feat: memory, web components and landing rework

[arjunkmrm]

Summary

• New capability components: <Memory> (memory_read/write/list with path-traversal guard), <WebSearch> (Exa), <WebFetch> (raw GET, truncated)
• README updated to mention the new components
• Landing rework: new headline "Write your own Claude Code. / Run it anywhere.", lede updated, second section "Build your own components." with a Goal halt-predicate sketch, "Copy prompt" CTA that copies the agentjsx SKILL.md as a primer for whichever agent the user paste it into, dev-only Docs link removed
• Hero code block widened, long lines now scroll horizontally instead of wrapping, TS badge replaces box icon, Memory + WebSearch + WebFetch added to the hero snippet, linear McpServer collapsed to one line
• New examples/goal-test/: userspace emulation of Claude Code's /goal — loop agent.run + separate judge infer until verdict says condition holds. Verified end-to-end with sonnet-4-6 (Hello → judge fails → re-prompt with reason → Ahoy → judge passes)
• Vite config: server.fs.allow: [".."] so the SKILL.md outside web/ can be ?raw-imported

Test plan

• npx tsc --noEmit at repo root: clean
• cd web && npx tsc -b: clean
• cd examples/goal-test && infisical run --silent -- npx tsx cli.tsx: passes the two-iteration loop, judge ok=true on iteration 2
• Vite dev server on :5173 / :5174: hot-reloads cleanly