@bongos/core 1.19.801 → 1.19.803

package/scripts/gds/artifact-staleness.js

@@ -0,0 +1,267 @@
+#!/usr/bin/env node
+// scripts/gds/artifact-staleness.js — are the gitignored generated artifacts on
+// THIS machine actually current with the commit it has checked out? (task 1001930,
+// ADR 0110 pt.5)
+//
+// WHY THIS EXISTS. ADR 0110 moved `docs/repo-map.md` and `docs/session-log-index.md`
+// off the git graph: they are 100%-generated, two parallel branches collided on them
+// constantly, so they became untracked files regenerated at deploy. That trade was
+// deliberate, and the ADR names the debt it took on in its own Consequences:
+//
+//   "A generator that breaks at deploy leaves a stale/absent file with no blocking
+//    backstop; the LOUD ::ERROR:: line is the alarm (mandatory, not optional)."
+//
+// That alarm is a line in the droplet's journald. Nobody reads journald. The per-PR
+// freshness gate cannot help either — the files are gitignored, so they never appear
+// in a diff for CI to check (`--check-whole`'s own comment says it plainly: "used by
+// nothing hard; deploy owns it"). So if the deploy regen step silently stops running,
+// every agent keeps reading a symbol map of a codebase that has moved on, and nothing
+// anywhere says so. This check is the backstop that posture depends on.
+//
+// WHAT IT COMPARES. For each registered artifact, its own generator's `--check-whole`
+// answers "would regenerating this file change it?" — run against the checkout that is
+// actually serving. A difference means the deploy generator did not run, or ran and
+// broke. Asking the generator rather than reimplementing the comparison is the point:
+// there is exactly one definition of "fresh" per artifact and it lives with the thing
+// that writes it, so this check cannot drift away from what deploy produces.
+//
+// ABSENT AND STALE ARE REPORTED SEPARATELY, and the split is operational, not cosmetic:
+// absent means the regen step never ran here at all (a deploy line was dropped, or a
+// fresh clone was never generated into), stale means it ran once and has not since.
+// Those send someone to two different places, so the report must not blur them.
+//
+// ── WHERE THIS MAY RUN — read before scheduling it anywhere new ───────────────
+//
+// ONLY where the artifacts are SUPPOSED to already exist: a deployed instance, or a
+// dev box after its source-fetch regen. NEVER in CI. A fresh `git clone` legitimately
+// has no `docs/repo-map.md` — it is gitignored, so it is not in the repository at all
+// — and this check would report it absent on every single run. A monitor that is red
+// every run teaches its reader to ignore it, which is the exact failure ADR 0110's
+// warn-only posture already has. That is why this ships as a scheduled routine on the
+// instance and NOT as a `.github/workflows/` job, unlike its sibling publish-watch.js
+// (which reads git and the registry, and so is true anywhere).
+//
+// THE MAIN GUARD. "Fresh" here means "matches what THIS commit generates". On a
+// deployed instance that commit is main, because deploy does `git reset --hard
+// origin/main` — so the question the caller actually wants ("does the served artifact
+// lag main's HEAD?") and the question `--check-whole` answers are the same question.
+// In a feature worktree they are NOT: the artifact would be compared against that
+// branch, and a pass or a fail would both be meaningless. So a checkout that is not at
+// origin/main is INCONCLUSIVE (exit 2), never a finding — a monitor that cannot tell
+// which question it just answered must say so rather than guess.
+'use strict';
+
+const { spawnSync } = require('child_process');
+const fs = require('fs');
+const path = require('path');
+
+const REPO_ROOT = path.resolve(__dirname, '..', '..');
+const GIT_TIMEOUT_MS = 15000;
+const CHECK_TIMEOUT_MS = 120000;
+
+// The gitignored, whole-file generated artifacts this check covers.
+//
+// MEMBERSHIP RULE — an artifact belongs here when all three hold: it is gitignored (so
+// no PR gate can see it), it is written by a deploy-time generator (so it can silently
+// stop being written), and its generator exposes `--check-whole` (so freshness has one
+// definition, owned by the writer). tests/artifact_staleness.mjs pins this list against
+// .gitignore and against the generators' real flags, so an artifact that gains a
+// `--check-whole` and is left out of here fails that test rather than going unwatched.
