docs: add backlog items for site PR validation and frontend js-yaml overrides

site/** pull requests get no build validation: ci-site.yml declares a
pull_request trigger but gates its only build job to master and tags, so
a bad site lockfile first fails on the post-merge run that deploys.

frontend pnpm override floors still admit js-yaml 3.15.0 and 5.2.0,
leaving three open advisories including the one PR 2141 closed for site/.

docs/backlog/frontend-js-yaml-overrides-stale.md

@@ -0,0 +1,37 @@
+---
+worth: yes
+where: frontend/pnpm-lock.yaml:61
+added: 2026-08-11
+---
+# frontend pnpm overrides still admit vulnerable js-yaml
+
+Three open Dependabot alerts against `frontend/pnpm-lock.yaml` survive because the pnpm override
+floors in that file are set below the patched versions.
+
+| override (line) | current floor | resolved version | advisory | patched in |
+|---|---|---|---|---|
+| `js-yaml@>=3.0.0 <4.0.0` (:61) | `>=3.14.2` | 3.15.0 (:4798, :12669) | GHSA-5p4m-2wfm-xmqj (high) | 3.15.1 |
+| `js-yaml@>=4.0.0 <5.0.0` (:62) | `>=5.0.0` | 5.2.0 (:4802, :12674) | GHSA-pm4m-ph32-ghv5 (high) | 5.2.2 |
+| | | | GHSA-724g-mxrg-4qvm (medium) | 5.2.1 |
+
+GHSA-5p4m-2wfm-xmqj is the same advisory PR #2141 closed for `site/`, patched there by moving to
+js-yaml 3.15.1. It remains open against the frontend lockfile, which is still on 3.15.0. Verified
+against the repository's Dependabot alerts, not inferred from version numbers.
+
+`.github/dependabot.yml` sets `open-pull-requests-limit: 0` on every npm directory, so no version
+update PR will ever reach these. Security updates bypass that limit, which is why #2141 existed at
+all, but the override floors pin the resolution regardless of what the bot proposes.
+
+Fix: raise both floors and re-lock.
+
+```yaml
+js-yaml@>=3.0.0 <4.0.0: '>=3.15.1 <4.0.0'
+js-yaml@>=4.0.0 <5.0.0: '>=5.2.2 <6.0.0'
+```
+
+Both js-yaml lines are build and test time only in the frontend (3.x arrives via
+`@istanbuljs/load-nyc-config`), so this is alert hygiene rather than a shipped vulnerability.
+
+Separately, and not covered here: `frontend/pnpm-lock.yaml` carries a wider queue of open alerts
+(brace-expansion, undici, fast-uri, postcss, svgo, webpack-dev-server, body-parser). Those were not
+audited and may or may not be reachable.

docs/backlog/site-pr-build-never-runs.md

@@ -0,0 +1,35 @@
+---
+worth: yes
+where: .github/workflows/ci-site.yml:24
+added: 2026-08-11
+---
+# site/** pull requests get no build validation
+
+`ci-site.yml` declares a `pull_request` trigger on `site/**` (lines 12-15), but the only job that
+installs and builds is gated:
+
+```yaml
+if: github.ref == 'refs/heads/master' || startsWith(github.ref, 'refs/tags/')
+```
+
+On a pull request `github.ref` is `refs/pull/N/merge`, so `build` skips, and `merge` (`needs: build`)
+and `deploy` skip with it. The trigger is dead: it produces a run in which every job is skipped, which
+renders in the checks list the same way a pass does. No required status checks are configured on
+master either, so nothing else catches it.
+
+Consequence: a `site/yarn.lock` that fails `yarn --frozen-lockfile` first fails on the master run
+after merge, which is the same run that rebuilds and deploys remark42.com. `site/Dockerfile:5-6` is
+the only place the lockfile is exercised, and it runs post-merge.
+
+Fix: add a PR-only job mirroring `.github/workflows/ci-build.yml`'s existing pattern, using
+`docker/build-push-action` with `context: ./site`, `load: true`, no `outputs:`, and no
+`docker/login-action`.
+
+Constraint any fix must preserve: the current gate exists because `build` pushes to ghcr using
+`secrets.PKG_TOKEN`, which must not run for pull requests, since fork PRs receive no secrets.
+Relaxing the `if` on its own is not sufficient and would break fork PRs.
+
+Surfaced while reviewing PR #2141 (a dependabot js-yaml lockfile bump), where all three site jobs
+reported as skipped and nothing verified the lockfile before merge. `site/**` PRs are mostly
+generated lockfile bumps, which is exactly the class a frozen-lockfile install catches and a human
+reviewer cannot.