test(check): escape the runs-on fixture with JSON.stringify, not a quote-only replace (#11942)

CodeQL js/incomplete-sanitization (alert #888 on #11929): the hand-rolled replace
only escaped double quotes, so a backslash in the fixture would have produced a
malformed YAML scalar. JSON.stringify covers every escape the double-quoted YAML
scalar needs. Test-only change (7/7 pass).

.github/workflows/ci.yml

@@ -606,13 +606,24 @@ jobs:
     # Dynamic runner: when the release captain flips the USE_VPS_RUNNER repo var to
     # 'true' (scripts/vps/release-runner-up.sh does it after the self-hosted VM is
     # online), the heavy jobs run on the dedicated 32-core VPS runners (label
-    # omni-release) instead of queueing on the 20-concurrent-job hosted pool.
+    # omni-build) instead of queueing on the 20-concurrent-job hosted pool.
     # Safety: fork PRs NEVER reach the self-hosted runner — the expression falls
     # back to ubuntu-latest unless the PR head repo is this repository (push /
     # dispatch events are own-origin by definition). Any failure path (VM down,
     # var unset/false) also falls back to ubuntu-latest.
-    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository)) && fromJSON('["self-hosted","omni-release"]') || 'ubuntu-latest' }}
+    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository)) && fromJSON('["self-hosted","omni-build"]') || 'ubuntu-latest' }}
     needs: changes
+    # The .113 pool runs ONE next-build with room to spare and two at the edge: the
+    # box has 31 GB and a single next-build peaks at 14–16 GB RSS. On 2026-08-28
+    # 13:50Z the kernel OOM-killed main's build while a PR build ran beside it
+    # (five Build jobs had been queued by a burst of PRs). Two lanes: main keeps
+    # its own so a release is never queued behind PR traffic; PR builds serialize
+    # among themselves. GitHub keeps one running + one pending per group and
+    # CANCELS older pendings — a cancelled PR build is re-runnable; a dead main
+    # build costs the publish its artefact and a 40-minute rebuild that OOMs.
+    concurrency:
+      group: heavy-build-${{ github.ref == 'refs/heads/main' && 'main' || 'pr' }}
+      cancel-in-progress: false
     if: ${{ github.event_name != 'pull_request' || (needs.changes.outputs.code == 'true' && github.event.pull_request.draft == false) }}
     steps:
       - uses: actions/checkout@v7
@@ -646,14 +657,14 @@ jobs:
         # Keep standalone/node_modules intact: package/electron jobs consume the
         # Next-traced standalone tree and must not replace it with root node_modules.
         run: |
-          tar -czf /tmp/e2e-build.tar.gz \
+          tar -czf "$RUNNER_TEMP/e2e-build.tar.gz" \
             --exclude='.build/next/cache' \
             .build/next
       - name: Upload Next.js build for downstream jobs
         uses: actions/upload-artifact@v7
         with:
           name: next-build
-          path: /tmp/e2e-build.tar.gz
+          path: ${{ runner.temp }}/e2e-build.tar.gz
           retention-days: 1
 
   package-artifact:
@@ -676,10 +687,14 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: next-build
-          path: /tmp/
+          # Workspace-relative on purpose: the matrix below includes windows-latest, whose
+          # default shell is pwsh, where $RUNNER_TEMP is empty (it is $env:RUNNER_TEMP) —
+          # #11896's first cut broke the Electron smoke on exactly that. A relative path
+          # works in bash and pwsh alike; hosted workspaces are ephemeral.
+          path: next-build-artifact
       - name: Extract Next.js build artifact
         run: |
-          tar -xzf /tmp/e2e-build.tar.gz
+          tar -xzf next-build-artifact/e2e-build.tar.gz
       # build:cli consumes the downloaded .build/next standalone artifact and assembles dist/;
       # it only rebuilds if the downloaded standalone artifact is missing.
       - run: npm run build:cli
@@ -767,10 +782,14 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: next-build
-          path: /tmp/
+          # Workspace-relative on purpose: the matrix below includes windows-latest, whose
+          # default shell is pwsh, where $RUNNER_TEMP is empty (it is $env:RUNNER_TEMP) —
+          # #11896's first cut broke the Electron smoke on exactly that. A relative path
+          # works in bash and pwsh alike; hosted workspaces are ephemeral.
+          path: next-build-artifact
       - name: Extract Next.js build artifact
         run: |
-          tar -xzf /tmp/e2e-build.tar.gz
+          tar -xzf next-build-artifact/e2e-build.tar.gz
       - name: Install Electron dependencies
         working-directory: electron
         run: npm install --no-audit --no-fund
@@ -1230,10 +1249,14 @@ jobs:
         uses: actions/download-artifact@v8
         with:
           name: next-build
-          path: /tmp/
+          # Workspace-relative on purpose: the matrix below includes windows-latest, whose
+          # default shell is pwsh, where $RUNNER_TEMP is empty (it is $env:RUNNER_TEMP) —
+          # #11896's first cut broke the Electron smoke on exactly that. A relative path
+          # works in bash and pwsh alike; hosted workspaces are ephemeral.
+          path: next-build-artifact
       - name: Extract Next.js build artifact
         run: |
-          tar -xzf /tmp/e2e-build.tar.gz
+          tar -xzf next-build-artifact/e2e-build.tar.gz
       # WS4.1: duration-balanced shards (LPT over config/quality/e2e-timings.json).
       # Measured skew of plain --shard was 14× (24m47s vs 1m47s) — E2E was the CI
       # critical path. The balancer self-verifies completeness and exits non-zero on

.github/workflows/nightly-release-green.yml

@@ -68,7 +68,7 @@ jobs:
     # this runs on the dedicated VPS runner — clean env (no operator OMNIROUTE_API_KEY,
     # no local noauth CLIs => zero machine-specific false positives) and no contention.
     # Nightly cron normally finds the var false (VM off) and falls back to hosted.
-    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && fromJSON('["self-hosted","omni-release"]')) || 'ubuntu-latest' }}
+    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && fromJSON('["self-hosted","omni-build"]')) || 'ubuntu-latest' }}
     env:
       JWT_SECRET: ci-nightly-secret-with-sufficient-length-for-validation
       API_KEY_SECRET: ci-nightly-api-key-secret-long
@@ -217,7 +217,7 @@ jobs:
     # On a push, only run for a push to main — a push to release/* is handled by
     # release-green above. Schedule/dispatch always run (they also sweep main).
     if: ${{ github.event_name != 'push' || github.ref_name == 'main' }}
-    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && fromJSON('["self-hosted","omni-release"]')) || 'ubuntu-latest' }}
+    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && fromJSON('["self-hosted","omni-build"]')) || 'ubuntu-latest' }}
     env:
       JWT_SECRET: ci-nightly-secret-with-sufficient-length-for-validation
       API_KEY_SECRET: ci-nightly-api-key-secret-long

.github/workflows/npm-publish.yml

@@ -23,11 +23,12 @@ on:
           - next
           - historic
       publish_mode:
-        description: "staged = npm stage publish (owner approves with 2FA after the staged boot-verify); direct = legacy immediate publish (emergency fallback only)"
+        description: "auto = publish through npm Trusted Publishing (OIDC, no token, no 2FA prompt — the default); staged = npm stage publish (owner approves with 2FA); direct = legacy token publish (emergency fallback only)"
         required: false
-        default: "staged"
+        default: "auto"
         type: choice
         options:
+          - auto
           - staged
           - direct
   workflow_call:
@@ -62,7 +63,7 @@ jobs:
     # mid-"Creating an optimized production build" while v3.8.48 had still fit in 16min.
     # This job never runs on `pull_request`, so the fork-safety clause is always true here;
     # it is kept verbatim so the expression stays greppable against ci.yml.
-    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository)) && fromJSON('["self-hosted","omni-release"]') || 'ubuntu-latest' }}
+    runs-on: ${{ (vars.USE_VPS_RUNNER == 'true' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository)) && fromJSON('["self-hosted","omni-build"]') || 'ubuntu-latest' }}
     outputs:
       version: ${{ steps.resolve.outputs.version }}
       tag: ${{ steps.resolve.outputs.tag }}
@@ -204,8 +205,11 @@ jobs:
             exit 0
           fi
           RUN=""
+          # $RUNNER_TEMP, never /tmp: on the .113 pool /tmp is a 12 GB tmpfs (RAM). Parking
+          # this 1.3 GB artefact there took 27–32 min of the 76-min publish job — the
+          # same bytes upload from disk in 2 min. RUNNER_TEMP is per-runner and on disk.
           for candidate in $CANDIDATES; do
-            if gh run download "$candidate" --repo "$REPO" --name next-build --dir /tmp/next-build 2>/dev/null; then
+            if gh run download "$candidate" --repo "$REPO" --name next-build --dir "$RUNNER_TEMP/next-build" 2>/dev/null; then
               RUN="$candidate"
               break
             fi
@@ -215,8 +219,8 @@ jobs:
             echo "::notice::none of the candidate runs still carries next-build (1-day retention) — falling back to a full build"
             exit 0
           fi
-          tar -xzf /tmp/next-build/e2e-build.tar.gz -C .
-          rm -rf /tmp/next-build
+          tar -xzf "$RUNNER_TEMP/next-build/e2e-build.tar.gz" -C .
+          rm -rf "$RUNNER_TEMP/next-build"
           if [ -f .build/next/standalone/server.js ]; then
             echo "✅ standalone tree restored from CI run $RUN — build:cli will skip next build"
           else
@@ -404,8 +408,34 @@ jobs:
           fi
           npm --version
 