+//
+// DELIBERATELY NOT HERE, and each for a reason that is not "we forgot":
+//   • modules/hall-ui/public/atlas.json (gen-atlas.js, ADR 0094) — gitignored and
+//     deploy-generated, but its payload embeds `generatedAt: new Date()`, so a
+//     regenerate-and-compare differs on every run by construction. Checking it needs a
+//     comparison over `views` only, or a freshness-by-age test against the last deploy
+//     — a different mechanism, not this one. gen-atlas has no --check mode at all today.
+//   • docs/onboarding/diagrams/*.svg|png — rendering them needs a headless Chromium the
+//     droplet does not have (the reason ADR 0110 deferred the diagram half in the first
+//     place), so freshness cannot be evaluated where this runs. Their FACTS are already
+//     backstopped by the diagram-drift routine, which reads assertions.json instead.
+const ARTIFACTS = Object.freeze([
+  Object.freeze({
+    rel: 'docs/repo-map.md',
+    generator: 'gen-repo-map.js',
+    what: 'the symbol skeleton agents navigate the codebase by (ADR 0063)',
+  }),
+  Object.freeze({
+    rel: 'docs/session-log-index.md',
+    generator: 'gen-session-index.js',
+    what: 'the full session-log index the CLAUDE.md §13 snippet links out to (ADR 0062 §8)',
+  }),
+]);
+
+function git(args, run = spawnSync) {
+  const raw = gitRaw(args, run);
+  return raw === null ? null : raw.trim();
+}
+
+// Untrimmed, because `git status --porcelain` is column-significant: a line is
+// `XY<space>PATH` and an unstaged modification's X is a SPACE (" M path"). Trimming the
+// buffer eats that leading space on the FIRST line only, so a slice(3) then swallows the
+// path's first character — which is exactly how this printed "cripts/gds/…" until
+// tests/artifact_staleness.mjs caught it.
+function gitRaw(args, run = spawnSync) {
+  const r = run('git', args, { cwd: REPO_ROOT, encoding: 'utf8', timeout: GIT_TIMEOUT_MS });
+  if (!r || r.error || r.status !== 0) return null;
+  return String(r.stdout || '');
+}
+
+// The paths in a porcelain listing, status columns removed.
+function porcelainPaths(raw) {
+  return raw.split('\n').filter((l) => l.length > 3).map((l) => l.slice(3).trim()).filter(Boolean);
+}
+
+// Is this checkout a fair thing to compare the artifacts against — at origin/main AND
+// clean? Returns { ok } or { ok:false, reason }; git being unable to answer is equally
+// inconclusive.
+//
+// The CLEAN half is not belt-and-braces, and it was found by running this check against
+// its own uncommitted source file: a generator regenerates from the WORKING TREE, so one
+// uncommitted file that moves a symbol makes repo-map.md "stale" against a main it in
+// fact matches. On the droplet the tree is clean by construction (deploy does
+// `git reset --hard`), so requiring it costs a correct run nothing — while without it
+// every hand-run from a working checkout invents a finding, which is precisely how a
+// monitor teaches its reader to ignore it. Gitignored files (the artifacts themselves)
+// never appear in `--porcelain`, so they cannot trip this.
+function atMain(run = spawnSync) {
+  const head = git(['rev-parse', 'HEAD'], run);
+  const main = git(['rev-parse', 'origin/main'], run);
+  if (!head) return { ok: false, reason: 'could not read HEAD (not a git checkout?)' };
+  if (!main) return { ok: false, reason: 'could not read origin/main (no remote ref here?)' };
+  if (head !== main) {
+    const branch = git(['rev-parse', '--abbrev-ref', 'HEAD'], run) || '(detached)';
+    return { ok: false, reason: `checkout is at ${branch} ${head.slice(0, 8)}, not origin/main ${main.slice(0, 8)}` };
+  }
+  const dirty = gitRaw(['status', '--porcelain'], run);
+  if (dirty === null) return { ok: false, reason: 'could not read the working tree state' };
+  const paths = porcelainPaths(dirty);
+  if (paths.length) {
+    const shown = paths.slice(0, 3).join(', ');
+    const more = paths.length > 3 ? `, +${paths.length - 3} more` : '';
+    return { ok: false, reason: `working tree is not clean (${shown}${more}) — the generators would compare against those edits, not main` };
+  }
+  return { ok: true, sha: head };
+}
+
+// Ask ONE artifact's own generator whether it is current.
