           ← /idea-triage — daily walk through the GDS idea_inbox (promote / discard / merge); Metic+

package/modules/lifecycle/db.js

@@ -893,7 +893,10 @@ async function listCriteriaForGoal(goalId, deps = {}) {
      counts AS (
        SELECT tc.criterion_id,
               count(*)::int                                        AS linked_tasks,
-              count(*) FILTER (WHERE t.status <> 'shipped')::int    AS unshipped_tasks
+              -- task 1002651 (migration 017's family): abandoned is a terminal
+              -- state (deliberately cut), not pending work — it must not hold a
+              -- criterion's "met — pending review" flip hostage forever.
+              count(*) FILTER (WHERE t.status NOT IN ('shipped', 'abandoned'))::int AS unshipped_tasks
          FROM task_criteria tc
          JOIN tasks t ON t.id = tc.task_id
         WHERE tc.criterion_id IN (SELECT id FROM crit)

package/modules/lifecycle/done-when.js

@@ -126,7 +126,7 @@ async function listPendingReviewCriteria({ versionId = null } = {}, deps = {}) {
          JOIN tasks t          ON t.id = tc.task_id
         WHERE c.satisfied = false
         GROUP BY c.id
-       HAVING count(*) FILTER (WHERE t.status <> 'shipped') = 0
+       HAVING count(*) FILTER (WHERE t.status NOT IN ('shipped', 'abandoned')) = 0
      ),
      goal_unsat AS (
        SELECT goal_id, count(*)::int AS unsatisfied
@@ -269,7 +269,9 @@ function buildCriterionProgress(versionId, rows, unattributedCount = 0, goalRows
     const counts = { total: c.tasks.length };
     for (const t of c.tasks) counts[t.status] = (counts[t.status] || 0) + 1;
     const shipped = counts.shipped || 0;
-    const remaining = c.tasks.filter((t) => t.status !== 'shipped');
+    // task 1002651: abandoned is terminal (cut work, migration 017's family) —
+    // it is not "remaining" and must not block the pending-review flip.
+    const remaining = c.tasks.filter((t) => t.status !== 'shipped' && t.status !== 'abandoned');
     // Auto-flag (ADR 0086 §6 / BV1.R63): a criterion is "met — pending review"
     // when it has ≥1 linked task and every one has shipped, but it has NOT yet
     // been confirmed (satisfied=false). Purely DERIVED — no stored column, no

package/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@bongos/core",
-  "version": "1.19.39",
+  "version": "1.19.41",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@bongos/core",
-      "version": "1.19.39",
+      "version": "1.19.41",
       "license": "AGPL-3.0-or-later",
       "dependencies": {
         "express": "^4.21.2",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bongos/core",
-  "version": "1.19.39",
+  "version": "1.19.41",
   "description": "Cloud Bongos — the AI-first build platform core (GDS + platform surfaces + module system), installed as a versioned dependency (ADR 0108).",
   "license": "AGPL-3.0-or-later",
   "main": "src/platform-server.js",

package/src/module-api.js

@@ -505,8 +505,12 @@ const { validateOrRespond, LIMITS, parseId, asyncHandler, corsPublicGet, parsePa
 //         landed since 1.19.37 with no explicit bump. run 31511537558. (task 1002620)
 //   1.19.39 — CI auto-patch (publish-on-merge, ADR 0161): carrier for merges
 //         landed since 1.19.38 with no explicit bump. run 31512284325. (task 1002620)
+//   1.19.40 — CI auto-patch (publish-on-merge, ADR 0161): carrier for merges
+//         landed since 1.19.39 with no explicit bump. run 31513136724. (task 1002620)
+//   1.19.41 — CI auto-patch (publish-on-merge, ADR 0161): carrier for merges
+//         landed since 1.19.40 with no explicit bump. run 31513793491. (task 1002620)
 // ---------------------------------------------------------------------------
-const CORE_VERSION = '1.19.39'; // CI auto-patch carrier (ADR 0161); see changelog above
+const CORE_VERSION = '1.19.41'; // CI auto-patch carrier (ADR 0161); see changelog above
 