+      # Trusted Publishing (OIDC): npm mints a short-lived credential for THIS run from
+      # GitHub's id-token — no NPM_TOKEN secret, no 2FA prompt, provenance included, and
+      # it is the bypass npm sanctions now that tokens which skip 2FA are being retired
+      # (gh.io/npm-gat-bypass2fa-deprecation). Requires the package's Trusted Publisher to
+      # be configured on npmjs.com (owner: diegosouzapw/OmniRoute, workflow
+      # npm-publish.yml) and a github-hosted runner — which is why this job exists.
+      # Without that configuration `npm publish` fails with ENEEDAUTH: re-dispatch with
+      # publish_mode=staged or direct. Automatic publishing was the flow up to v3.8.48;
+      # v3.8.49 moved to staged (WS1.3) to keep a leaked token from publishing alone —
+      # OIDC gives the same guarantee without the manual approve.
+      - name: Publish to npm (Trusted Publishing / OIDC — automatic)
+        if: github.event_name != 'workflow_dispatch' || inputs.publish_mode == 'auto'
+        env:
+          VERSION: ${{ needs.publish.outputs.version }}
+          TAG: ${{ needs.publish.outputs.tag }}
+        run: |
+          set -euo pipefail
+          TARBALL="omniroute-${VERSION}.tgz"
+          test -f "$TARBALL" || { echo "tarball $TARBALL did not arrive from the publish job" >&2; ls -la; exit 1; }
+          # Deliberately NO NODE_AUTH_TOKEN in this step: npm >= 11.5 detects the GitHub
+          # OIDC token itself. Always pass --tag explicitly (defense in depth: an older
+          # VERSION can never claim `@latest`).
+          npm publish "$TARBALL" --provenance --access public --tag "$TAG" --ignore-scripts
+          echo "✅ Published omniroute@$VERSION (dist-tag=$TAG) via Trusted Publishing"
+
       - name: Publish to npm (staged — owner approves with 2FA)
-        if: github.event_name != 'workflow_dispatch' || inputs.publish_mode != 'direct'
+        # Only on an explicit request now: Trusted Publishing below is the default.
+        if: github.event_name == 'workflow_dispatch' && inputs.publish_mode == 'staged'
         env:
           VERSION: ${{ needs.publish.outputs.version }}
           TAG: ${{ needs.publish.outputs.tag }}

changelog.d/features/npm-trusted-publishing-oidc.md

@@ -0,0 +1,4 @@
+- The npm publish is automatic again, through npm Trusted Publishing (OIDC): the hosted
+  `stage-npm` job publishes with a short-lived credential minted from GitHub's id-token —
+  no `NPM_TOKEN`, no 2FA prompt, provenance attached. `publish_mode=staged` (owner
+  approves with 2FA) and `direct` (token) remain available on `workflow_dispatch`.

changelog.d/fixes/11866-alibaba-allowlist-test-timebomb.md

@@ -0,0 +1,5 @@
+- Fixed the Alibaba free-tier allowlist test that went red on its own once the
+  shipped catalog's `validUntil` (2026-08-27) passed, leaving every PR and `main`
+  with a failing `Unit Tests (1/8)`. The test now builds its own packs with dates
+  it controls, and covers the expired-pack fallback that production has actually
+  been serving.

changelog.d/maintenance/11866-config-expiry-time-bomb-test.md

@@ -0,0 +1,4 @@
+- Added a unit test that fails seven days before any dated pack under `config/`
+  (`validUntil` and sibling keys) lapses, naming the file and key. The Alibaba
+  free-tier pack expired on 2026-08-27 and turned every PR red the next morning
+  with no commit involved; renewal now happens on someone's terms, not the clock's.

changelog.d/maintenance/11878-check-workflows-provenance-self-hosted.md

@@ -0,0 +1,3 @@
+- `check:workflows` now fails (under `--strict`/`--ratchet`) when any job routed to a
+  self-hosted runner publishes with `--provenance` — npm rejects that with `422` at the
+  registry, which in v3.8.50 only surfaced after the tag and Docker images were public.

changelog.d/maintenance/11892-runner-janitor-act-not-advise.md

@@ -0,0 +1,5 @@
+- `scripts/ops/runner-janitor.sh` now proves a path is idle with one `lsof`
+  snapshot and removes stale leftovers itself (tmpfs after 3 h — it is RAM — disk
+  after 24 h), kills orphan `next-build` processes, prunes checkouts of stopped
+  runners, and alerts on memory pressure; `--dry-run` shows exactly what it would
+  do. `docs/ops/RUNNER_BOX.md` reconciled to the measured box (31 GB, 10 listeners).

changelog.d/maintenance/11896-ci-artifact-download-to-disk.md

@@ -0,0 +1,5 @@
+- The `next-build` artefact (1.3 GB) is now written and read under `$RUNNER_TEMP`
+  (per-runner, on disk) instead of `/tmp`, which on the self-hosted pool is a
+  12 GB tmpfs in RAM. Landing it there took 27–32 of the publish job's 76 minutes,
+  and the fixed `/tmp/e2e-build.tar.gz` name let E2E jobs on different runners
+  overwrite each other's download.

changelog.d/maintenance/11897-ci-heavy-build-lane.md

@@ -0,0 +1,3 @@
+- The CI `build` job now runs in two concurrency lanes — `main` and pull requests —
+  so a release build is never queued behind (or OOM-killed beside) PR builds on the
+  self-hosted pool, which holds one `next-build` comfortably and two at the edge.

changelog.d/maintenance/ci-omni-build-runner-label.md

@@ -0,0 +1,4 @@
+- Every CI job that runs a `next build` (`build`, the npm `publish`, both release-green
+  validations) now targets the `omni-build` runner label, which only two of the eight
+  self-hosted runners carry. The box holds one build comfortably and two at the edge; a
+  third now queues on GitHub instead of being OOM-killed by the kernel.

docs/ops/RELEASE_CHECKLIST.md

@@ -1,12 +1,12 @@
 ---
 title: "Release Checklist"
-version: 3.8.40
-lastUpdated: 2026-06-28
+version: 3.8.51
+lastUpdated: 2026-08-28
 ---
 
 # Release Checklist
 
-> **Last updated:** 2026-06-28 — v3.8.40
+> **Last updated:** 2026-08-28 — v3.8.51
 > Streamlined release flow that leverages Claude Code skills for automation.
 >
 > **Keep the queue/branch green between releases:** see [RELEASE_GREEN.md](./RELEASE_GREEN.md)
@@ -37,7 +37,21 @@ npm run test:e2e           # optional but recommended
 /capture-release-evidences-cc
 ```
 
-## npm Staged Publishing (default since v3.8.49 — WS1.3/D2)
+## npm Trusted Publishing (default since v3.8.51) — staged on request, direct as fallback
+
+`npm-publish.yml` publishes through **npm Trusted Publishing (OIDC)** by default: the
+`stage-npm` job (github-hosted) exchanges GitHub's id-token for a short-lived npm
+credential for that run — no long-lived npm token in the repository secrets, no 2FA prompt, provenance attached.
+That is the bypass npm sanctions now that tokens which skip 2FA are being retired;
+it restores the fully automatic flow the project had up to v3.8.48 while keeping the
+WS1.3 guarantee (a leaked token cannot publish alone — there is no token).
+
+**One-time setup (owner):** npmjs.com → package `omniroute` → Settings → *Trusted
+Publisher* → GitHub: owner `diegosouzapw`, repo `OmniRoute`, workflow `npm-publish.yml`
+(environment: none). Until that exists, the automatic step fails with `ENEEDAUTH`:
+re-dispatch with `publish_mode=staged` (below) or `direct`.
+
+### Staged publishing (on request — `publish_mode=staged`)
 
 The npm-publish workflow no longer publishes directly: it boots the packed tarball
 (`check:pack-boot`) and then runs `npm stage publish` — the exact bytes are parked on

docs/ops/RUNNER_BOX.md

@@ -4,32 +4,66 @@ title: Self-Hosted Runner Box Operations
 
 # Self-Hosted Runner Box Operations (.113 pool)
 
-The self-hosted pool (`self-hosted, omni-release` labels) runs on the 16 GB box at
-`192.168.0.113`. Two failure modes recurred on release days and were, until v3.8.49,
-manual discipline; the **janitor script codifies them** (WS3.3 of the quality plan):
+The self-hosted pool (`self-hosted, omni-release` on all eight runners; `omni-build` on two) runs on the **.113** box.
+Measured 2026-08-28 (v3.8.50 postmortem, Parte III):
 
-1. **Orphaned temp/work dirs** filling the disk → disk-full SQLite errors mid-job.
-2. **>4 concurrent runners** → OOM-killed jobs (8-wide killed jobs twice on the
-   v3.8.47 release day; 4-wide is the proven ceiling).
+| resource  | value                                                            | what it means for scheduling                                                                                                                  |
+| --------- | ---------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------- |
+| RAM / CPU | **31 GB / 32 cores** (was 16 GB when this doc was first written) | one `next-build` peaks at **~14 GB** → 2 concurrent heavy builds saturate the box, 3 take it down (2026-08-28 06:42Z: load 56, two jobs lost) |
+| swap      | 15 GB                                                            | it swapped its way through the v3.8.50 publish; pressure shows in `/proc/pressure/memory`                                                     |
+| `/tmp`    | **12 GB tmpfs = RAM**                                            | anything parked there is memory; leftovers are swept after 3 h                                                                                |
+| disk      | 188 GB                                                           | `_work` checkouts of 8 runners reach ~70 GB with no cap                                                                                       |
+| runners   | **10 listeners**: 8 OmniRoute + OmniHeuris + OmniMind            | all share the memory above                                                                                                                    |
 
 ## Install the janitor (one-time, on the box)
 