+//   fresh    exit 0 — regenerating would change nothing
+//   stale    exit 2 — it would (absent counts as stale; `present` says which)
+//   unknown  anything else — the generator itself could not answer, which is not a
+//            finding about the artifact and must never be reported as one.
+function checkArtifact(a, { root = REPO_ROOT, run = spawnSync, exists = fs.existsSync } = {}) {
+  const present = exists(path.join(root, a.rel));
+  const r = run(process.execPath, [path.join(root, 'scripts', 'gds', a.generator), '--check-whole'],
+    { cwd: root, encoding: 'utf8', timeout: CHECK_TIMEOUT_MS });
+  if (!r || r.error) {
+    return { ...a, state: 'unknown', present, detail: r && r.error ? r.error.message : 'generator did not run' };
+  }
+  if (r.status === 0) return { ...a, state: 'fresh', present };
+  if (r.status === 2) return { ...a, state: present ? 'stale' : 'absent', present };
+  const said = String(r.stderr || r.stdout || '').trim().split('\n').pop() || `exit ${r.status}`;
+  return { ...a, state: 'unknown', present, detail: said };
+}
+
+function check({ root = REPO_ROOT, run = spawnSync, exists = fs.existsSync, artifacts = ARTIFACTS, requireMain = true } = {}) {
+  if (requireMain) {
+    const m = atMain(run);
+    if (!m.ok) { const e = new Error(m.reason); e.inconclusive = true; throw e; }
+  }
+  const results = artifacts.map((a) => checkArtifact(a, { root, run, exists }));
+  return {
+    results,
+    fresh: results.filter((r) => r.state === 'fresh'),
+    stale: results.filter((r) => r.state === 'stale'),
+    absent: results.filter((r) => r.state === 'absent'),
+    unknown: results.filter((r) => r.state === 'unknown'),
+  };
+}
+
+// Plain language: this lands in Discord #ship-news, where the audience is the owner,
+// not whoever wrote the generator.
+function report(r) {
+  const bad = [...r.absent, ...r.stale];
+  if (!bad.length && !r.unknown.length) {
+    return `Generated navigation docs are current — ${r.fresh.length} artifact(s) match this commit.`;
+  }
+  const lines = [];
+  if (bad.length) {
+    lines.push(`⚠ ${bad.length} generated file(s) on this instance no longer match the code being served.`);
+    lines.push('');
+    for (const a of r.absent) lines.push(`  • ${a.rel} — MISSING entirely. Nothing has generated it here. It is ${a.what}.`);
+    for (const a of r.stale) lines.push(`  • ${a.rel} — out of date: it no longer matches what this commit would produce. It is ${a.what}.`);
+    lines.push('');
+    lines.push('These are rebuilt by the deploy step, so this means that step stopped running or broke.');
+    lines.push('Nothing blocks on them, which is why it went unreported until now — anyone reading these');
+    lines.push('files is being given an out-of-date picture. Check the deploy log for its regen lines');
+    lines.push('(they print ::ERROR:: when a generator fails).');
+  }
+  if (r.unknown.length) {
+    if (bad.length) lines.push('');
+    lines.push(`Could not check ${r.unknown.length} artifact(s):`);
+    for (const a of r.unknown) lines.push(`  • ${a.rel} — ${a.detail}`);
+  }
+  return lines.join('\n');
+}
+
+// Deliver the finding here rather than through run-routine.js's `sink: discord`.
+// run-routine treats a non-zero step as a FAILED step and aborts BEFORE it routes
+// stdout, so exiting 1 on a finding would throw that finding away. publish-watch.js
+// and artifact-audit.js carry the same shape for the same reason; their routine steps
+// all stay `sink: none` and the script performs its own side effect.
+async function deliver(text) {
+  const { postWebhook } = require(path.join(__dirname, '..', 'discord', 'lib.js'));
+  await postWebhook('ship_news', text);
+}
+
+async function main() {
+  const argv = process.argv.slice(2);
+  const json = argv.includes('--json');
+
+  let r;
+  try {
+    r = check({ requireMain: !argv.includes('--any-ref') });
+  } catch (err) {
+    // Exit 2 = COULD NOT RUN, kept distinct from 1 (ran, found something) so a
+    // scheduler can go red on the former and merely report the latter.
