refactor(test): derive test_seo sample from prerender + harden deploy smoke test

[miquelmatoses]

Summary

Two reliability follow-ups from the 2026-05-26 audit.

1. test_seo.py sample derived from the prerender (no hardcoded slugs)

The audit found SAMPLED_ROUTES contained three blog slugs that no longer existed. CI never caught it because these tests skipif when dist/ is absent and CI does not run build:full on the backend job. PR #36 swapped the dead slugs by hand, but the fragile pattern remained.

Now the sample is derived from the prerendered dist/ at import time:

def _discover_routes() -> list[str]:
    if not _has_prerendered():
        return []
    routes = [r for r in TOP_LEVEL_CANDIDATES if _html_path(r).is_file()]
    for lang in LANGS:
        blog_dir = DIST / "blog" if lang == "" else DIST / lang / "blog"
        slugs = sorted(p.name for p in blog_dir.iterdir()
                       if p.is_dir() and (p / "index.html").is_file())
        if slugs:
            prefix = "blog" if lang == "" else f"{lang}/blog"
            routes.append(f"{prefix}/{slugs[0]}")
    return routes

• Deterministic (alphabetical, first per category) so it fails identically for everyone.
• Degrades gracefully: empty when dist/ is not built (skipif still fires), individual top-level pages skipped if missing instead of a hard FileNotFoundError.
• Coverage 7 -> 13 routes (one blog article per language present).

Robustness gate (verified locally): removing the currently-sampled article from dist/ keeps the suite green, because the sample re-derives to the next existing article.

$ mv dist/blog/anonymity-...-why-it-matters /tmp/   # simulate vanished slug
$ pytest -q api/tests/test_seo.py
57 passed   # sample auto-advanced to big-five-personality-across-cultures-...

2. Drift inventory (FASE 2)

Location	Finding	Action
api/tests/test_seo.py	3 dead hardcoded blog slugs	Fixed — dynamic derivation
api/jobs/pagespeed_ingest.py SEED_URLS	2 dead fallback blog slugs (PSI would measure 404 pages)	Fixed — replaced with live articles + rot-risk comment
scripts/update_blog_article_{2,3}.py	body cross-link to /blog/big-five-vs-disc-vs-belbin (dead)	Documented, not fixed — run-once historical generators; editing them does not change already-published content. This is content debt (a broken internal link in published articles), tracked separately, not test fragility
api/seo_mcp/, src/	no hardcoded article slugs found	none

3. Backend deploy smoke test hardened

PR #36's backend deploy went green on the server (Caddy validated, service active, api.cercol.team/blog 200, crawl_logs flowing) but the Action failed: the smoke test waited only ~10s (5 x 2s) while uvicorn took ~31s to boot two workers on the load-saturated shared VPS (the YELLOW 6 audit finding). False negative. Window is now 20 x 3s = up to 60s.

Test plan

• pytest -q api/tests/test_seo.py (full dist) — 57 passed
• Robustness gate — remove sampled article, still 57 passed
• pytest -q api/ — 176 passed
• vitest run — 204 passed
• ruff — covered by CI (not installed locally)
• CI frontend job runs build:full + test_seo against the complete dist (authoritative gate)

No production behavior change. Test + CI-config only; the auto-triggered backend deploy should now pass its (longer) smoke test.

🤖 Generated with Claude Code