 ```bash
-sudo mkdir -p /opt/omniroute-ops
-sudo cp scripts/ops/runner-janitor.sh /opt/omniroute-ops/
-sudo chmod +x /opt/omniroute-ops/runner-janitor.sh
-( sudo crontab -l 2>/dev/null; echo '*/30 * * * * /opt/omniroute-ops/runner-janitor.sh >> /var/log/runner-janitor.log 2>&1' ) | sudo crontab -
+scp scripts/ops/runner-janitor.sh root@192.168.0.113:/opt/omniroute-ops/runner-janitor.sh
+ssh root@192.168.0.113 'chmod +x /opt/omniroute-ops/runner-janitor.sh; apt-get install -y lsof'
+# cron (root): every 30 min, log to /var/log/runner-janitor.log
+*/30 * * * * MAX_ACTIVE_RUNNERS=8 /opt/omniroute-ops/runner-janitor.sh >> /var/log/runner-janitor.log 2>&1
 ```
 
-What it does every 30min: sweeps runner temp leftovers older than 24h, alerts at
-≥85% root-disk usage, and alerts when more than the runner ceiling (default 4, tunable
-via the script's own environment) of `Runner.Listener` processes are up. Alerts land in `/var/log/runner-janitor.log`
-with a non-zero exit (grep for `⚠`).
+`lsof` is required: the janitor proves a path is idle with one snapshot of open
+files before removing it, and without the tool it removes nothing and says so
+(exit 1). Try any change with `--dry-run` first — it prints exactly what it would
+do and touches nothing.
+
+What it does every run: sweeps our own leftovers (`runner-*`, `omniroute-*`,
+`next-build*`, `e2e-build.tar.gz`) after **3 h on tmpfs** and 24 h on disk
+`_work/_temp`; kills a `next-build` older than 75 min (no job runs that long — on
+2026-08-27 one ran 70 min after GitHub had declared its job lost); prunes 48 h-old
+checkouts of runners whose unit is **stopped**; alerts on disk ≥ 85 %, memory PSI
+`full/avg60` ≥ 10 %, and more listeners than `MAX_ACTIVE_RUNNERS` (with an
+omniroute/other breakdown). Exit 1 = attention needed; read the log.
+
+## Runner units: KillMode
+
+The runner's default `KillMode=process` leaves `Runner.Worker → npm → next-build`
+alive when a unit is stopped or restarted — an orphan build keeps eating RAM and
+CPU with no job attached. Every OmniRoute unit carries a drop-in
+(`/etc/systemd/system/actions.runner.diegosouzapw-OmniRoute.<name>.service.d/10-killmode.conf`)
+with `KillMode=mixed`: SIGTERM to the listener first, SIGKILL to the whole cgroup at
+`TimeoutStop`. It takes effect on the unit's next restart — restart **one runner at
+a time, only when idle**, with the idle check and the restart in the same command.
 
 ## Operating rules
 
-- **Ceiling: 4 runners** on the 16 GB box. Runners 5–8 stay STOPPED except for
-  explicit off-peak experiments — never during a release window.
-- Stopping a runner mid-job cancels the job (observed live): `systemctl stop`
-  only when its runner is idle (`Runner.Listener` without a `Runner.Worker` child).
+- **Heavy-build ceiling: 2 at a time — enforced by label.** Every job that runs a
+  `next build` (`ci.yml` `build`, `npm-publish.yml` `publish`, both `nightly-release-green`
+  validations) targets `[self-hosted, omni-build]`, and only **two** runners carry that
+  label (`omniroute-113-5`, `omniroute-113-6`, added through the runners API — no
+  re-registration). The other six keep `omni-release` and take nothing heavy; GitHub
+  queues a third build instead of the kernel killing one. Pair with the `heavy-build-*`
+  concurrency lanes in `ci.yml`. To add capacity, label another runner — never raise
+  the count past what 31 GB holds (one next-build ≈ 14–16 GB).
+- **Never clean `/tmp` or `_work` by hand while any runner is busy.** A
+  check-then-delete with a gap between the two is how a live Build job lost its
+  `_work` on 2026-08-27. The janitor does the check and the removal in one step;
+  let it.
+- Stopping a runner mid-job cancels the job (observed live): `systemctl stop` only
+  when its listener has no `Runner.Worker` child — and do it in one command.
+- Workflows must not park artefacts in `/tmp` (it is RAM). Download to
+  `$RUNNER_TEMP` (on disk, per runner) — the 1.3 GB `next-build` artefact took 27–32
+  minutes to land on the tmpfs and 2 minutes to upload from disk.
 - The `.15` VPS is homologation-only — never runs CI runners.

scripts/check/check-workflows.mjs

@@ -42,6 +42,7 @@ import { execFileSync, spawnSync } from "node:child_process";
 import fs from "node:fs";
 import path from "node:path";
 import { pathToFileURL } from "node:url";
+import { findProvenanceOnSelfHosted, formatProvenanceFinding } from "./lib/provenanceRunner.mjs";
 
 const ROOT = process.cwd();
 const WORKFLOWS_DIR = path.join(ROOT, ".github", "workflows");
@@ -275,6 +276,23 @@ export function runZizmor(workflowsDir) {
 // Main
 // ---------------------------------------------------------------------------
 
+/**
+ * Hard rule (not a lint count): `--provenance` inside a job that runs on a
+ * self-hosted runner. npm answers 422 at the registry, and in v3.8.50 that
+ * answer only came after the tag, the GitHub Release and the Docker images were
+ * already out. Blocks under --strict AND --ratchet (the CI mode); plain mode
+ * reports it like everything else.
+ * @param {string[]} files absolute workflow paths
+ */
+export function runProvenanceRunnerCheck(files) {
+  const findings = [];
+  for (const file of files) {
+    const text = fs.readFileSync(file, "utf8");
+    findings.push(...findProvenanceOnSelfHosted(text, path.relative(ROOT, file)));
+  }
+  return findings;
+}
+
 function main() {
   const hasActionlint = isBinaryAvailable("actionlint");
   const hasZizmor = isBinaryAvailable("zizmor");
@@ -350,6 +368,16 @@ function main() {
     }
   }
 
+  const provenanceFindings = runProvenanceRunnerCheck(workflowFiles);
+  if (provenanceFindings.length > 0) {
+    console.error(
+      `[check-workflows] provenance×self-hosted: ${provenanceFindings.length} finding(s) — HARD RULE:`
+    );
+    provenanceFindings.forEach((f) => console.error(`  ${formatProvenanceFinding(f)}`));
+  } else if (!QUIET) {
+    console.log("[check-workflows] provenance×self-hosted: OK (0 findings)");
+  }
+
   const total = actionlintCount + zizmorCount;
   process.stdout.write(`workflowFindings=${total}\n`);
   process.stdout.write(`actionlintFindings=${actionlintCount}\n`);
@@ -357,6 +385,15 @@ function main() {
   // Read this line with the count above: a finding total is only reproducible against the
   // version that produced it. See zizmorVersion().
   process.stdout.write(`zizmorVersion=${hasZizmor ? zizmorVersion() : "absent"}\n`);
+  process.stdout.write(`provenanceRunnerFindings=${provenanceFindings.length}\n`);
+  if ((STRICT || RATCHET) && provenanceFindings.length > 0) {
+    console.error(
+      `\n[check-workflows] FAIL — ${provenanceFindings.length} job(s) publish with --provenance from a self-hosted runner.\n` +
+        "  npm rejects that with 422 at the registry. Move the upload step to a github-hosted job\n" +
+        "  (see .github/workflows/npm-publish.yml `stage-npm` for the pattern)."
+    );
+    process.exit(1);
+  }
 
   if (STRICT && total > 0) {
     console.error(`\n[check-workflows] FAIL — ${total} workflow finding(s) total (--strict mode).`);

scripts/check/lib/configExpiry.mjs

@@ -0,0 +1,90 @@
+/**
+ * scripts/check/lib/configExpiry.mjs
+ *
+ * Finds dated validity fields in JSON config packs so a test can fail BEFORE
+ * they lapse. Origin: config/alibaba-free-tier-allowlist.json carried
+ * `"validUntil": "2026-08-27"`; on 2026-08-28 the loader started (correctly)
+ * rejecting the pack and a test that asserted "the shipped pack loads" turned
+ * every PR and main red with no commit involved (#11866). A time bomb, not a
+ * regression — and the only kind of defect a diff review can never catch.
+ *
+ * Pure helpers; the repo-wide assertion lives in
+ * tests/unit/config-expiry-time-bomb.test.ts.
+ */
+import fs from "node:fs";
+import path from "node:path";
+
+export const EXPIRY_KEY =
+  /^(validUntil|valid_until|validTo|valid_to|expiresAt|expires_at|expiry|expires)$/;
+const DAY_MS = 86_400_000;
+
+/**
+ * Walks a parsed JSON value and returns every string-valued expiry field.