+    console.error(`artifact-staleness could not complete: ${err && err.message ? err.message : err}`);
+    if (err && err.inconclusive) {
+      console.error('  Freshness is only meaningful against the commit being served. Re-run on the');
+      console.error('  deployed instance, or pass --any-ref to compare against whatever is checked out.');
+    }
+    process.exit(2);
+  }
+
+  const text = report(r);
+  const found = r.stale.length + r.absent.length;
+
+  if (json) { console.log(JSON.stringify(r, null, 2)); process.exit(found ? 1 : 0); }
+
+  // A healthy run is deliberately quiet — it tells a person who ran it by hand
+  // (stderr) and posts nothing. A daily "all fine" is how a channel becomes wallpaper.
+  if (!found) {
+    console.error(text);
+    // An artifact we could not check is not a finding, but it must not be silent
+    // either: a permanently-unknown artifact is unwatched without anyone deciding so.
+    process.exit(r.unknown.length ? 2 : 0);
+  }
+
+  console.log(text);
+  if (argv.includes('--no-post')) process.exit(1);
+
+  try {
+    await deliver(text);
+    console.error('posted the finding to Discord #ship-news.');
+    process.exit(1);
+  } catch (err) {
+    // Found something and could not say it — the original defect one layer down, so
+    // it is the LOUDEST outcome: exit 3 is >= 2, which turns the scheduled run red.
+    console.error(`\nFOUND STALE ARTIFACTS AND COULD NOT DELIVER THE FINDING: ${err && err.message ? err.message : err}`);
+    console.error('The finding is above. Set DISCORD_SHIP_NEWS_WEBHOOK so this reaches #ship-news.');
+    process.exit(3);
+  }
+}
+
+if (require.main === module) main();
+
+module.exports = { ARTIFACTS, atMain, checkArtifact, check, report };

package/src/module-api.js

@@ -71,7 +71,7 @@ const { responsibilityFor, ROLE_RESPONSIBILITIES } = require('./role-responsibil
 // there. scripts/gds/bump-version.js still rewrites the literal below; it appends
 // the entry to that file. Look for a version's history there, not here.
 // ---------------------------------------------------------------------------
-const CORE_VERSION = '1.19.801'; // CI auto-patch carrier (ADR 0161); changelog: docs/module-api-changelog.md
+const CORE_VERSION = '1.19.803'; // CI auto-patch carrier (ADR 0161); changelog: docs/module-api-changelog.md
 
 // A namespaced logger so a module's log lines are attributable + consistent.
 // Usage: const log = api.logger('dev-box'); log.info('mounted');

package/tests/artifact_staleness.mjs

@@ -0,0 +1,202 @@
+// tests/artifact_staleness.mjs
+//
+// Covers scripts/gds/artifact-staleness.js (task 1001930, ADR 0110 pt.5) — the backstop
+// for the gitignored, deploy-generated navigation docs.
+//
+// ADR 0110 untracked docs/repo-map.md + docs/session-log-index.md to kill a merge-collision
+// class, and accepted a debt in writing: a deploy generator that breaks leaves a stale or
+// absent file with no blocking gate, alarming only via a ::ERROR:: line in journald. This
+// check is that backstop, so the properties worth pinning are the ones that decide whether
+// anyone will still be reading it in six months:
+//
+//   1. it distinguishes CANNOT-CHECK from FOUND-SOMETHING (a monitor that cries wolf when
+//      it is merely in the wrong place is one nobody reads — the exact ADR 0110 failure);
+//   2. absent and stale stay separate (they send a reader to two different places);
+//   3. the registry cannot silently fall behind the set of gitignored artifacts.
+//
+// `run`/`exists` are injected throughout, so nothing here touches a real generator, git
+// repo, or network. The end-to-end behaviour against the REAL generators was verified by
+// hand at ship time (fresh / absent / stale / both), which fakes cannot prove.
+//
+// Run:  node --test tests/artifact_staleness.mjs
+
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { createRequire } from 'node:module';
+import { readFileSync, readdirSync } from 'node:fs';
+
+const require = createRequire(import.meta.url);
+const { ARTIFACTS, atMain, checkArtifact, check, report } = require('../scripts/gds/artifact-staleness.js');
+
+const SHA = 'a'.repeat(40);
+const OTHER = 'b'.repeat(40);
+
+// A fake `run` for git: a table keyed by the git subcommand+args it is asked.
