Home page: fix 8 stale/wrong facts

- StatsBar: '+21% F1 bridge queries' -> '-44% Token Reduction' (real benchmark)
- StatsBar: '55 unit tests' -> '50' (actual: 28 test_core + 22 test_novelties)
- HeroSection: 'dual-pipeline' -> '3-pipeline', '40%' -> '44%' tokens
- HowItWorks step 02: 'both pipelines' -> 'all 3 pipelines' with LLM-Only named
- HowItWorks step 04: 'Claude (or...)' -> 'Any of 12 LLM providers'; add Judge+BERTScore
- BentoShowcase: 'both pipelines' -> 'all 3 pipelines'; 'HotpotQA' -> 'Wikipedia science'
- CTASection: 'Baseline RAG and GraphRAG' -> 'all 3 pipelines'; add Judge+BERTScore

web/src/components/home/BentoShowcase.tsx

@@ -25,8 +25,8 @@ export function BentoShowcase() {
                 Live Playground
               </h3>
               <p style={{ color: "rgba(255,255,255,0.8)", fontSize: "0.9375rem", maxWidth: "400px" }}>
-                Ask any question and watch both pipelines race side-by-side.
-                Real-time token counting, cost tracking, and quality metrics.
+                Ask any science question and watch all 3 pipelines run simultaneously —
+                LLM-Only, Basic RAG, and GraphRAG. Real-time tokens, cost, and accuracy.
               </p>
               <div className="flex gap-2 mt-4">
                 <span className="badge" style={{ background: "rgba(255,255,255,0.2)", color: "white", fontSize: "0.6875rem" }}>12 LLM Providers</span>
@@ -43,7 +43,7 @@ export function BentoShowcase() {
                 Benchmarks
               </h3>
               <p className="body-sm" style={{ color: "var(--color-on-dark-soft)" }}>
-                Run HotpotQA benchmarks with F1, EM, and radar charts.
+                Run Wikipedia science benchmarks with F1, LLM-Judge, BERTScore, and radar charts.
               </p>
             </div>
           </Link>

web/src/components/home/CTASection.tsx

@@ -42,8 +42,8 @@ export function CTASection() {
           maxWidth: "480px", margin: "0 auto 40px",
           lineHeight: 1.6,
         }}>
-          Run a live comparison between Baseline RAG and GraphRAG.
-          Measure F1, tokens, cost, and latency — all in your browser.
+          Run a live comparison across all 3 pipelines — LLM-Only, Basic RAG, and GraphRAG.
+          Measure tokens, cost, LLM-Judge accuracy, and BERTScore — all in your browser.
         </p>
         <div className="flex flex-wrap gap-4 justify-center">
           <Link href="/playground" className="btn btn-primary btn-lg no-underline">

web/src/components/home/HeroSection.tsx

@@ -36,8 +36,8 @@ export function HeroSection() {
             </h1>
 
             <p className="body-lg mb-10 animate-fade-in-up delay-200" style={{ maxWidth: "540px", color: "#3d3d3a" }}>
-              A dual-pipeline system that routes queries through knowledge graphs
-              when it matters — cutting tokens by 40%, speeding up multi-hop questions,
+              A 3-pipeline system that routes queries through knowledge graphs
+              when it matters — cutting tokens by 44%, speeding up multi-hop questions,
               and delivering measurably better answers.
             </p>
 

web/src/components/home/HowItWorks.tsx

@@ -11,7 +11,7 @@ const STEPS = [
   {
     number: "02",
     title: "Dual Pipeline Activation",
-    description: "Both pipelines execute simultaneously: Baseline RAG (vector search → LLM) and GraphRAG (entity extraction → graph traversal → LLM).",
+    description: "All 3 pipelines execute simultaneously: LLM-Only (no retrieval), Baseline RAG (vector search → LLM), and GraphRAG (entity extraction → graph traversal → LLM).",
     detail: "Schema-bounded extraction ensures valid entities | GSQL multi-hop traversal",
     color: "#0072CE",
   },
@@ -25,7 +25,7 @@ const STEPS = [
   {
     number: "04",
     title: "LLM Generation & Evaluation",
-    description: "Claude (or any of 12 providers) generates answers from the structured context. RAGAS evaluates F1, Exact Match, and quality metrics in real-time.",
+    description: "Any of 12 LLM providers (Gemini, GPT-4, Llama, etc.) generates answers. Evaluated with F1, Exact Match, LLM-as-a-Judge (PASS/FAIL), and BERTScore in real-time.",
     detail: "Cost tracking | Token counting | Latency measurement",
     color: "#cc785c",
   },

web/src/components/home/StatsBar.tsx

@@ -2,11 +2,11 @@
 
 export function StatsBar() {
   const stats = [
-    { value: "+21%", label: "F1 on Bridge Queries", color: "#FF6B00" },
+    { value: "−44%", label: "Token Reduction vs RAG", color: "#FF6B00" },
     { value: "4", label: "AI Factory Layers", color: "#002B49" },
     { value: "12", label: "LLM Providers", color: "#0072CE" },
     { value: "5", label: "Novel Features", color: "#cc785c" },
-    { value: "55", label: "Unit Tests", color: "#5db8a6" },
+    { value: "50", label: "Unit Tests", color: "#5db8a6" },
   ];
 
   return (