+ * @returns {{ file: string, keyPath: string, raw: string, expiresAt: number|null }[]}
+ */
+export function collectExpiryFields(value, file, keyPath = []) {
+  const out = [];
+  if (Array.isArray(value)) {
+    value.forEach((v, i) => out.push(...collectExpiryFields(v, file, [...keyPath, String(i)])));
+    return out;
+  }
+  if (!value || typeof value !== "object") return out;
+  for (const [key, v] of Object.entries(value)) {
+    const kp = [...keyPath, key];
+    if (EXPIRY_KEY.test(key) && typeof v === "string") {
+      const ms = Date.parse(v);
+      out.push({ file, keyPath: kp.join("."), raw: v, expiresAt: Number.isFinite(ms) ? ms : null });
+    } else if (v && typeof v === "object") {
+      out.push(...collectExpiryFields(v, file, kp));
+    }
+  }
+  return out;
+}
+
+/** @returns {"expired"|"expiring"|"ok"|"unparseable"} */
+export function classifyExpiry(field, nowMs, warnDays = 7) {
+  if (field.expiresAt === null) return "unparseable";
+  if (field.expiresAt < nowMs) return "expired";
+  if (field.expiresAt < nowMs + warnDays * DAY_MS) return "expiring";
+  return "ok";
+}
+
+/** All *.json under dir, recursively, skipping node_modules. Sorted for stable output. */
+export function walkJsonFiles(dir) {
+  const out = [];
+  const stack = [dir];
+  while (stack.length > 0) {
+    const current = stack.pop();
+    let entries;
+    try {
+      entries = fs.readdirSync(current, { withFileTypes: true });
+    } catch {
+      continue;
+    }
+    for (const e of entries) {
+      const full = path.join(current, e.name);
+      if (e.isDirectory()) {
+        if (e.name !== "node_modules") stack.push(full);
+      } else if (e.isFile() && e.name.endsWith(".json")) {
+        out.push(full);
+      }
+    }
+  }
+  return out.sort();
+}
+
+/**
+ * Scans every JSON file under `dir`; `file` in the result is relative to `dir`
+ * with forward slashes, so allowlists can key on it portably.
+ */
+export function scanConfigExpiry(dir) {
+  return walkJsonFiles(dir).flatMap((f) => {
+    let parsed;
+    try {
+      parsed = JSON.parse(fs.readFileSync(f, "utf8"));
+    } catch {
+      return []; // not this scanner's job to validate JSON
+    }
+    return collectExpiryFields(parsed, path.relative(dir, f).split(path.sep).join("/"));
+  });
+}

scripts/check/lib/provenanceRunner.mjs

@@ -0,0 +1,83 @@
+/**
+ * scripts/check/lib/provenanceRunner.mjs
+ *
+ * npm refuses `--provenance` from a self-hosted runner:
+ *
+ *   422 Unprocessable Entity - Error verifying sigstore provenance bundle:
+ *   Unsupported GitHub Actions runner environment: "self-hosted".
+ *   Only "github-hosted" runners are supported when publishing with provenance.
+ *
+ * v3.8.50 hit this at the very end of a 76-minute publish job — after the tag,
+ * the GitHub Release and the Docker images were already public — because
+ * `USE_VPS_RUNNER` had been turned on (2026-08-02) with no release in between to
+ * surface it. The combination is greppable, so it must fail in CI the moment a
+ * workflow introduces it, not four weeks later at the registry.
+ *
+ * Pure: takes workflow YAML text, returns the offending (job, step) pairs.
+ */
+import { load as yamlLoad } from "js-yaml";
+
+const SELF_HOSTED = /\bself-hosted\b/;
+const EXPRESSION = /\$\{\{/;
+// Lookahead, not \b: `--provenance-file=…` is a different flag (a pre-built
+// bundle) and must not match — a word boundary sits between "e" and "-".
+const PROVENANCE = /(^|\s)--provenance(?=\s|=|$)/m;
+
+/**
+ * Classifies a job's `runs-on` value.
+ * @returns {"self-hosted"|"hosted"|"unknown"}
+ *   "unknown" = an expression with no literal `self-hosted` in it (e.g.
+ *   `${{ matrix.os }}`); the check does not guess, it skips.
+ */
+export function classifyRunsOn(runsOn) {
+  if (runsOn == null) return "unknown";
+  if (typeof runsOn === "string") {
+    if (SELF_HOSTED.test(runsOn)) return "self-hosted";
+    return EXPRESSION.test(runsOn) ? "unknown" : "hosted";
+  }
+  if (Array.isArray(runsOn)) {
+    return runsOn.some((v) => typeof v === "string" && SELF_HOSTED.test(v))
+      ? "self-hosted"
+      : "hosted";
+  }
+  if (typeof runsOn === "object") {
+    // { group: ..., labels: ... } form
+    const labels = runsOn.labels;
+    return classifyRunsOn(Array.isArray(labels) ? labels : labels == null ? "" : String(labels));
+  }
+  return "unknown";
+}
+
+/**
+ * @param {string} yamlText
+ * @param {string} fileName  used only for reporting
+ * @returns {{ file: string, job: string, step: string }[]}
+ */
+export function findProvenanceOnSelfHosted(yamlText, fileName = "<workflow>") {
+  let doc;
+  try {
+    doc = yamlLoad(yamlText);
+  } catch {
+    // actionlint owns syntax; an unparseable file is not this rule's finding.
+    return [];
+  }
+  const jobs =
+    doc && typeof doc === "object" && doc.jobs && typeof doc.jobs === "object" ? doc.jobs : {};
+  const findings = [];
+  for (const [jobName, job] of Object.entries(jobs)) {
+    if (!job || typeof job !== "object") continue;
+    if (classifyRunsOn(job["runs-on"]) !== "self-hosted") continue;
+    const steps = Array.isArray(job.steps) ? job.steps : [];
+    steps.forEach((step, i) => {
+      if (step && typeof step.run === "string" && PROVENANCE.test(step.run)) {
+        findings.push({ file: fileName, job: jobName, step: step.name || `#${i + 1}` });
+      }
+    });
+  }
+  return findings;
+}
+
+/** Human-readable line per finding, used by the CLI. */
+export function formatProvenanceFinding(f) {
+  return `${f.file}: job "${f.job}", step "${f.step}" runs \`--provenance\` on a self-hosted runner — npm rejects that (422). Move the upload to a github-hosted job.`;
+}

scripts/ops/runner-janitor.sh

@@ -1,53 +1,172 @@
 #!/usr/bin/env bash
-# runner-janitor — self-hosted runner box hygiene (WS3.3, v3.8.49 quality plan).
+# runner-janitor — self-hosted runner box hygiene for the .113 pool.
 #
-# The .113 runner box has recurring failure modes that until now were manual
-# discipline: orphaned tmpfs/work dirs filling the disk, and >4 concurrent
-# runners OOM-killing jobs (16 GB box; incidents on the v3.8.47 release day).
-# Install via cron on the box (see docs/ops/RUNNER_BOX.md):
-#   */30 * * * *  /opt/omniroute-ops/runner-janitor.sh >> /var/log/runner-janitor.log 2>&1
+# Runs from cron every 30 min (see docs/ops/RUNNER_BOX.md). It ACTS on what it
+# can prove is safe and ALERTS on what needs an operator decision. Reads of
+# "is this in use?" and the removal happen in the same command, never in two
+# passes: a check-then-delete with a gap is how a live Build job lost its _work
+# on 2026-08-27.
 #
+# Measured box (2026-08-28): 31 GB RAM, 32 cores, 15 GB swap, /tmp = 12 GB
+# tmpfs (RAM!), 188 GB disk. A single `next-build` peaks at ~14 GB, so two
+# concurrent heavy builds saturate the box and three take it down (06:42Z that
+# day: load 56, two jobs lost). The v3.8.50 postmortem (Parte III) has the numbers.
+#
+# What it does, in order:
+#   1) sweep stale artefacts our tooling leaves behind — tmpfs bases after 3 h
+#      (they hold RAM), disk _work/_temp bases after 24 h; only names we create,
+#      only when no process has them open
+#   2) kill zombie builds: a `next-build` older than ZOMBIE_BUILD_MAX_MIN has no
+#      job attached (a real Build step measures ~26 min). On 2026-08-27 one ran
+#      70 minutes after GitHub had already declared its job lost, eating 3.6 GB
+#      and a full core set. KillMode=mixed on the units covers systemctl
+#      stop/restart; this covers the lost-connection path.
+#   3) prune 48 h-old checkouts under _work of runners whose unit is INACTIVE
+#      (stopped runners cannot be mid-job; active ones are never touched)
+#   4) alert: root disk >= DISK_ALERT_PCT, memory PSI full/avg60 >= threshold,
+#      Runner.Listener count above the ceiling (with a per-project breakdown —
+#      the box also hosts OmniHeuris and OmniMind runners)
+#
+# Usage: runner-janitor.sh [--dry-run] [--help]
 # Exit codes: 0 healthy · 1 attention needed (printed to stdout for the log).
 set -euo pipefail
 