+function fakeGit({ head = SHA, main = SHA, porcelain = '', branch = 'main', fail = null } = {}) {
+  return (cmd, args) => {
+    const key = args.join(' ');
+    if (fail && key.includes(fail)) return { status: 1, stdout: '', stderr: 'boom' };
+    if (key === 'rev-parse HEAD') return { status: 0, stdout: `${head}\n` };
+    if (key === 'rev-parse origin/main') return { status: 0, stdout: `${main}\n` };
+    if (key === 'rev-parse --abbrev-ref HEAD') return { status: 0, stdout: `${branch}\n` };
+    if (key === 'status --porcelain') return { status: 0, stdout: porcelain };
+    return { status: 0, stdout: '' };
+  };
+}
+
+// ── the guard: when is a comparison even meaningful? ────────────────────────────
+
+test('at origin/main with a clean tree is comparable', () => {
+  assert.equal(atMain(fakeGit()).ok, true);
+});
+
+test('a feature branch is INCONCLUSIVE — the artifact would be judged against the wrong code', () => {
+  const r = atMain(fakeGit({ head: OTHER, branch: 'claude/some-task' }));
+  assert.equal(r.ok, false);
+  assert.match(r.reason, /claude\/some-task/);
+  assert.match(r.reason, /not origin\/main/);
+});
+
+test('a DIRTY tree is inconclusive — a generator regenerates from the working tree, not from main', () => {
+  // The real bug this caught: running the check against its own uncommitted source file
+  // reported repo-map.md stale when it in fact matched main.
+  const r = atMain(fakeGit({ porcelain: ' M scripts/gds/artifact-staleness.js\n' }));
+  assert.equal(r.ok, false);
+  assert.match(r.reason, /not clean/);
+  assert.match(r.reason, /scripts\/gds\/artifact-staleness\.js/);
+});
+
+test('every porcelain status column is parsed whole — an unstaged " M" must not lose its first letter', () => {
+  // The SECOND bug this file caught: the reader trimmed the buffer, and an unstaged
+  // modification's leading space is significant, so the first line printed "cripts/…".
+  // One case per column shape, first-line position included, since only the first line
+  // was ever affected.
+  for (const line of [' M scripts/gds/artifact-staleness.js', 'M  scripts/gds/artifact-staleness.js', '?? scripts/gds/artifact-staleness.js', 'A  scripts/gds/artifact-staleness.js']) {
+    const r = atMain(fakeGit({ porcelain: `${line}\n` }));
+    assert.match(r.reason, /\(scripts\/gds\/artifact-staleness\.js\)/, `mangled the path for "${line}"`);
+  }
+});
+
+test('a long dirty list is truncated with a count rather than dumped', () => {
+  const porcelain = [' M a.js', ' M b.js', ' M c.js', ' M d.js', ' M e.js'].join('\n') + '\n';
+  const r = atMain(fakeGit({ porcelain }));
+  assert.match(r.reason, /a\.js, b\.js, c\.js, \+2 more/);
+});
+
+test('git being unable to answer is inconclusive, never a pass', () => {
+  assert.equal(atMain(fakeGit({ fail: 'rev-parse origin/main' })).ok, false);
+  assert.equal(atMain(fakeGit({ fail: 'status --porcelain' })).ok, false);
+  assert.equal(atMain(fakeGit({ fail: 'rev-parse HEAD' })).ok, false);
+});
+
+test('check() REFUSES rather than reporting a finding when the checkout is not comparable', () => {
+  assert.throws(
+    () => check({ run: fakeGit({ head: OTHER }), exists: () => true }),
+    (e) => e.inconclusive === true,
+    'an inconclusive run must be distinguishable from a finding',
+  );
+});
+
+// ── one artifact's verdict ──────────────────────────────────────────────────────
+
+const A = { rel: 'docs/repo-map.md', generator: 'gen-repo-map.js', what: 'the symbol skeleton' };
+const gen = (status, extra = {}) => () => ({ status, stdout: '', stderr: '', ...extra });
+
+test('generator exit 0 → fresh', () => {
+  assert.equal(checkArtifact(A, { run: gen(0), exists: () => true }).state, 'fresh');
+});
+
+test('generator exit 2 + file present → stale; + file absent → absent', () => {
