fix(e2e): exclude test fixtures from @parity content crawls + robust ephemeral loop timing (#1771 follow-up)

[Adam-S-Daniel]

Why the loops stayed red (the vicious cycle + #1723 Cat-2 regression)

The #1771 step-4 ephemeral prod loops (cms-publish-loop-prod-mutate, cms-media-roundtrip) were still failing in real prod after #1797/#1801/#1804. The proven root cause, reproduced locally:

The ephemeral posts are created through the genuinely-UI-driven Decap "+ New Post" leg. Decap writes only the fields the posts collection declares (admin/config*.yml), which has no sitemap/robots field and whose test_fixture is a widget: hidden, default: false the editor can't toggle. So the post that actually lands on main carries published: true + a future date + test_fixture: false and no sitemap/robots keys — verified against the real Create Post commit:

---
title: E2E Media Roundtrip 1779924060949
slug: e2e-media-roundtrip-1779924060949
date: 2099-12-31 00:00:00 +0000
featured_image: /assets/images/uploads/e2e-media-roundtrip-1779924060949.png
published: true
test_fixture: false      # ← NOT true
---

It therefore looks exactly like real public content. The @parity content-crawl specs enumerate every published /blog/ post:

• console-clean.spec.js runs in BOTH the required e2e-admin (@admin-read) and parity checks;
• image-alt-text.spec.js crawls every sitemap URL (the post lands in the sitemap — no sitemap: false).

When a crashed run leaves an orphan on main (a post whose uploaded image was already deleted), /blog/<slug>/ 404s a resource → console-clean/image-alt-text red-fail → every cms PR is BLOCKED, including the loop's OWN create PR (verified: create PR #1808 BLOCKED on e2e-admin/parity/finalize failing on /blog/e2e-media-roundtrip-…/). The create PR never merges → the post never serves → the loop's serve leg times out → the run fails → leaves another orphan → repeat. This is the exact #1723 Category-2 class (transient main state poisoning a shared required check), reintroduced by the ephemeral redesign.

Crucially, the design's intended test_fixture: true / sitemap: false markers exist only in the composePost afterAll fallback — they were never on the UI-created posts — so a test_fixture-only exclusion would not have worked.

Fix A (primary — breaks the cycle): public-content crawls exclude test fixtures

A single shared, well-named enumeration predicate, e2e/public-content.js#isTestFixturePost, is now the source of truth used by console-clean, image-alt-text, and sitemap. It excludes a post on ANY of:

1. test_fixture: true (the documented marker — _e2e/unpublish canaries, composePost fallback),
2. sitemap: false (fixtures opt out of the sitemap),
3. the structural e2e- slug signature (/^\d{4}-\d{2}-\d{2}-e2e-/ filename, ^e2e- URL slug — already the codebase-wide fixture detector in admin/posts-list-enhance.js).

Signal 3 is load-bearing: it catches the UI-created posts that carry neither flag, because the slug:/date: the spec types are reliably present. Real public posts (even one titled "E2E Testing Guide" with a non-e2e- slug) are not excluded.

Locked by e2e/public-content.test.js, including the exact UI-shaped no-flag post and a console-clean-style end-to-end enumeration.

Local proof (negative control): with the exclusion temporarily disabled, both console-clean (a 5th test for the canary URL → console.error: Failed to load resource … 404) and image-alt-text (<img> with no alt on the canary) FAIL on a UI-shaped canary that is in the built sitemap and renders at /blog/…/. With the exclusion on, the canary URL is never enumerated and both stay green.

Fix B: robust ephemeral create→serve→delete→404 timing

• reopenForPublishedDelete 6 → 13 min (it timed out waiting for Decap to re-sync to the published file after the create PR's squash auto-merge + branch removal; under contention 6 min is too tight).
• create-PR waitForMerge 5 → 10 min (absorbs GitHub API lag between deploy completing and the PR flipping merged:true; the merge has normally already landed by the URL-serve gate).
• TEST_TIMEOUT_MS raised to fit the worst-case leg sum within the job timeouts: prod 40 → 46 min (50-min job), media 55 → 80 min (95-min job).

The loop still genuinely exercises the real Decap UI create→publish→serve→UI-delete→404 chain; the afterAll existence-only API safety-net is unchanged.

Fix C: harden orphan sweeping

sweep-stale-cms-prs.yml gains a dedicated ephemeral_threshold_hours (default 3h) for the ephemeral prod-loop orphan tier (vs 6h for human-hostable tiers) plus a 3-hourly cron, so a killed run's orphan is gone within ~3-6h. 3h is deliberately above the longest run's 95-min job timeout so the sweep can never reap a post still under active test. With Fix A this is defence-in-depth (orphans no longer poison checks) + tidiness.

Docs / lint plumbing

• Reconciled the inaccurate "born with test_fixture: true" claims in AGENTS.md, prod-mutate-fixture.js, and cms-posts-list-enhance.spec.js to reflect what the UI actually writes; documented the Cat-2 regression + the public-content crawl scope.
• select-specs.js wires e2e/public-content.js as a dependency of the three crawl specs (and the parity-preview selector) + maps the new unit test.

Validation

• npm run lint (0 errors) + npm run format:check clean.
• All constraint lints green: silent-catch-lint, select-specs, cms-recursion-churn, cms-editor-ui, fixture-baseline, canary-content, workflow-prod-loop-serialized, plus the new public-content + cms-posts-list-enhance + prod-mutate-fixture.
• Both ephemeral specs parse/list under --project chromium-desktop-3k --list.
• Local TARGET=local runs of console-clean / image-alt-text / sitemap / draft-isolation pass with AND without a UI-shaped canary present.

🤖 Generated with Claude Code

[github-actions]

🗑️ Preview environment cleaned up.