-MAX_ACTIVE_RUNNERS="${MAX_ACTIVE_RUNNERS:-4}"
-DISK_ALERT_PCT="${DISK_ALERT_PCT:-85}"
-WORK_DIR_MAX_AGE_HOURS="${WORK_DIR_MAX_AGE_HOURS:-24}"
-STATUS=0
-
-echo "[janitor] $(date -u +%FT%TZ) start"
-
-# 1) Sweep stale runner temp/work leftovers (>24h — no legitimate job runs that long).
-# Hardened for a root cron on world-writable paths: never follow a symlinked base
-# (a compromised runner could plant one), -P + -xdev so the sweep cannot traverse
-# out of the filesystem, and patterns narrowed to names OUR tooling creates
-# (no generic tmp* — unrelated system temp files are out of scope).
-for base in /tmp /home/*/actions-runner*/_work/_temp; do
-  [ -d "$base" ] || continue
-  [ -L "$base" ] && { echo "[janitor] skip symlinked base: $base"; continue; }
-  find -P "$base" -xdev -maxdepth 1 \( -name 'runner-*' -o -name 'omniroute-*' \) \
-    ! -type l -mmin +$((WORK_DIR_MAX_AGE_HOURS * 60)) -exec rm -rf {} + 2>/dev/null || true
+DRY_RUN=0
+for arg in "$@"; do
+  case "$arg" in
+    --dry-run) DRY_RUN=1 ;;
+    -h|--help)
+      sed -n '2,32p' "$0" | sed 's/^# \{0,1\}//'
+      exit 0 ;;
+    *) echo "unknown argument: $arg" >&2; exit 2 ;;
+  esac
 done
-echo "[janitor] stale temp sweep done"
 
-# 2) Disk pressure — alert loudly before SQLITE_FULL kills jobs mid-run.
-USAGE=$(df --output=pcent / | tail -1 | tr -dc '0-9')
-if [ "$USAGE" -ge "$DISK_ALERT_PCT" ]; then
-  echo "[janitor] ⚠ ROOT DISK ${USAGE}% >= ${DISK_ALERT_PCT}% — clean before the next heavy run"
-  STATUS=1
+MAX_ACTIVE_RUNNERS="${MAX_ACTIVE_RUNNERS:-8}"
+DISK_ALERT_PCT="${DISK_ALERT_PCT:-85}"
+TMPFS_MAX_AGE_HOURS="${TMPFS_MAX_AGE_HOURS:-3}"
+WORK_TEMP_MAX_AGE_HOURS="${WORK_TEMP_MAX_AGE_HOURS:-24}"
+WORK_CHECKOUT_MAX_AGE_HOURS="${WORK_CHECKOUT_MAX_AGE_HOURS:-48}"
+ZOMBIE_BUILD_MAX_MIN="${ZOMBIE_BUILD_MAX_MIN:-75}"
+ZOMBIE_BUILD_COMM="${ZOMBIE_BUILD_COMM:-next-build}"
+PSI_FULL_AVG60_ALERT="${PSI_FULL_AVG60_ALERT:-10}"
+# Overridable so the unit test can point everything at a fixture tree.
+JANITOR_TMP_BASES="${JANITOR_TMP_BASES-/tmp}"
+JANITOR_WORK_TEMP_BASES="${JANITOR_WORK_TEMP_BASES-/opt/actions-runner*/_work/_temp /home/*/actions-runner*/_work/_temp}"
+JANITOR_RUNNER_DIRS="${JANITOR_RUNNER_DIRS-/opt/actions-runner*}"
+JANITOR_PSI_FILE="${JANITOR_PSI_FILE:-/proc/pressure/memory}"
+JANITOR_DF_PATH="${JANITOR_DF_PATH:-/}"
+
+STATUS=0
+say() { echo "[janitor] $*"; }
+
+# "Is anything using this?" — ONE snapshot of every open path on the box
+# (lsof -Fn), then a prefix match per candidate. `lsof +D <dir>` walks the whole
+# tree instead and took minutes on a 5 GB leftover — unusable from cron. An
+# absent lsof means "cannot prove idle": the sweep keeps the path and says so.
+LSOF_BIN="${JANITOR_LSOF:-lsof}"
+have_busy_tools() { command -v "$LSOF_BIN" >/dev/null 2>&1; }
+SNAP=""
+cleanup() { [ -n "$SNAP" ] && rm -f -- "$SNAP"; }
+trap cleanup EXIT
+# One lsof for the whole run (~13 s / 83k lines on the box), kept ONLY for the
+# bases we sweep — 460 candidates grepping a re-printed 83k-line string was the
+# slow part, not lsof itself.
+snapshot_open_paths() {
+  have_busy_tools || return 0
+  SNAP=$(mktemp) || return 0
+  local prefixes="" b
+  for b in $JANITOR_TMP_BASES $JANITOR_WORK_TEMP_BASES; do [ -d "$b" ] && prefixes="$prefixes"$'\n'"$b/"; done
+  # -F n: one "n<path>" line per open file; -w: no warnings
+  "$LSOF_BIN" -w -Fn 2>/dev/null | sed -n 's/^n//p' | grep -F -f <(printf '%s' "$prefixes" | sed '/^$/d') > "$SNAP" 2>/dev/null || true
+}
+is_busy() {
+  local p="$1"
+  [ -n "$SNAP" ] && [ -s "$SNAP" ] || return 1
+  # exact path, or anything beneath it when it is a directory
+  grep -qxF -- "$p" "$SNAP" && return 0
+  [ -d "$p" ] && grep -qF -- "$p/" "$SNAP"
+}
+
+# sweep <base> <max-age-minutes>: only names our tooling creates, never through
+# a symlinked base, never across a filesystem, and remove+check in one step.
+sweep() {
+  local base="$1" max_min="$2" p
+  [ -d "$base" ] || return 0
+  [ -L "$base" ] && { say "skip symlinked base: $base"; return 0; }
+  while IFS= read -r -d '' p; do
+    if ! have_busy_tools; then say "cannot prove idle (lsof missing — apt install lsof), kept: $p"; STATUS=1; continue; fi
+    if is_busy "$p"; then say "busy, kept: $p"; continue; fi
+    if [ "$DRY_RUN" -eq 1 ]; then say "would remove ($(( max_min / 60 ))h+): $p"; else rm -rf -- "$p" && say "removed ($(( max_min / 60 ))h+): $p"; fi
+  done < <(find -P "$base" -xdev -mindepth 1 -maxdepth 1 \
+      \( -name 'runner-*' -o -name 'omniroute-*' -o -name 'next-build*' -o -name 'e2e-build.tar.gz' \) \
+      ! -type l -mmin "+$max_min" -print0 2>/dev/null || true)
+}
+
+say "$(date -u +%FT%TZ) start${DRY_RUN:+ (dry-run=$DRY_RUN)} busy-tools=$(have_busy_tools && echo ok || echo MISSING)"
+
+# 1) stale artefacts — tmpfs is RAM, so it gets the short fuse
+snapshot_open_paths
+for base in $JANITOR_TMP_BASES; do sweep "$base" $(( TMPFS_MAX_AGE_HOURS * 60 )); done
+for base in $JANITOR_WORK_TEMP_BASES; do sweep "$base" $(( WORK_TEMP_MAX_AGE_HOURS * 60 )); done
+say "stale temp sweep done"
+
+# 2) zombie builds
+ZOMBIES=0
+while read -r pid etimes comm; do
+  [ -n "${pid:-}" ] || continue
+  if [ "$etimes" -gt $(( ZOMBIE_BUILD_MAX_MIN * 60 )) ]; then
+    say "⚠ zombie build pid=$pid comm=$comm age=$(( etimes / 60 ))min > ${ZOMBIE_BUILD_MAX_MIN}min — no job runs this long"
+    if [ "$DRY_RUN" -eq 1 ]; then say "[dry-run] would: kill -TERM $pid (then -KILL)"; else
+      kill -TERM "$pid" 2>/dev/null || true; sleep 10
+      kill -0 "$pid" 2>/dev/null && { kill -KILL "$pid" 2>/dev/null || true; say "  needed SIGKILL"; }
+    fi
+    ZOMBIES=$(( ZOMBIES + 1 )); STATUS=1
+  fi
+done < <(ps -eo pid=,etimes=,comm= 2>/dev/null | awk -v c="$ZOMBIE_BUILD_COMM" '$3 ~ ("^" c) {print $1, $2, $3}' || true)
+say "zombie builds: $ZOMBIES"
+
+# 3) old checkouts of STOPPED runners
+for d in $JANITOR_RUNNER_DIRS; do
+  [ -d "$d" ] && [ -f "$d/.runner" ] || continue
+  agent=$(grep -o '"agentName": *"[^"]*"' "$d/.runner" 2>/dev/null | sed 's/.*"\([^"]*\)"$/\1/')
+  [ -n "$agent" ] || continue
+  unit=$(systemctl list-units --plain --no-legend "actions.runner.*.${agent}.service" 2>/dev/null | awk 'NR==1{print $1}')
+  [ -n "$unit" ] || continue
+  if systemctl is-active --quiet "$unit"; then continue; fi
+  while IFS= read -r -d '' co; do
+    if [ "$DRY_RUN" -eq 1 ]; then say "would prune checkout of stopped runner $agent: $co"; else rm -rf -- "$co" && say "pruned checkout of stopped runner $agent: $co"; fi
+  done < <(find -P "$d/_work" -xdev -mindepth 2 -maxdepth 2 -type d -mmin "+$(( WORK_CHECKOUT_MAX_AGE_HOURS * 60 ))" -print0 2>/dev/null || true)
+done
+
+# 4a) disk
+USAGE=$(df --output=pcent "$JANITOR_DF_PATH" 2>/dev/null | tail -1 | tr -dc '0-9')
+if [ "${USAGE:-0}" -ge "$DISK_ALERT_PCT" ]; then
+  say "⚠ ROOT DISK ${USAGE}% >= ${DISK_ALERT_PCT}% — clean before the next heavy run"; STATUS=1
 else
-  echo "[janitor] disk ${USAGE}% OK"
+  say "disk ${USAGE:-?}% OK"
 fi
 
-# 3) Concurrency ceiling — 8-wide OOMed the 16 GB box twice on release day;
-#    4 is the proven ceiling. This CODIFIES the rule that was manual discipline.
+# 4b) memory pressure (PSI) — the box swapped its way through the v3.8.50 publish
+if [ -r "$JANITOR_PSI_FILE" ]; then
+  FULL60=$(awk '/^full/ {for(i=1;i<=NF;i++) if ($i ~ /^avg60=/) {sub("avg60=","",$i); print $i}}' "$JANITOR_PSI_FILE" 2>/dev/null || echo "")
+  if [ -n "$FULL60" ] && awk -v v="$FULL60" -v t="$PSI_FULL_AVG60_ALERT" 'BEGIN{exit !(v+0 >= t+0)}'; then
+    say "⚠ MEMORY PRESSURE psi full/avg60=${FULL60}% >= ${PSI_FULL_AVG60_ALERT}% — too many heavy jobs at once"; STATUS=1
+  else
+    say "memory psi full/avg60=${FULL60:-n/a}% OK"
+  fi
+fi
+
+# 4c) concurrency ceiling — alert with a breakdown; the fix is fewer/labelled
+#     runners (an operator decision), not killing listeners from cron.
 ACTIVE=$(pgrep -fc "Runner.Listener" || true)
+OMNI=$(pgrep -fc "actions-runner-omniroute[^ ]*/bin[^ ]*/Runner.Listener" || true)
 if [ "${ACTIVE:-0}" -gt "$MAX_ACTIVE_RUNNERS" ]; then