+  assert.equal(checkArtifact(A, { run: gen(2), exists: () => true }).state, 'stale');
+  assert.equal(checkArtifact(A, { run: gen(2), exists: () => false }).state, 'absent');
+});
+
+test('a generator that cannot answer is UNKNOWN, never stale — it is not a finding about the artifact', () => {
+  const crashed = checkArtifact(A, { run: gen(1, { stderr: 'SyntaxError: bad\n' }), exists: () => true });
+  assert.equal(crashed.state, 'unknown');
+  assert.match(crashed.detail, /SyntaxError/);
+  const spawnFailed = checkArtifact(A, { run: () => ({ error: new Error('ENOENT') }), exists: () => true });
+  assert.equal(spawnFailed.state, 'unknown');
+});
+
+// ── the report a human actually reads ───────────────────────────────────────────
+
+const runFor = (byGenerator) => (cmd, args) => {
+  if (cmd === 'git') return fakeGit()(cmd, args);
+  const which = String(args[0] || '').split('/').pop();
+  return { status: byGenerator[which], stdout: '', stderr: '' };
+};
+
+test('all fresh → one quiet line, and no alarm language', () => {
+  const r = check({ run: runFor({ 'gen-repo-map.js': 0, 'gen-session-index.js': 0 }), exists: () => true });
+  assert.equal(r.stale.length + r.absent.length, 0);
+  const text = report(r);
+  assert.match(text, /current/);
+  assert.ok(!text.includes('⚠'), 'a healthy run must not look like a finding');
+});
+
+test('absent and stale are reported as different things, not one blurred count', () => {
+  const r = check({
+    run: runFor({ 'gen-repo-map.js': 2, 'gen-session-index.js': 2 }),
+    exists: (p) => !String(p).endsWith('repo-map.md'), // repo-map gone, session-index present
+  });
+  assert.deepEqual(r.absent.map((a) => a.rel), ['docs/repo-map.md']);
+  assert.deepEqual(r.stale.map((a) => a.rel), ['docs/session-log-index.md']);
+  const text = report(r);
+  assert.match(text, /docs\/repo-map\.md — MISSING entirely/);
+  assert.match(text, /docs\/session-log-index\.md — out of date/);
+  assert.match(text, /deploy step/, 'the report must say where to go look');
+});
+
+test('an unchecked artifact is surfaced even when nothing is stale — unwatched must not read as healthy', () => {
+  const r = check({ run: runFor({ 'gen-repo-map.js': 0, 'gen-session-index.js': 1 }), exists: () => true });
+  assert.equal(r.stale.length + r.absent.length, 0);
+  assert.equal(r.unknown.length, 1);
+  assert.match(report(r), /Could not check 1 artifact/);
+});
+
+// ── the registry cannot fall behind reality ─────────────────────────────────────
+
+test('every registered artifact is really gitignored — a tracked file belongs to the PR gate instead', () => {
+  const ignore = readFileSync(new URL('../.gitignore', import.meta.url), 'utf8')
+    .split('\n').map((l) => l.trim());
+  for (const a of ARTIFACTS) {
+    assert.ok(ignore.includes(a.rel), `${a.rel} must be gitignored to belong in this check`);
+  }
+});
+
+test('every registered generator really exposes --check-whole — the contract this check is built on', () => {
+  for (const a of ARTIFACTS) {
+    const src = readFileSync(new URL(`../scripts/gds/${a.generator}`, import.meta.url), 'utf8');
+    assert.match(src, /--check-whole/, `${a.generator} must support --check-whole`);
+  }
+});
+
+test('the two ADR 0110 artifacts are both covered, and the known exclusions stay deliberate', () => {
+  assert.deepEqual(ARTIFACTS.map((a) => a.rel).sort(), ['docs/repo-map.md', 'docs/session-log-index.md']);
+  // atlas.json and the diagram renders are gitignored + deploy-generated too, but neither
+  // can be answered by this mechanism (a timestamped payload; a render needing Chromium).
+  // The reasons are written out at ARTIFACTS — this asserts they were a decision, so that
+  // adding either later is a deliberate act rather than a silent widening.