 // A namespaced logger so a module's log lines are attributable + consistent.
 // Usage: const log = api.logger('dev-box'); log.info('mounted');

package/tests/criterion_progress.mjs

@@ -247,6 +247,40 @@ test('isPendingReview: agrees with the rollup it was extracted from', () => {
   assert.equal(b.pending_review, false);
 });
 
+// --- task 1002651 (idea 1000635): abandoned tasks are terminal, not "remaining" ---
+
+test('1002651 one shipped + one abandoned → remaining=0, criterion flips to pending review', () => {
+  const rows = [
+    row(1, 20, 'c-ab', false, 300, 'landed', 'shipped'),
+    row(1, 20, 'c-ab', false, 301, 'cut', 'abandoned'),
+  ];
+  const c = buildCriterionProgress('V', rows, 0).criteria[0];
+  assert.equal(c.counts.remaining, 0, 'the abandoned task no longer counts as unshipped');
+  assert.deepEqual(c.remaining, [], 'the remaining list omits deliberately-cut work');
+  assert.equal(c.counts.abandoned, 1, 'the per-status bucket still reports it honestly');
+  assert.equal(c.pending_review, true, 'the criterion surfaces for human confirm instead of hanging forever');
+});
+
+test('1002651 abandoned does not hide genuinely unshipped work', () => {
+  const rows = [
+    row(1, 21, 'c-mix', false, 310, 'landed', 'shipped'),
+    row(1, 21, 'c-mix', false, 311, 'cut', 'abandoned'),
+    row(1, 21, 'c-mix', false, 312, 'in-flight', 'ready'),
+  ];
+  const c = buildCriterionProgress('V', rows, 0).criteria[0];
+  assert.equal(c.counts.remaining, 1);
+  assert.deepEqual(c.remaining.map((t) => t.id), [312], 'only the live task remains');
+  assert.equal(c.pending_review, false, 'live work still blocks the flip');
+});
+
+test('1002651 an all-abandoned criterion surfaces for review rather than auto-satisfying silently', () => {
+  const rows = [row(1, 22, 'c-cut', false, 320, 'cut', 'abandoned')];
+  const c = buildCriterionProgress('V', rows, 0).criteria[0];
+  // linked>0 + unshipped=0 → pending review: a human confirms or rejects what
+  // "met" means for a criterion whose entire linked work was cut.
+  assert.equal(c.pending_review, true);
+});
+
 console.log('');
 console.log(`criterionProgress aggregation tests: ${passed} passed, ${failed} failed`);
 if (failed > 0) process.exit(1);

package/tests/goal_achievement.mjs

@@ -138,7 +138,9 @@ await test('listPendingReviewCriteria: SQL encodes the auto-flag (all shipped) +
   const sql = String(find(pool.queries, /WITH pending AS/).sql).replace(/\s+/g, ' ');
   // The derivation: satisfied=false AND every linked task shipped (zero non-shipped).
   assert.match(sql, /c\.satisfied = false/, 'only unconfirmed criteria are pending');
-  assert.match(sql, /HAVING count\(\*\) FILTER \(WHERE t\.status <> 'shipped'\) = 0/, 'all linked tasks must be shipped');
+  // task 1002651: abandoned is terminal (cut work) — it must not hold the queue
+  // entry hostage, so the filter counts only live-and-unshipped tasks.
+  assert.match(sql, /HAVING count\(\*\) FILTER \(WHERE t\.status NOT IN \('shipped', 'abandoned'\)\) = 0/, 'all LIVE linked tasks must be shipped');
   assert.match(sql, /is_last_in_goal/, 'surfaces whether confirming completes the goal');
 });
 

package/tests/goal_routes.mjs

@@ -173,6 +173,17 @@ await test('listCriteriaForGoal: pending_review false once the criterion is conf
   assert.equal(out.pending_review, false);
 });
 