-  echo "[janitor] ⚠ ${ACTIVE} Runner.Listener processes > ceiling ${MAX_ACTIVE_RUNNERS} — stop the extra runners (systemctl stop actions.runner.<name>)"
+  say "⚠ ${ACTIVE} Runner.Listener processes (omniroute=${OMNI:-0}, other=$(( ${ACTIVE:-0} - ${OMNI:-0} ))) > ceiling ${MAX_ACTIVE_RUNNERS} — stop idle extras: systemctl stop <unit> only when it has no Runner.Worker child"
   STATUS=1
 else
-  echo "[janitor] runners active: ${ACTIVE:-0}/${MAX_ACTIVE_RUNNERS} OK"
+  say "runners active: ${ACTIVE:-0}/${MAX_ACTIVE_RUNNERS} (omniroute=${OMNI:-0}) OK"
 fi
 
-echo "[janitor] done status=$STATUS"
+say "done status=$STATUS"
 exit "$STATUS"

tests/unit/alibaba-free-tier-allowlist.test.ts

@@ -7,6 +7,9 @@
  */
 import { test } from "node:test";
 import assert from "node:assert/strict";
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 import {
   ALIBABA_FREE_TIER_TEXT_CAPABLE_MODELS,
   ALIBABA_NO_FREE_TIER_TEXT_MODELS,
@@ -27,18 +30,90 @@ test("built-in allowlist includes operator free models and excludes paid blockli
   assert.equal(isAlibabaBuiltinFreeTierTextModel("qwen3.7-max"), false);
 });
 
-test("allowlist JSON pack overrides embedded lists when valid", () => {
+/**
+ * The shipped `config/alibaba-free-tier-allowlist.json` carries a `validUntil`,
+ * so asserting against it made this test a time bomb: it went red on its own on
+ * 2026-08-28, the day after the pack expired, and stayed red on every PR and on
+ * `main` (#11866). Nothing had changed — the clock moved.
+ *
+ * Production was never affected: an expired pack falls back to the embedded
+ * list by design. So the contract worth pinning is the BEHAVIOR on both sides of
+ * the expiry, with packs this test owns and dates it controls — never the
+ * freshness of the catalog that ships in the repo.
+ */
+function withAllowlistPack(
+  pack: Record<string, unknown>,
+  assertions: () => void
+): void {
+  const dir = mkdtempSync(join(tmpdir(), "alibaba-allowlist-"));
+  const packPath = join(dir, "allowlist.json");
+  writeFileSync(packPath, JSON.stringify(pack), "utf8");
+
   const previousPath = process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH;
-  const packPath = `${process.cwd()}/config/alibaba-free-tier-allowlist.json`;
   process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH = packPath;
   resetAlibabaFreeTierAllowlistCache();
+  try {
+    assertions();
+  } finally {
+    if (previousPath) process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH = previousPath;
+    else delete process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH;
+    resetAlibabaFreeTierAllowlistCache();
+    rmSync(dir, { recursive: true, force: true });
+  }
+}
 
-  const pack = loadAlibabaFreeTierAllowlistPack();
-  assert.ok(pack);
-  assert.ok(isAlibabaFreeTierAllowlistPackValid(pack!));
-  assert.ok(pack!.capable.includes("qwen3.6-plus"));
+test("allowlist JSON pack overrides embedded lists while it is still valid", () => {
+  withAllowlistPack(
+    {
+      asOf: "2026-07-28",
+      validUntil: "2999-01-01",
+      capable: ["pack-only-capable-model", "qwen3.6-plus"],
+      noFreeTier: ["pack-only-paid-model"],
+    },
+    () => {
+      const pack = loadAlibabaFreeTierAllowlistPack();
+      assert.ok(pack, "a pack inside its validity window must load");
+      assert.ok(isAlibabaFreeTierAllowlistPackValid(pack!));
+      assert.ok(pack!.capable.includes("qwen3.6-plus"));
+      // Positive anchor: the pack must actually REPLACE the embedded list, not
+      // merely load. `pack-only-capable-model` exists nowhere else.
+      assert.equal(isAlibabaBuiltinFreeTierTextModel("pack-only-capable-model"), true);
+      assert.equal(isAlibabaBuiltinNoFreeTierTextModel("pack-only-paid-model"), true);
+    }
+  );
+});
 
-  if (previousPath) process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH = previousPath;
-  else delete process.env.ALIBABA_FREE_TIER_ALLOWLIST_PATH;
-  resetAlibabaFreeTierAllowlistCache();
+test("an expired allowlist pack is ignored and the embedded list serves instead", () => {
+  // This is the path production has actually been on since 2026-08-27, and it
+  // had no coverage at all — which is why the expiry surfaced as a red test
+  // rather than as a deliberate, understood fallback.
+  withAllowlistPack(
+    {
+      asOf: "2026-07-28",
+      validUntil: "2026-08-27",
+      capable: ["pack-only-capable-model"],
+      noFreeTier: ["pack-only-paid-model"],
+    },
+    () => {
+      assert.equal(loadAlibabaFreeTierAllowlistPack(), null, "expired pack must not load");
+      assert.equal(isAlibabaBuiltinFreeTierTextModel("pack-only-capable-model"), false);
+      // The embedded list must be what answers once the pack is rejected.
+      assert.equal(isAlibabaBuiltinFreeTierTextModel("qwen3.6-plus"), true);
+      assert.equal(isAlibabaBuiltinNoFreeTierTextModel("qwen3.7-max"), true);
+    }
+  );
+});
+
+test("isAlibabaFreeTierAllowlistPackValid compares against the instant it is given", () => {
+  const pack = { asOf: "2026-07-28", validUntil: "2026-08-27", capable: ["x"], noFreeTier: [] };
+  assert.equal(isAlibabaFreeTierAllowlistPackValid(pack, Date.parse("2026-08-26")), true);
+  assert.equal(isAlibabaFreeTierAllowlistPackValid(pack, Date.parse("2026-08-28")), false);
+  // No expiry declared means the pack never goes stale on its own.
+  assert.equal(
+    isAlibabaFreeTierAllowlistPackValid(
+      { asOf: "2026-07-28", capable: ["x"], noFreeTier: [] },
+      Date.parse("2999-01-01")
+    ),
+    true
+  );
 });

tests/unit/check-workflows-provenance-runner.test.ts

@@ -0,0 +1,145 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { readFileSync, readdirSync } from "node:fs";
+import { join } from "node:path";
+
+import {
+  classifyRunsOn,
+  findProvenanceOnSelfHosted,
+} from "../../scripts/check/lib/provenanceRunner.mjs";
+
+/**
+ * v3.8.50, 10th publish attempt, 76 minutes in — after the tag, the GitHub
+ * Release and the Docker images were already public:
+ *
+ *   422 Unprocessable Entity - Error verifying sigstore provenance bundle:
+ *   Unsupported GitHub Actions runner environment: "self-hosted".
+ *
+ * `USE_VPS_RUNNER` had routed the publish job to the .113 pool on 2026-08-02;
+ * no release happened between 07-30 and 08-28, so nothing surfaced it. The
+ * pairing is pure text, so it must fail the workflow lint on the PR that
+ * introduces it.
+ */
+const ROOT = join(import.meta.dirname, "../..");
+const WORKFLOWS = join(ROOT, ".github/workflows");
+
+// The exact runs-on expression npm-publish.yml used when it broke.
+const VPS_EXPR =
+  "${{ (vars.USE_VPS_RUNNER == 'true' && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository)) && fromJSON('[\"self-hosted\",\"omni-release\"]') || 'ubuntu-latest' }}";
+
+function workflow(runsOn: string, run: string, extra = ""): string {
+  return [
+    "name: t",
+    "on: push",
+    "jobs:",
+    "  publish:",
+    `    runs-on: ${runsOn}`,
+    extra,
+    "    steps:",
+    "      - name: upload",
+    `        run: ${run}`,
+    "",
+  ].join("\n");
+}
+
+test("classifyRunsOn: literal, array, object-with-labels and the fromJSON expression are self-hosted", () => {
+  assert.equal(classifyRunsOn("self-hosted"), "self-hosted");
+  assert.equal(classifyRunsOn(["self-hosted", "omni-release"]), "self-hosted");
+  assert.equal(classifyRunsOn({ group: "Default", labels: ["self-hosted"] }), "self-hosted");
+  assert.equal(classifyRunsOn(VPS_EXPR), "self-hosted");
+});
+
+test("classifyRunsOn: hosted labels are hosted, opaque expressions are unknown (never guessed)", () => {
+  assert.equal(classifyRunsOn("ubuntu-latest"), "hosted");
+  assert.equal(classifyRunsOn(["ubuntu-latest"]), "hosted");
+  assert.equal(classifyRunsOn("${{ matrix.os }}"), "unknown");
+  assert.equal(classifyRunsOn(undefined), "unknown");
+});
+
+test("flags --provenance inside a job routed to the self-hosted pool", () => {
+  const found = findProvenanceOnSelfHosted(
+    workflow(
+      JSON.stringify(VPS_EXPR),
+      'npm stage publish --provenance --access public --tag "$TAG"'