+  const src = readFileSync(new URL('../scripts/gds/artifact-staleness.js', import.meta.url), 'utf8');
+  assert.match(src, /atlas\.json/, 'the atlas exclusion must stay documented');
+  assert.match(src, /diagrams/, 'the diagram-render exclusion must stay documented');
+});
+
+test('this check must NOT be wired into CI — a fresh clone legitimately has none of these files', () => {
+  // The artifacts are gitignored, so a CI checkout never has them and this would be red on
+  // every single run — the "monitor nobody reads" failure ADR 0110 already has once. The
+  // mistake is an easy one to make by analogy with publish-watch.js, which IS a workflow
+  // (it reads git + the registry, so it is true anywhere). Assert no workflow invokes it.
+  const src = readFileSync(new URL('../scripts/gds/artifact-staleness.js', import.meta.url), 'utf8');
+  assert.match(src, /NEVER in CI/, 'the file must state where it may run');
+
+  let names = [];
+  try { names = readdirSync(new URL('../.github/workflows', import.meta.url)); } catch { names = []; }
+  const callers = names.filter((n) => {
+    try {
+      return readFileSync(new URL(`../.github/workflows/${n}`, import.meta.url), 'utf8').includes('artifact-staleness');
+    } catch { return false; }
+  });
+  assert.deepEqual(callers, [], `no workflow may run this check; found: ${callers.join(', ')}`);
+});

package/tests/hall_record_world.mjs

@@ -412,6 +412,25 @@ test('the figure band drops absent figures rather than drawing a dash', async ()
   assert.deepEqual(labels, ['credits', 'Estimated', 'Priority'], 'the figures it DOES carry are all drawn, in order');
 });
 
+// task 1003964 — 'Actual' is claimed-session time, and the bare label promises
+// more than that: a re-graded task's recovery session is not in the number,
+// because a re-grade has no claim to seal a session_logs row against. The
+// decision is fine; the word alone hiding it is not. Rendered, not grepped —
+// a title that never survives figureBand looks identical in the source.
+test('the Actual figure says what it counts, and only that figure carries a title', async () => {
+  const r = renderTask({ ...BASE_TASK });
+  await r.boot(); await settle();
+  const cells = [...r.html().matchAll(/<div class="rec-fig"([^>]*)>[\s\S]*?<span class="rec-fig__l">([^<]+)</g)]
+    .map((m) => ({ attrs: m[1], label: m[2] }));
+  const actual = cells.find((c) => c.label === 'Actual');
+  assert.ok(actual, 'the figure is drawn at all');
+  assert.match(actual.attrs, /title="[^"]*claimed sessions[^"]*"/, 'it names what it measures');
+  assert.match(actual.attrs, /re-grade/, 'and what it deliberately leaves out');
+  for (const c of cells.filter((x) => x.label !== 'Actual')) {
+    assert.equal(c.attrs.trim(), '', `${c.label} carries no stray title`);
+  }
+});
+
 test('under two figures there is no band at all — one figure is not a row', async () => {
   const r = renderTask({ ...BASE_TASK, credits_reward: null, est_minutes: null, actual_minutes: null });
   await r.boot(); await settle();

package/tests/regrade_actual_minutes.mjs

@@ -0,0 +1,100 @@
+// tests/regrade_actual_minutes.mjs — a re-grade does not re-measure, and the
+// panel is not fed the figure it would misread (task 1003964, goal 1000118).
+//
+// THE MECHANISM. actual_minutes is recomputed in exactly one place — resolveClaim
+// (modules/lifecycle/db-ship.js) — as the SUM of every session_logs row for the
+// task, and session_logs rows are sealed per CLAIM. A --regrade has no claim,
+// which is the defining property of that path (#678): it never reaches
+// resolveClaim, seals no row, and does not recompute. A re-graded task therefore
+// keeps the figure its FIRST, grade-failed attempt measured (observed on task
+// 1002343: actual_minutes=19 from 2026-09-04, shipped on the 2026-09-07 redo).
+//
+// THE POSTURE, and why it is a decision rather than a patch. A claimless window
+// has no defensible start time. The prior grade's timestamp puts the entire gap
+// between attempts into the sum — for task 1002343 that is ~3 days, roughly
+// 4300 minutes in place of 19, which is worse than stale and poisons the same
+// calibration the fix was meant to protect. The ship's own session digest
+// (scripts/gds/session-digest-build.js) measures a Claude Code SESSION, which
+// routinely spans several tasks and is absent outside an interactive one. An
+// opt-in --minutes flag would be omitted by default, which is the present
+// silence with a flag beside it. So: actual_minutes means CLAIMED-session time,
+// stated at the recompute, in the schema map, in the calibration reader, and on
+// the hall figure — the four places someone meets the number.