+await test('listCriteriaForGoal: the unshipped filter excludes abandoned (task 1002651)', async () => {
+  const pool = critPool([critRow()]);
+  await listCriteriaForGoal(3, { pool });
+  const sel = find(pool.queries, CRIT_SQL_RE);
+  assert.match(
+    sel.sql,
+    /FILTER \(WHERE t\.status NOT IN \('shipped', 'abandoned'\)\)/,
+    'an abandoned linked task must not hold the pending-review flip hostage (migration 017 family)'
+  );
+});
+
 await test('listCriteriaForGoal: pending_review false with NO linked tasks (vacuous truth guard)', async () => {
   const pool = critPool([critRow({ satisfied: false, linked_tasks: 0, unshipped_tasks: 0 })]);
   const [out] = await listCriteriaForGoal(3, { pool });

package/tests/skill_grade_audit.mjs

@@ -0,0 +1,50 @@
+// tests/skill_grade_audit.mjs — the /grade-audit skill exists and covers the
+// accountability sweep from the grader-hardening evaluation §6 (task 1002657):
+// the override ledger, dropped blocker/major findings on passing grades (with
+// the advisory-Narc attribution filter), the false-pass spot-check, and the
+// outage roll-up handoff — describing and queueing ONLY (ADR 0162/0158: it
+// never confirms a task or flips status).
+import assert from 'node:assert/strict';
+import { test } from 'node:test';
+import { readFileSync, existsSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import path from 'node:path';
+
+const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
+const SKILL = path.join(ROOT, '.claude', 'skills', 'grade-audit', 'SKILL.md');
+
+// .claude/skills is INSTANCE-materialized; a neutral core checkout carries none
+// (task 1002470) — skip there, same guard shape as file_map.mjs.
+const skillsDirExists = existsSync(path.join(ROOT, '.claude', 'skills'));
+
+test('/grade-audit exists, covers the sweep legs, and stays describe-and-queue only', (t) => {
+  if (!skillsDirExists) return t.skip('.claude/skills not materialized in this checkout');
+  assert.ok(existsSync(SKILL), '.claude/skills/grade-audit/SKILL.md must exist (task 1002657)');
+  const md = readFileSync(SKILL, 'utf8');
+
+  // Data surfaces.
+  assert.match(md, /\/api\/gds\/public\/recent-shipped/, 'sweeps the recent-shipped window');
+  assert.match(md, /include=grade/, 'reads each task\'s persisted grade');
+  assert.match(md, /\/api\/gds\/override-requests/, 'cross-references the override ledger');
+
+  // Leg 1 — override ledger: shipped-with-failing-grade, infra vs real split.
+  assert.match(md, /panel_outcome='unavailable'/, 'splits infra outage from a real fail');
+  assert.match(md, /no approved (override-)?request/i, 'names tasks that landed with no request row');
+
+  // Leg 2 — dropped findings with the worker-attribution filter.
+  assert.match(md, /blocker|major/, 'targets blocker/major findings on passing grades');
+  assert.match(md, /Narc/i, 'carries the advisory-Narc attribution filter');
+  assert.match(md, /capture\.js/, 'queues findings via capture.js');
+  assert.match(md, /\[grade-audit\]/, 'uses the deterministic title marker for idempotency');
+
+  // Leg 3 — false-pass spot-check (the 2/8 runtime-blindness classes).
+  assert.match(md, /runtime/i, 'spot-checks for runtime-behavior blindness');
+
+  // Leg 4 — outage roll-up handoff.
+  assert.match(md, /\/grader-health/, 'hands outage windows to /grader-health');
+
+  // The hard constraint.
+  assert.match(md, /never confirm/i, 'states it never confirms');
+  assert.match(md, /ADR 0162/, 'anchors the no-new-gates constraint');
+  assert.ok(!/curl[^\n]*\/confirm/.test(md), 'carries no ready-to-paste confirm curl');
+});