+    ),
+    "npm-publish.yml"
+  );
+  assert.deepEqual(found, [{ file: "npm-publish.yml", job: "publish", step: "upload" }]);
+});
+
+test("also catches the literal label and the --provenance-file form", () => {
+  assert.equal(
+    findProvenanceOnSelfHosted(workflow("self-hosted", "npm publish --provenance")).length,
+    1
+  );
+  assert.equal(
+    findProvenanceOnSelfHosted(
+      workflow("[self-hosted, omni-release]", "npm publish --provenance-file=./p.json")
+    ).length,
+    0,
+    "--provenance-file is a different flag (a pre-built bundle) and is not what the registry rejects"
+  );
+  assert.equal(
+    findProvenanceOnSelfHosted(workflow("self-hosted", "npm publish --provenance=true")).length,
+    1
+  );
+});
+
+test("does not flag hosted jobs, unknown runners, or self-hosted jobs without the flag", () => {
+  assert.deepEqual(
+    findProvenanceOnSelfHosted(workflow("ubuntu-latest", "npm publish --provenance")),
+    []
+  );
+  assert.deepEqual(
+    findProvenanceOnSelfHosted(workflow("${{ matrix.os }}", "npm publish --provenance")),
+    []
+  );
+  assert.deepEqual(
+    findProvenanceOnSelfHosted(workflow("self-hosted", "npm publish --access public")),
+    []
+  );
+  // The word only in a step NAME or a comment is not a finding.
+  assert.deepEqual(
+    findProvenanceOnSelfHosted(
+      [
+        "name: t",
+        "on: push",
+        "jobs:",
+        "  j:",
+        "    runs-on: self-hosted",
+        "    steps:",
+        "      - name: provenance note",
+        "        run: echo hi # --provenance later",
+        "",
+      ].join("\n")
+    ),
+    [],
+    "a comment after the command is still part of the run string — accept that the regex is conservative"
+  );
+});
+
+test("reusable-workflow jobs (uses:) and unparseable YAML are not this rule's findings", () => {
+  const reusable = [
+    "name: t",
+    "on: push",
+    "jobs:",
+    "  j:",
+    "    uses: ./.github/workflows/x.yml",
+    "",
+  ].join("\n");
+  assert.deepEqual(findProvenanceOnSelfHosted(reusable), []);
+  assert.deepEqual(findProvenanceOnSelfHosted("jobs: [unclosed"), []);
+});
+
+test("regression guard: no workflow in this repo publishes with --provenance from a self-hosted runner", () => {
+  const files = readdirSync(WORKFLOWS).filter((f) => /\.ya?ml$/.test(f));
+  assert.ok(files.length > 10, "expected the real workflow set");
+  const findings = files.flatMap((f) =>
+    findProvenanceOnSelfHosted(readFileSync(join(WORKFLOWS, f), "utf8"), f)
+  );
+  assert.deepEqual(
+    findings,
+    [],
+    `npm rejects provenance from self-hosted runners (422) — move the upload to a github-hosted job: ${JSON.stringify(findings)}`
+  );
+});

tests/unit/config-expiry-time-bomb.test.ts

@@ -0,0 +1,135 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+import {
+  classifyExpiry,
+  collectExpiryFields,
+  scanConfigExpiry,
+} from "../../scripts/check/lib/configExpiry.mjs";
+
+/**
+ * Time bombs: config packs with a `validUntil` (or sibling key) that lapse with
+ * no commit involved. The Alibaba free-tier pack expired on 2026-08-27 and from
+ * the 28th every PR and main carried a red Unit Tests shard (#11866). Nothing a
+ * diff review could have caught.
+ *
+ * This suite fails SEVEN DAYS BEFORE any pack under config/ lapses, naming the
+ * file and key, so renewal happens on someone's terms instead of the clock's.
+ */
+const ROOT = join(import.meta.dirname, "../..");
+const CONFIG_DIR = join(ROOT, "config");
+const DAY = 86_400_000;
+const WARN_DAYS = 7;
+
+/**
+ * Packs known to be expired/expiring, each pinned to the issue that owns the
+ * renewal decision. An entry whose pack is no longer expiring FAILS below as a
+ * stale allowlist entry — remove it when the pack is renewed.
+ */
+const ALLOWLIST: Record<string, string> = {
+  "alibaba-free-tier-allowlist.json":
+    "#11866 — validUntil 2026-08-27 has passed; the loader already falls back to the embedded list, and renewing the curated free-tier pack is an operator data decision, not a test fix",
+};
+
+const NOW = Date.UTC(2026, 7, 28); // 2026-08-28, fixed: this suite must not itself depend on the clock
+const day = (offset: number) => new Date(NOW + offset * DAY).toISOString().slice(0, 10);
+
+test("collectExpiryFields: finds nested and array-nested expiry keys, ignores non-string values", () => {
+  const fields = collectExpiryFields(
+    {
+      validUntil: day(3),
+      nested: { expiresAt: day(30), other: "x" },
+      list: [{ expiry: day(-1) }, { expires: 12345 }],
+      expires_at: "not a date",
+    },
+    "pack.json"
+  );
+  assert.deepEqual(
+    fields.map((f) => [f.keyPath, f.expiresAt === null ? null : "date"]),
+    [
+      ["validUntil", "date"],
+      ["nested.expiresAt", "date"],
+      ["list.0.expiry", "date"],
+      ["expires_at", null],
+    ]
+  );
+});
+
+test("classifyExpiry: expired / expiring inside the warning window / ok / unparseable", () => {
+  const f = (raw: string) => ({
+    file: "p",
+    keyPath: "validUntil",
+    raw,
+    expiresAt: Number.isFinite(Date.parse(raw)) ? Date.parse(raw) : null,
+  });
+  assert.equal(classifyExpiry(f(day(-1)), NOW, WARN_DAYS), "expired");
+  assert.equal(
+    classifyExpiry(f(day(0)), NOW, WARN_DAYS),
+    "expiring",
+    "lapsing today is already too late to be 'ok'"
+  );
+  assert.equal(classifyExpiry(f(day(6)), NOW, WARN_DAYS), "expiring");
+  assert.equal(classifyExpiry(f(day(8)), NOW, WARN_DAYS), "ok");
+  assert.equal(classifyExpiry(f("never"), NOW, WARN_DAYS), "unparseable");
+});
+
+test("scanConfigExpiry: walks a config tree, skips node_modules and invalid JSON, keys files portably", () => {
+  const dir = mkdtempSync(join(tmpdir(), "cfg-expiry-"));
+  try {
+    mkdirSync(join(dir, "sub"), { recursive: true });
+    mkdirSync(join(dir, "node_modules", "dep"), { recursive: true });
+    writeFileSync(join(dir, "a.json"), JSON.stringify({ validUntil: day(3) }));
+    writeFileSync(join(dir, "sub", "b.json"), JSON.stringify({ deep: { expiresAt: day(40) } }));
+    writeFileSync(
+      join(dir, "node_modules", "dep", "c.json"),
+      JSON.stringify({ validUntil: day(-5) })
+    );
+    writeFileSync(join(dir, "broken.json"), "{ not json");
+    writeFileSync(join(dir, "notes.txt"), JSON.stringify({ validUntil: day(-5) }));
+    const found = scanConfigExpiry(dir).map((f) => `${f.file}:${f.keyPath}`);
+    assert.deepEqual(found, ["a.json:validUntil", "sub/b.json:deep.expiresAt"]);
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test(`repo: no pack under config/ lapses within ${WARN_DAYS} days unless its renewal is tracked`, (t) => {
+  const fields = scanConfigExpiry(CONFIG_DIR);
+  // Positive anchor: the scanner must be seeing SOMETHING, or a renamed key
+  // would silently turn this whole suite into a no-op.
+  assert.ok(
+    fields.length >= 1,
+    "expected at least one dated pack under config/ (the Alibaba allowlist) — if the key was renamed, extend EXPIRY_KEY"
+  );
+
+  const failures: string[] = [];
+  const seenAllowlisted = new Set<string>();
+  for (const f of fields) {
+    const status = classifyExpiry(f, Date.now(), WARN_DAYS);
+    const tracked = ALLOWLIST[f.file];
+    if (status === "unparseable") {
+      t.diagnostic(`${f.file} ${f.keyPath}="${f.raw}" is not a date — not monitored`);
+      continue;
+    }
+    if (status === "ok") continue;
+    if (tracked) {
+      seenAllowlisted.add(f.file);
+      t.diagnostic(`${f.file} ${f.keyPath}=${f.raw} is ${status} — tracked: ${tracked}`);
+      continue;
+    }
+    failures.push(
+      `${f.file} → ${f.keyPath}=${f.raw} is ${status}: renew the pack (or track it in ALLOWLIST with its issue)`
+    );
+  }
+  for (const file of Object.keys(ALLOWLIST)) {
+    if (!seenAllowlisted.has(file)) {
+      failures.push(
+        `stale ALLOWLIST entry: ${file} is no longer expired/expiring — remove it (${ALLOWLIST[file]})`
+      );
+    }
+  }
+  assert.deepEqual(failures, [], failures.join("\n"));
+});

tests/unit/runner-janitor.test.ts