+// The hall half is pinned in tests/hall_record_world.mjs, which owns the render
+// harness that proves the title survives figureBand rather than merely existing.
+//
+// WHAT THIS FILE PINS is the half that can silently become untrue. The bug row
+// was filed believing the grader scores an est/actual ratio against the stale
+// figure; on current main it cannot, because the live panel never receives
+// actual_minutes at all. That is the whole reason the posture is safe, so it is
+// the thing that needs a tripwire: wire actual_minutes into the panel prompt and
+// this file reds, and the decision has to be made again with the risk real.
+import { strict as assert } from 'node:assert';
+import { test } from 'node:test';
+import { createRequire } from 'node:module';
+import { readFileSync } from 'node:fs';
+
+const require = createRequire(import.meta.url);
+const { buildPrompt } = require('../modules/grading/grader-prompt.js');
+
+// A task carrying an unmistakable actual_minutes — a value that cannot appear by
+// coincidence in a prompt full of ids, line counts and minute figures.
+const TASK = {
+  id: 1003964,
+  title: 'A re-graded task keeps the first attempt figure',
+  kind: 'bug',
+  discipline: 'engineer',
+  est_minutes: 60,
+  actual_minutes: 31337,
+  scoping_confidence: 4,
+  description: 'the description',
+};
+
+const BUNDLE = {
+  task: TASK,
+  valueSummary: 'a value summary',
+  diff: 'diff --git a/x.js b/x.js\n+one line\n',
+  changedFiles: ['x.js'],
+  diffStats: { files: 1, insertions: 1, deletions: 0 },
+};
+
+test('the panel prompt does not carry actual_minutes — the tripwire', () => {
+  // THE reason a stale figure is harmless today. If this ever starts failing,
+  // the panel is scoring a number that on a re-grade comes from a superseded
+  // attempt, and "documented as excluded" stops being a sufficient answer.
+  const prompt = buildPrompt(BUNDLE);
+  assert.ok(!prompt.includes('31337'), 'the stored figure never reaches the panel');
+  assert.doesNotMatch(prompt, /actual[_ ]minutes/i, 'nor the field by name');
+  // And the assertion is not vacuous: the prompt DOES carry the estimate, which
+  // is the neighbouring field and the one the panel is meant to weigh.
+  assert.match(prompt, /60 min/, 'est_minutes is what the panel is given instead');
+});
+
+test('a task with no actual_minutes at all builds the same prompt', () => {
+  // The inverse of the tripwire: if the two prompts ever differ, the field is
+  // reaching the panel through some path this file is not watching.
+  const withFigure = buildPrompt(BUNDLE);
+  const without = buildPrompt({ ...BUNDLE, task: { ...TASK, actual_minutes: null } });
+  assert.equal(withFigure, without, 'actual_minutes changes nothing about what the panel reads');
+});
+
+test('scoreEstActualBand is reachable only from the retroactive audit', () => {
+  // The one scorer that DOES read the figure. It lives in gradeDeterministic,
+  // which the grader module's own comment calls the legacy v1 entry — its single
+  // non-test caller is scripts/gds/audit-ships.js, a process audit run by hand
+  // over already-shipped tasks, not a gate anything passes through. A new caller
+  // is the same risk as wiring the figure into the prompt.
+  const grader = require('../modules/grading/grader.js');
+  assert.equal(typeof grader.gradeDeterministic, 'function', 'the legacy entry still exists');
+  assert.equal(grader.gradeSubagent.length > 0, true, 'and the live entry is a different function');
+
+  // gradeSubagent / gradeSubagentV3MultiShot take a single options object, so the
+  // guard that matters is which KEYS they destructure. Neither may take minutes.
+  const src = readFileSync(new URL('../modules/grading/grader.js', import.meta.url), 'utf8');
+  for (const fn of ['async function gradeSubagent(', 'async function gradeSubagentV3MultiShot(']) {
+    const at = src.indexOf(fn);
+    assert.ok(at !== -1, `${fn} is still the entry point this pins`);
+    const sig = src.slice(at, src.indexOf(')', at));
+    assert.ok(!/actualMinutes/.test(sig), `${fn} does not accept actualMinutes`);
+  }
+});