Launch checklist — v0.1

Verified items say how they were verified. Anything unchecked is a real to-do, not an omission.

Data and research integrity

• [x] make validate passes; counts within limits and shown honestly on / (338 words, 115 meanings, 85 affixes, 50 atoms, 70 edges).
• [x] Every etymology edge has a source in the bibliography; Wiktionary-only edges ≤ 0.6 (validator-enforced).
• [x] Contested items are badged wherever they appear (6 of 70 edges).
• [x] The family constraint is tested with deliberately malformed edges, and re-asserted when the report figures are built.
• [x] "Semantic atoms are an engineering interlingua, not a theory of human cognition." appears in the UI (/atoms/, /about/) and in docs/SEMANTICS.md.
• [x] The report and dataset page state that the meanings are a selection, that every record is a draft, and that CDIAL/DEDR entry numbers are absent.
• [ ] Expert review of any record (draft → reviewed). None has happened; see docs/ROADMAP.md.

Build and release

• [x] 29 automated tests pass (make test). CI runs validate, tests, the prohibited-feature grep, typecheck, build, the link check and checksum verification.
• [x] next build succeeds; about 650 static pages are generated.
• [x] scripts/check_links.py: 29,480 internal links across 656 pages, none broken.
• [x] The dataset release (CSV, JSON, JSON Schemas, zip) is regenerated on every build with SHA-256 checksums; all verified.
• [x] No prohibited features: the grep for openai|anthropic|embedding|vector|auth|stripe|langchain in code and data is clean.
• [x] CITATION.cff, LICENSE (Apache-2.0) and data/LICENSE (CC BY-SA 4.0) are present.

Quality (Lighthouse, headless Edge, simulated mobile throttling, local static build)

Performance varied by up to 25 points between runs of the same page, so treat it as a range, not a score.

• [x] Accessibility ≥ 90: 100 on every page measured, after fixing two real findings (chart marks missing a role; a dimmed chip below contrast).
• [ ] Performance on pages whose content is largely Hindi or Telugu is about 70 under simulated slow 4G. The cause is the two Indic font files (~390 KB), which load only on pages that show that text. Glyph subsetting was measured and rejected (docs/DECISIONS.md, entry 12). The localhost run also overstates the cost, because fonts arrive before first paint. Re-measure on production, with real-user data, before optimising further.
• [ ] Telugu and Devanagari render correctly on a physical phone. Verified in headless Edge at 390 px (conjuncts form correctly), not on a device.
• [x] Light and dark themes both reviewed in screenshots.
• [ ] Interactive behaviour (hover readouts, the search palette, the theme toggle, table views) exercised in a real browser. It is implemented and compiled, not clicked through.

Deploy (IONOS VPS; see docs/DEPLOY.md)

• [x] Release / prune / rollback logic exercised locally against a fake server (7 deploys keep 5; rollback edge cases); both shell scripts pass bash -n.
• [ ] nginx -t and deploy/setup-server.sh run on the real server. They are written but not yet run against one.
• [ ] DNS A (and AAAA only if IPv6 works) records for suffeffix.com and www point at the VPS; IONOS firewall policy allows 22, 80 and 443.
• [ ] Certificate issued; certbot renew --dry-run succeeds.
• [ ] GitHub secrets VPS_HOST, VPS_SSH_KEY, VPS_KNOWN_HOSTS added; the first deploy ran and the live check returned 200.
• [ ] Production headers confirmed with curl -I (HSTS, CSP), including the http→https and www→apex redirects.
• [ ] SSH password logins disabled after key login is confirmed.
• [ ] sitemap.xml submitted to Search Console and Bing; social card checked.
• [ ] Lighthouse re-run against production and recorded above.