LIA

Frontend/documentation release — no backend change, no DB migration.

Added — /story field report & landing proof

• /story field-report page (6 languages): how LIA is built — nearly 100% of the code written by an AI under human direction (written rulebook, blocking automated checks, a review cycle where the human decides, adversarial audits), the structural trade-offs, operations, and the measured 8.0/10 across 20 areas audit profile, planned weaknesses included. Wired into the header nav (Présentation · Story · Philosophie · Technique · Blog · FAQ), both footers, a new landing RexSection and the FAQ ("How was LIA built?", 6 languages, mirrored in docs/knowledge/).
• Animated hero chat demo: three rotating scenarios mirroring the real display modes with per-mode title-bar chips — HITL (humorous email draft, Send/Edit controls, approval), HTML cards (rich weather card + proactive cross-domain initiative), Markdown (multi-agent place search, numbered-list reply). Static under prefers-reduced-motion.
• Engineering proof band: 19+ agents · 76 tools · 7 providers · 99+ voice languages · 10,000+ tests · 100+ ADRs · 120+ releases · 8.0/10 audit — every figure sourced from the codebase (LANDING_STATS documents each origin).
• Today-briefing feature card — the home-page briefing was absent from every showcase surface.

Changed — landing redesign & fact corrections

• Two-mode diagram rebuilt to the real LangGraph topology (ADR-070): router fork, five numbered Pipeline steps with the human-approval step highlighted, ReAct reason→act→observe loop, "or" divider, streaming-response convergence.
• Landing: split hero (ambient gradients replace the background image; version + update date kept as a freshness signal), compacted rem-breakpoint feature grid, featured-card use-case grid, portrait-friendly screenshot frame, FadeInOnScroll threshold fix (tall sections revealed thousands of px late), SSR-rendered counters, rewritten "Under the hood" intro.
• Facts corrected on every public surface (×6 locales + README + guides + llms.txt): 7 LLM providers, not 8 ("Mistral" was an Ollama-served model in the pricing seed), 76 tools, 394 Prometheus metrics, 22 Grafana dashboards, 54 LLM config types, 14 personalities, 100+ ADRs, Gemini embeddings (ADR-069) instead of OpenAI, honest latency framing (first-SSE-event + LLM-cascade note).
• French landing voice unified to tutoiement; de/es/it/zh use-case descriptions rewritten from telegraphic fragments to full sentences; guides version-harmonized (why 3.5 / how 2.6 across all six languages).

Fixed

• Anonymous visitors silently ejected from all public pages (production bug): the 401 handler's public-route list only covered login/register/root, so every logged-out visitor opening /why, /how, /blog or /faq bounced to /login after the session probe — invisible to developers, whose browsers always carry a session cookie. List extended to all public pages and locked by a filesystem-completeness test (31 new vitest tests): a new public page missing from the list now fails CI.
• Cloudflare Insights beacon blocked by CSP (console error on every prod page) — allowed and pinned as the single permitted external script host (ADR-098 policy+test discipline).
• SEO JsonLd repaired: sameAs pointed to a non-existent repository, softwareVersion was hardcoded to 1.8.2, WebSite description said 15 agents; llms.txt fully rewritten.
• Presentation subtitle said 13 slides; the carousel ships 15 (×6 locales).

Removed

• Dead landing components HeroBackground and StatsSection (superseded by the live demo and ProofSection); landing.stats.* / landing.hero.cta_secondary keys pruned from all six locales.

Tests: 200 vitest green (169 + 31 new), tsc strict & ESLint clean, i18n parity across all 6 locales, every change verified on the rendered app (desktop, mobile, dark mode, anonymous session).