@@ -0,0 +1,196 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import { spawnSync } from "node:child_process";
+import {
+  existsSync,
+  mkdirSync,
+  mkdtempSync,
+  readFileSync,
+  rmSync,
+  statSync,
+  utimesSync,
+  writeFileSync,
+} from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+/**
+ * scripts/ops/runner-janitor.sh runs from cron on the .113 runner box. This
+ * suite pins its safety contract against a fixture tree — never the real /tmp:
+ * every base, the runner dirs, the PSI file and the df path are redirected, the
+ * zombie pattern is set to a name no process has, and the ceilings are lifted
+ * so the outcome does not depend on the box the test happens to run on.
+ */
+const ROOT = path.resolve(import.meta.dirname, "..", "..");
+const SCRIPT = path.join(ROOT, "scripts", "ops", "runner-janitor.sh");
+const HOUR = 3_600_000;
+// The sweep needs lsof to PROVE a path is idle (one snapshot of open paths). Hosted CI
+// images ship both; a bare devbox may not. Each branch below asserts what must
+// hold in that environment — without the tools the contract is "delete nothing,
+// say why", which is exactly the behaviour worth pinning.
+const HAVE_BUSY_TOOLS =
+  spawnSync("bash", ["-c", "command -v lsof"], { stdio: "ignore" }).status === 0;
+
+function fixture() {
+  const base = mkdtempSync(path.join(os.tmpdir(), "janitor-fixture-"));
+  const old = new Date(Date.now() - 5 * HOUR);
+  const mk = (name: string, dir: boolean, when: Date | null) => {
+    const p = path.join(base, name);
+    if (dir) {
+      mkdirSync(p);
+      writeFileSync(path.join(p, "x"), "x");
+    } else writeFileSync(p, "x");
+    if (when) utimesSync(p, when, when);
+    return p;
+  };
+  return {
+    base,
+    staleTar: mk("e2e-build.tar.gz", false, old), // fixed-name artefact ci.yml/npm-publish leave behind
+    staleBuild: mk("next-build-abc", true, old),
+    staleUpgrade: mk("omniroute-install-upgrade-xyz", true, old),
+    fresh: mk("omniroute-batch-api-fresh", true, null), // in use right now
+    unrelated: mk("somebody-elses.log", false, old), // not ours — never touched
+  };
+}
+
+function run(args: string[], base: string, extraEnv: Record<string, string> = {}) {
+  return spawnSync("bash", [SCRIPT, ...args], {
+    encoding: "utf8",
+    stdio: ["ignore", "pipe", "pipe"],
+    env: {
+      ...process.env,
+      JANITOR_TMP_BASES: base,
+      JANITOR_WORK_TEMP_BASES: "",
+      JANITOR_RUNNER_DIRS: path.join(base, "no-runners-here-*"),
+      JANITOR_PSI_FILE: path.join(base, "no-psi"),
+      JANITOR_DF_PATH: base,
+      ZOMBIE_BUILD_COMM: "janitor-test-no-such-process",
+      MAX_ACTIVE_RUNNERS: "9999",
+      DISK_ALERT_PCT: "101",
+      ...extraEnv,
+    },
+  });
+}
+
+describe("runner-janitor.sh", () => {
+  it("is executable bash with strict mode and prints usage on --help", () => {
+    assert.ok(existsSync(SCRIPT));
+    assert.ok(statSync(SCRIPT).mode & 0o111, "must be chmod +x (cron runs it directly)");
+    const body = readFileSync(SCRIPT, "utf8");
+    assert.ok(body.startsWith("#!/usr/bin/env bash"));
+    assert.ok(body.includes("set -euo pipefail"));
+    const help = run(["--help"], os.tmpdir());
+    assert.equal(help.status, 0, help.stderr);
+    assert.match(help.stdout, /--dry-run/);
+  });
+
+  it("without lsof it cannot prove idle, so it deletes nothing and says why (exit 1)", () => {
+    const f = fixture();
+    try {
+      const r = run([], f.base, { JANITOR_LSOF: "/nonexistent/lsof" });
+      assert.equal(r.status, 1, "a janitor that cannot do its job must show up in the cron log");
+      assert.match(r.stdout, /busy-tools=MISSING/);
+      assert.match(
+        r.stdout,
+        /cannot prove idle \(lsof missing — apt install lsof\), kept: .*e2e-build\.tar\.gz/
+      );
+      for (const p of [f.staleTar, f.staleBuild, f.staleUpgrade, f.fresh, f.unrelated]) {
+        assert.ok(existsSync(p), `must not delete ${p} when idleness cannot be proven`);
+      }
+    } finally {
+      rmSync(f.base, { recursive: true, force: true });
+    }
+  });
+
+  it("--dry-run names what it WOULD remove and removes nothing", (t) => {
+    if (!HAVE_BUSY_TOOLS) return t.skip("lsof absent on this box — sweep branch covered in CI");
+    const f = fixture();
+    try {
+      const r = run(["--dry-run"], f.base);
+      assert.equal(r.status, 0, r.stderr + r.stdout);
+      assert.match(r.stdout, /busy-tools=ok/);
+      for (const p of [f.staleTar, f.staleBuild, f.staleUpgrade]) {
+        assert.match(
+          r.stdout,
+          new RegExp(`would remove \\(3h\\+\\): ${p.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}`)
+        );
+        assert.doesNotMatch(
+          r.stdout,
+          new RegExp(`removed \\(3h\\+\\): ${p.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}`),
+          "dry-run must never claim it removed something"
+        );
+        assert.ok(existsSync(p), `dry-run must not delete ${p}`);
+      }
+      assert.doesNotMatch(
+        r.stdout,
+        /omniroute-batch-api-fresh/,
+        "a fresh dir is never a candidate"
+      );
+      assert.doesNotMatch(r.stdout, /somebody-elses\.log/, "only names our tooling creates");
+      assert.match(r.stdout, /zombie builds: 0/);
+      assert.match(r.stdout, /done status=0/);
+    } finally {
+      rmSync(f.base, { recursive: true, force: true });
+    }
+  });
+
+  it("for real: sweeps the three stale artefacts, keeps the fresh one and the stranger", (t) => {
+    if (!HAVE_BUSY_TOOLS) return t.skip("lsof absent on this box — sweep branch covered in CI");
+    const f = fixture();
+    try {
+      const r = run([], f.base);
+      assert.equal(r.status, 0, r.stderr + r.stdout);
+      assert.ok(!existsSync(f.staleTar), "stale e2e-build.tar.gz must go (it is RAM on tmpfs)");
+      assert.ok(!existsSync(f.staleBuild), "stale next-build dir must go");
+      assert.ok(!existsSync(f.staleUpgrade), "stale install-upgrade dir must go");
+      assert.ok(existsSync(f.fresh), "a fresh dir must survive");
+      assert.ok(existsSync(f.unrelated), "files we did not create must survive even when old");
+    } finally {
+      rmSync(f.base, { recursive: true, force: true });
+    }
+  });
+
+  it("tmpfs fuse is shorter than the disk fuse (RAM vs disk), both overridable", () => {
+    const f = fixture();
+    try {
+      // With a 6h tmpfs fuse the 5h-old artefacts are NOT stale yet.
+      const r = run(["--dry-run"], f.base, { TMPFS_MAX_AGE_HOURS: "6" });
+      assert.doesNotMatch(
+        r.stdout,
+        /would remove|removed \(|cannot prove idle/,
+        "nothing is stale under a 6h fuse, so no candidate is even examined"
+      );
+      const body = readFileSync(SCRIPT, "utf8");
+      assert.match(body, /TMPFS_MAX_AGE_HOURS:-3\}/, "tmpfs default must stay short — it is RAM");
+      assert.match(body, /WORK_TEMP_MAX_AGE_HOURS:-24\}/);
+    } finally {
+      rmSync(f.base, { recursive: true, force: true });
+    }
+  });
+
+  it("alerts (exit 1) on disk and memory pressure thresholds without touching files", () => {
+    const f = fixture();
+    try {
+      writeFileSync(
+        path.join(f.base, "psi"),
+        "some avg10=0.00 avg60=0.00 avg300=0.00 total=1\nfull avg10=0.00 avg60=23.50 avg300=9.00 total=1\n"
+      );
+      const r = run(["--dry-run"], f.base, {
+        JANITOR_PSI_FILE: path.join(f.base, "psi"),
+        DISK_ALERT_PCT: "0",
+      });
+      assert.equal(r.status, 1, "attention needed must be exit 1 for the cron log");
+      assert.match(r.stdout, /MEMORY PRESSURE psi full\/avg60=23\.50%/);
+      assert.ok(existsSync(f.fresh) && existsSync(f.unrelated));
+      assert.match(r.stdout, /ROOT DISK \d+% >= 0%/);
+      assert.ok(existsSync(f.staleTar), "alerting never deletes");
+    } finally {
+      rmSync(f.base, { recursive: true, force: true });
+    }
+  });
+
+  it("rejects unknown arguments instead of silently running", () => {
+    const r = run(["--yolo"], os.tmpdir());
+    assert.equal(r.status, 2);
+  });
+});